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>
76 #include <asm/sections.h>
77 #include <asm/tlbflush.h>
78 #include <asm/div64.h>
81 #include "page_reporting.h"
83 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
84 typedef int __bitwise fpi_t;
86 /* No special request */
87 #define FPI_NONE ((__force fpi_t)0)
90 * Skip free page reporting notification for the (possibly merged) page.
91 * This does not hinder free page reporting from grabbing the page,
92 * reporting it and marking it "reported" - it only skips notifying
93 * the free page reporting infrastructure about a newly freed page. For
94 * example, used when temporarily pulling a page from a freelist and
95 * putting it back unmodified.
97 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
100 * Place the (possibly merged) page to the tail of the freelist. Will ignore
101 * page shuffling (relevant code - e.g., memory onlining - is expected to
102 * shuffle the whole zone).
104 * Note: No code should rely on this flag for correctness - it's purely
105 * to allow for optimizations when handing back either fresh pages
106 * (memory onlining) or untouched pages (page isolation, free page
109 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
112 static DEFINE_MUTEX(pcp_batch_high_lock);
113 #define MIN_PERCPU_PAGELIST_FRACTION (8)
115 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
116 DEFINE_PER_CPU(int, numa_node);
117 EXPORT_PER_CPU_SYMBOL(numa_node);
120 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
122 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
124 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
125 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
126 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
127 * defined in <linux/topology.h>.
129 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
130 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
133 /* work_structs for global per-cpu drains */
136 struct work_struct work;
138 static DEFINE_MUTEX(pcpu_drain_mutex);
139 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
141 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
142 volatile unsigned long latent_entropy __latent_entropy;
143 EXPORT_SYMBOL(latent_entropy);
147 * Array of node states.
149 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
150 [N_POSSIBLE] = NODE_MASK_ALL,
151 [N_ONLINE] = { { [0] = 1UL } },
153 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
154 #ifdef CONFIG_HIGHMEM
155 [N_HIGH_MEMORY] = { { [0] = 1UL } },
157 [N_MEMORY] = { { [0] = 1UL } },
158 [N_CPU] = { { [0] = 1UL } },
161 EXPORT_SYMBOL(node_states);
163 atomic_long_t _totalram_pages __read_mostly;
164 EXPORT_SYMBOL(_totalram_pages);
165 unsigned long totalreserve_pages __read_mostly;
166 unsigned long totalcma_pages __read_mostly;
168 int percpu_pagelist_fraction;
169 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
170 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
171 EXPORT_SYMBOL(init_on_alloc);
173 DEFINE_STATIC_KEY_FALSE(init_on_free);
174 EXPORT_SYMBOL(init_on_free);
176 static bool _init_on_alloc_enabled_early __read_mostly
177 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
178 static int __init early_init_on_alloc(char *buf)
181 return kstrtobool(buf, &_init_on_alloc_enabled_early);
183 early_param("init_on_alloc", early_init_on_alloc);
185 static bool _init_on_free_enabled_early __read_mostly
186 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
187 static int __init early_init_on_free(char *buf)
189 return kstrtobool(buf, &_init_on_free_enabled_early);
191 early_param("init_on_free", early_init_on_free);
194 * A cached value of the page's pageblock's migratetype, used when the page is
195 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
196 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
197 * Also the migratetype set in the page does not necessarily match the pcplist
198 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
199 * other index - this ensures that it will be put on the correct CMA freelist.
201 static inline int get_pcppage_migratetype(struct page *page)
206 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
208 page->index = migratetype;
211 #ifdef CONFIG_PM_SLEEP
213 * The following functions are used by the suspend/hibernate code to temporarily
214 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
215 * while devices are suspended. To avoid races with the suspend/hibernate code,
216 * they should always be called with system_transition_mutex held
217 * (gfp_allowed_mask also should only be modified with system_transition_mutex
218 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
219 * with that modification).
222 static gfp_t saved_gfp_mask;
224 void pm_restore_gfp_mask(void)
226 WARN_ON(!mutex_is_locked(&system_transition_mutex));
227 if (saved_gfp_mask) {
228 gfp_allowed_mask = saved_gfp_mask;
233 void pm_restrict_gfp_mask(void)
235 WARN_ON(!mutex_is_locked(&system_transition_mutex));
236 WARN_ON(saved_gfp_mask);
237 saved_gfp_mask = gfp_allowed_mask;
238 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
241 bool pm_suspended_storage(void)
243 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
247 #endif /* CONFIG_PM_SLEEP */
249 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
250 unsigned int pageblock_order __read_mostly;
253 static void __free_pages_ok(struct page *page, unsigned int order,
257 * results with 256, 32 in the lowmem_reserve sysctl:
258 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
259 * 1G machine -> (16M dma, 784M normal, 224M high)
260 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
261 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
262 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
264 * TBD: should special case ZONE_DMA32 machines here - in those we normally
265 * don't need any ZONE_NORMAL reservation
267 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
268 #ifdef CONFIG_ZONE_DMA
271 #ifdef CONFIG_ZONE_DMA32
275 #ifdef CONFIG_HIGHMEM
281 static char * const zone_names[MAX_NR_ZONES] = {
282 #ifdef CONFIG_ZONE_DMA
285 #ifdef CONFIG_ZONE_DMA32
289 #ifdef CONFIG_HIGHMEM
293 #ifdef CONFIG_ZONE_DEVICE
298 const char * const migratetype_names[MIGRATE_TYPES] = {
306 #ifdef CONFIG_MEMORY_ISOLATION
311 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
312 [NULL_COMPOUND_DTOR] = NULL,
313 [COMPOUND_PAGE_DTOR] = free_compound_page,
314 #ifdef CONFIG_HUGETLB_PAGE
315 [HUGETLB_PAGE_DTOR] = free_huge_page,
317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
318 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
322 int min_free_kbytes = 1024;
323 int user_min_free_kbytes = -1;
324 #ifdef CONFIG_DISCONTIGMEM
326 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
327 * are not on separate NUMA nodes. Functionally this works but with
328 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
329 * quite small. By default, do not boost watermarks on discontigmem as in
330 * many cases very high-order allocations like THP are likely to be
331 * unsupported and the premature reclaim offsets the advantage of long-term
332 * fragmentation avoidance.
334 int watermark_boost_factor __read_mostly;
336 int watermark_boost_factor __read_mostly = 15000;
338 int watermark_scale_factor = 10;
340 static unsigned long nr_kernel_pages __initdata;
341 static unsigned long nr_all_pages __initdata;
342 static unsigned long dma_reserve __initdata;
344 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
345 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
346 static unsigned long required_kernelcore __initdata;
347 static unsigned long required_kernelcore_percent __initdata;
348 static unsigned long required_movablecore __initdata;
349 static unsigned long required_movablecore_percent __initdata;
350 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
351 static bool mirrored_kernelcore __meminitdata;
353 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
355 EXPORT_SYMBOL(movable_zone);
358 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
359 unsigned int nr_online_nodes __read_mostly = 1;
360 EXPORT_SYMBOL(nr_node_ids);
361 EXPORT_SYMBOL(nr_online_nodes);
364 int page_group_by_mobility_disabled __read_mostly;
366 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
368 * During boot we initialize deferred pages on-demand, as needed, but once
369 * page_alloc_init_late() has finished, the deferred pages are all initialized,
370 * and we can permanently disable that path.
372 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
375 * Calling kasan_free_pages() only after deferred memory initialization
376 * has completed. Poisoning pages during deferred memory init will greatly
377 * lengthen the process and cause problem in large memory systems as the
378 * deferred pages initialization is done with interrupt disabled.
380 * Assuming that there will be no reference to those newly initialized
381 * pages before they are ever allocated, this should have no effect on
382 * KASAN memory tracking as the poison will be properly inserted at page
383 * allocation time. The only corner case is when pages are allocated by
384 * on-demand allocation and then freed again before the deferred pages
385 * initialization is done, but this is not likely to happen.
387 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
389 if (!static_branch_unlikely(&deferred_pages))
390 kasan_free_pages(page, order);
393 /* Returns true if the struct page for the pfn is uninitialised */
394 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
396 int nid = early_pfn_to_nid(pfn);
398 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
405 * Returns true when the remaining initialisation should be deferred until
406 * later in the boot cycle when it can be parallelised.
408 static bool __meminit
409 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
411 static unsigned long prev_end_pfn, nr_initialised;
414 * prev_end_pfn static that contains the end of previous zone
415 * No need to protect because called very early in boot before smp_init.
417 if (prev_end_pfn != end_pfn) {
418 prev_end_pfn = end_pfn;
422 /* Always populate low zones for address-constrained allocations */
423 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
427 * We start only with one section of pages, more pages are added as
428 * needed until the rest of deferred pages are initialized.
431 if ((nr_initialised > PAGES_PER_SECTION) &&
432 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
433 NODE_DATA(nid)->first_deferred_pfn = pfn;
439 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
441 static inline bool early_page_uninitialised(unsigned long pfn)
446 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
452 /* Return a pointer to the bitmap storing bits affecting a block of pages */
453 static inline unsigned long *get_pageblock_bitmap(struct page *page,
456 #ifdef CONFIG_SPARSEMEM
457 return section_to_usemap(__pfn_to_section(pfn));
459 return page_zone(page)->pageblock_flags;
460 #endif /* CONFIG_SPARSEMEM */
463 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
465 #ifdef CONFIG_SPARSEMEM
466 pfn &= (PAGES_PER_SECTION-1);
468 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
469 #endif /* CONFIG_SPARSEMEM */
470 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
474 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
475 * @page: The page within the block of interest
476 * @pfn: The target page frame number
477 * @mask: mask of bits that the caller is interested in
479 * Return: pageblock_bits flags
481 static __always_inline
482 unsigned long __get_pfnblock_flags_mask(struct page *page,
486 unsigned long *bitmap;
487 unsigned long bitidx, word_bitidx;
490 bitmap = get_pageblock_bitmap(page, pfn);
491 bitidx = pfn_to_bitidx(page, pfn);
492 word_bitidx = bitidx / BITS_PER_LONG;
493 bitidx &= (BITS_PER_LONG-1);
495 word = bitmap[word_bitidx];
496 return (word >> bitidx) & mask;
499 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
502 return __get_pfnblock_flags_mask(page, pfn, mask);
505 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
507 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
538 word = READ_ONCE(bitmap[word_bitidx]);
540 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
541 if (word == old_word)
547 void set_pageblock_migratetype(struct page *page, int migratetype)
549 if (unlikely(page_group_by_mobility_disabled &&
550 migratetype < MIGRATE_PCPTYPES))
551 migratetype = MIGRATE_UNMOVABLE;
553 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
554 page_to_pfn(page), MIGRATETYPE_MASK);
557 #ifdef CONFIG_DEBUG_VM
558 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
562 unsigned long pfn = page_to_pfn(page);
563 unsigned long sp, start_pfn;
566 seq = zone_span_seqbegin(zone);
567 start_pfn = zone->zone_start_pfn;
568 sp = zone->spanned_pages;
569 if (!zone_spans_pfn(zone, pfn))
571 } while (zone_span_seqretry(zone, seq));
574 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
575 pfn, zone_to_nid(zone), zone->name,
576 start_pfn, start_pfn + sp);
581 static int page_is_consistent(struct zone *zone, struct page *page)
583 if (!pfn_valid_within(page_to_pfn(page)))
585 if (zone != page_zone(page))
591 * Temporary debugging check for pages not lying within a given zone.
593 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
595 if (page_outside_zone_boundaries(zone, page))
597 if (!page_is_consistent(zone, page))
603 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
609 static void bad_page(struct page *page, const char *reason)
611 static unsigned long resume;
612 static unsigned long nr_shown;
613 static unsigned long nr_unshown;
616 * Allow a burst of 60 reports, then keep quiet for that minute;
617 * or allow a steady drip of one report per second.
619 if (nr_shown == 60) {
620 if (time_before(jiffies, resume)) {
626 "BUG: Bad page state: %lu messages suppressed\n",
633 resume = jiffies + 60 * HZ;
635 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
636 current->comm, page_to_pfn(page));
637 __dump_page(page, reason);
638 dump_page_owner(page);
643 /* Leave bad fields for debug, except PageBuddy could make trouble */
644 page_mapcount_reset(page); /* remove PageBuddy */
645 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
649 * Higher-order pages are called "compound pages". They are structured thusly:
651 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
653 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
654 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
656 * The first tail page's ->compound_dtor holds the offset in array of compound
657 * page destructors. See compound_page_dtors.
659 * The first tail page's ->compound_order holds the order of allocation.
660 * This usage means that zero-order pages may not be compound.
663 void free_compound_page(struct page *page)
665 mem_cgroup_uncharge(page);
666 __free_pages_ok(page, compound_order(page), FPI_NONE);
669 void prep_compound_page(struct page *page, unsigned int order)
672 int nr_pages = 1 << order;
675 for (i = 1; i < nr_pages; i++) {
676 struct page *p = page + i;
677 set_page_count(p, 0);
678 p->mapping = TAIL_MAPPING;
679 set_compound_head(p, page);
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
684 atomic_set(compound_mapcount_ptr(page), -1);
685 if (hpage_pincount_available(page))
686 atomic_set(compound_pincount_ptr(page), 0);
689 #ifdef CONFIG_DEBUG_PAGEALLOC
690 unsigned int _debug_guardpage_minorder;
692 bool _debug_pagealloc_enabled_early __read_mostly
693 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
694 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
695 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled);
698 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
700 static int __init early_debug_pagealloc(char *buf)
702 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
704 early_param("debug_pagealloc", early_debug_pagealloc);
706 static int __init debug_guardpage_minorder_setup(char *buf)
710 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
711 pr_err("Bad debug_guardpage_minorder value\n");
714 _debug_guardpage_minorder = res;
715 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
718 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
720 static inline bool set_page_guard(struct zone *zone, struct page *page,
721 unsigned int order, int migratetype)
723 if (!debug_guardpage_enabled())
726 if (order >= debug_guardpage_minorder())
729 __SetPageGuard(page);
730 INIT_LIST_HEAD(&page->lru);
731 set_page_private(page, order);
732 /* Guard pages are not available for any usage */
733 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
738 static inline void clear_page_guard(struct zone *zone, struct page *page,
739 unsigned int order, int migratetype)
741 if (!debug_guardpage_enabled())
744 __ClearPageGuard(page);
746 set_page_private(page, 0);
747 if (!is_migrate_isolate(migratetype))
748 __mod_zone_freepage_state(zone, (1 << order), migratetype);
751 static inline bool set_page_guard(struct zone *zone, struct page *page,
752 unsigned int order, int migratetype) { return false; }
753 static inline void clear_page_guard(struct zone *zone, struct page *page,
754 unsigned int order, int migratetype) {}
758 * Enable static keys related to various memory debugging and hardening options.
759 * Some override others, and depend on early params that are evaluated in the
760 * order of appearance. So we need to first gather the full picture of what was
761 * enabled, and then make decisions.
763 void init_mem_debugging_and_hardening(void)
765 if (_init_on_alloc_enabled_early) {
766 if (page_poisoning_enabled())
767 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
768 "will take precedence over init_on_alloc\n");
770 static_branch_enable(&init_on_alloc);
772 if (_init_on_free_enabled_early) {
773 if (page_poisoning_enabled())
774 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
775 "will take precedence over init_on_free\n");
777 static_branch_enable(&init_on_free);
780 #ifdef CONFIG_DEBUG_PAGEALLOC
781 if (!debug_pagealloc_enabled())
784 static_branch_enable(&_debug_pagealloc_enabled);
786 if (!debug_guardpage_minorder())
789 static_branch_enable(&_debug_guardpage_enabled);
793 static inline void set_buddy_order(struct page *page, unsigned int order)
795 set_page_private(page, order);
796 __SetPageBuddy(page);
800 * This function checks whether a page is free && is the buddy
801 * we can coalesce a page and its buddy if
802 * (a) the buddy is not in a hole (check before calling!) &&
803 * (b) the buddy is in the buddy system &&
804 * (c) a page and its buddy have the same order &&
805 * (d) a page and its buddy are in the same zone.
807 * For recording whether a page is in the buddy system, we set PageBuddy.
808 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
810 * For recording page's order, we use page_private(page).
812 static inline bool page_is_buddy(struct page *page, struct page *buddy,
815 if (!page_is_guard(buddy) && !PageBuddy(buddy))
818 if (buddy_order(buddy) != order)
822 * zone check is done late to avoid uselessly calculating
823 * zone/node ids for pages that could never merge.
825 if (page_zone_id(page) != page_zone_id(buddy))
828 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
833 #ifdef CONFIG_COMPACTION
834 static inline struct capture_control *task_capc(struct zone *zone)
836 struct capture_control *capc = current->capture_control;
838 return unlikely(capc) &&
839 !(current->flags & PF_KTHREAD) &&
841 capc->cc->zone == zone ? capc : NULL;
845 compaction_capture(struct capture_control *capc, struct page *page,
846 int order, int migratetype)
848 if (!capc || order != capc->cc->order)
851 /* Do not accidentally pollute CMA or isolated regions*/
852 if (is_migrate_cma(migratetype) ||
853 is_migrate_isolate(migratetype))
857 * Do not let lower order allocations polluate a movable pageblock.
858 * This might let an unmovable request use a reclaimable pageblock
859 * and vice-versa but no more than normal fallback logic which can
860 * have trouble finding a high-order free page.
862 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
870 static inline struct capture_control *task_capc(struct zone *zone)
876 compaction_capture(struct capture_control *capc, struct page *page,
877 int order, int migratetype)
881 #endif /* CONFIG_COMPACTION */
883 /* Used for pages not on another list */
884 static inline void add_to_free_list(struct page *page, struct zone *zone,
885 unsigned int order, int migratetype)
887 struct free_area *area = &zone->free_area[order];
889 list_add(&page->lru, &area->free_list[migratetype]);
893 /* Used for pages not on another list */
894 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
895 unsigned int order, int migratetype)
897 struct free_area *area = &zone->free_area[order];
899 list_add_tail(&page->lru, &area->free_list[migratetype]);
904 * Used for pages which are on another list. Move the pages to the tail
905 * of the list - so the moved pages won't immediately be considered for
906 * allocation again (e.g., optimization for memory onlining).
908 static inline void move_to_free_list(struct page *page, struct zone *zone,
909 unsigned int order, int migratetype)
911 struct free_area *area = &zone->free_area[order];
913 list_move_tail(&page->lru, &area->free_list[migratetype]);
916 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
919 /* clear reported state and update reported page count */
920 if (page_reported(page))
921 __ClearPageReported(page);
923 list_del(&page->lru);
924 __ClearPageBuddy(page);
925 set_page_private(page, 0);
926 zone->free_area[order].nr_free--;
930 * If this is not the largest possible page, check if the buddy
931 * of the next-highest order is free. If it is, it's possible
932 * that pages are being freed that will coalesce soon. In case,
933 * that is happening, add the free page to the tail of the list
934 * so it's less likely to be used soon and more likely to be merged
935 * as a higher order page
938 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
939 struct page *page, unsigned int order)
941 struct page *higher_page, *higher_buddy;
942 unsigned long combined_pfn;
944 if (order >= MAX_ORDER - 2)
947 if (!pfn_valid_within(buddy_pfn))
950 combined_pfn = buddy_pfn & pfn;
951 higher_page = page + (combined_pfn - pfn);
952 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
953 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
955 return pfn_valid_within(buddy_pfn) &&
956 page_is_buddy(higher_page, higher_buddy, order + 1);
960 * Freeing function for a buddy system allocator.
962 * The concept of a buddy system is to maintain direct-mapped table
963 * (containing bit values) for memory blocks of various "orders".
964 * The bottom level table contains the map for the smallest allocatable
965 * units of memory (here, pages), and each level above it describes
966 * pairs of units from the levels below, hence, "buddies".
967 * At a high level, all that happens here is marking the table entry
968 * at the bottom level available, and propagating the changes upward
969 * as necessary, plus some accounting needed to play nicely with other
970 * parts of the VM system.
971 * At each level, we keep a list of pages, which are heads of continuous
972 * free pages of length of (1 << order) and marked with PageBuddy.
973 * Page's order is recorded in page_private(page) field.
974 * So when we are allocating or freeing one, we can derive the state of the
975 * other. That is, if we allocate a small block, and both were
976 * free, the remainder of the region must be split into blocks.
977 * If a block is freed, and its buddy is also free, then this
978 * triggers coalescing into a block of larger size.
983 static inline void __free_one_page(struct page *page,
985 struct zone *zone, unsigned int order,
986 int migratetype, fpi_t fpi_flags)
988 struct capture_control *capc = task_capc(zone);
989 unsigned long buddy_pfn;
990 unsigned long combined_pfn;
991 unsigned int max_order;
995 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
997 VM_BUG_ON(!zone_is_initialized(zone));
998 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1000 VM_BUG_ON(migratetype == -1);
1001 if (likely(!is_migrate_isolate(migratetype)))
1002 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1004 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1005 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1008 while (order < max_order) {
1009 if (compaction_capture(capc, page, order, migratetype)) {
1010 __mod_zone_freepage_state(zone, -(1 << order),
1014 buddy_pfn = __find_buddy_pfn(pfn, order);
1015 buddy = page + (buddy_pfn - pfn);
1017 if (!pfn_valid_within(buddy_pfn))
1019 if (!page_is_buddy(page, buddy, order))
1022 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1023 * merge with it and move up one order.
1025 if (page_is_guard(buddy))
1026 clear_page_guard(zone, buddy, order, migratetype);
1028 del_page_from_free_list(buddy, zone, order);
1029 combined_pfn = buddy_pfn & pfn;
1030 page = page + (combined_pfn - pfn);
1034 if (order < MAX_ORDER - 1) {
1035 /* If we are here, it means order is >= pageblock_order.
1036 * We want to prevent merge between freepages on isolate
1037 * pageblock and normal pageblock. Without this, pageblock
1038 * isolation could cause incorrect freepage or CMA accounting.
1040 * We don't want to hit this code for the more frequent
1041 * low-order merging.
1043 if (unlikely(has_isolate_pageblock(zone))) {
1046 buddy_pfn = __find_buddy_pfn(pfn, order);
1047 buddy = page + (buddy_pfn - pfn);
1048 buddy_mt = get_pageblock_migratetype(buddy);
1050 if (migratetype != buddy_mt
1051 && (is_migrate_isolate(migratetype) ||
1052 is_migrate_isolate(buddy_mt)))
1055 max_order = order + 1;
1056 goto continue_merging;
1060 set_buddy_order(page, order);
1062 if (fpi_flags & FPI_TO_TAIL)
1064 else if (is_shuffle_order(order))
1065 to_tail = shuffle_pick_tail();
1067 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1070 add_to_free_list_tail(page, zone, order, migratetype);
1072 add_to_free_list(page, zone, order, migratetype);
1074 /* Notify page reporting subsystem of freed page */
1075 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1076 page_reporting_notify_free(order);
1080 * A bad page could be due to a number of fields. Instead of multiple branches,
1081 * try and check multiple fields with one check. The caller must do a detailed
1082 * check if necessary.
1084 static inline bool page_expected_state(struct page *page,
1085 unsigned long check_flags)
1087 if (unlikely(atomic_read(&page->_mapcount) != -1))
1090 if (unlikely((unsigned long)page->mapping |
1091 page_ref_count(page) |
1093 (unsigned long)page->mem_cgroup |
1095 (page->flags & check_flags)))
1101 static const char *page_bad_reason(struct page *page, unsigned long flags)
1103 const char *bad_reason = NULL;
1105 if (unlikely(atomic_read(&page->_mapcount) != -1))
1106 bad_reason = "nonzero mapcount";
1107 if (unlikely(page->mapping != NULL))
1108 bad_reason = "non-NULL mapping";
1109 if (unlikely(page_ref_count(page) != 0))
1110 bad_reason = "nonzero _refcount";
1111 if (unlikely(page->flags & flags)) {
1112 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1113 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1115 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1118 if (unlikely(page->mem_cgroup))
1119 bad_reason = "page still charged to cgroup";
1124 static void check_free_page_bad(struct page *page)
1127 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1130 static inline int check_free_page(struct page *page)
1132 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1135 /* Something has gone sideways, find it */
1136 check_free_page_bad(page);
1140 static int free_tail_pages_check(struct page *head_page, struct page *page)
1145 * We rely page->lru.next never has bit 0 set, unless the page
1146 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1148 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1150 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1154 switch (page - head_page) {
1156 /* the first tail page: ->mapping may be compound_mapcount() */
1157 if (unlikely(compound_mapcount(page))) {
1158 bad_page(page, "nonzero compound_mapcount");
1164 * the second tail page: ->mapping is
1165 * deferred_list.next -- ignore value.
1169 if (page->mapping != TAIL_MAPPING) {
1170 bad_page(page, "corrupted mapping in tail page");
1175 if (unlikely(!PageTail(page))) {
1176 bad_page(page, "PageTail not set");
1179 if (unlikely(compound_head(page) != head_page)) {
1180 bad_page(page, "compound_head not consistent");
1185 page->mapping = NULL;
1186 clear_compound_head(page);
1190 static void kernel_init_free_pages(struct page *page, int numpages)
1194 /* s390's use of memset() could override KASAN redzones. */
1195 kasan_disable_current();
1196 for (i = 0; i < numpages; i++)
1197 clear_highpage(page + i);
1198 kasan_enable_current();
1201 static __always_inline bool free_pages_prepare(struct page *page,
1202 unsigned int order, bool check_free)
1206 VM_BUG_ON_PAGE(PageTail(page), page);
1208 trace_mm_page_free(page, order);
1210 if (unlikely(PageHWPoison(page)) && !order) {
1212 * Do not let hwpoison pages hit pcplists/buddy
1213 * Untie memcg state and reset page's owner
1215 if (memcg_kmem_enabled() && PageKmemcg(page))
1216 __memcg_kmem_uncharge_page(page, order);
1217 reset_page_owner(page, order);
1222 * Check tail pages before head page information is cleared to
1223 * avoid checking PageCompound for order-0 pages.
1225 if (unlikely(order)) {
1226 bool compound = PageCompound(page);
1229 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1232 ClearPageDoubleMap(page);
1233 for (i = 1; i < (1 << order); i++) {
1235 bad += free_tail_pages_check(page, page + i);
1236 if (unlikely(check_free_page(page + i))) {
1240 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1243 if (PageMappingFlags(page))
1244 page->mapping = NULL;
1245 if (memcg_kmem_enabled() && PageKmemcg(page))
1246 __memcg_kmem_uncharge_page(page, order);
1248 bad += check_free_page(page);
1252 page_cpupid_reset_last(page);
1253 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1254 reset_page_owner(page, order);
1256 if (!PageHighMem(page)) {
1257 debug_check_no_locks_freed(page_address(page),
1258 PAGE_SIZE << order);
1259 debug_check_no_obj_freed(page_address(page),
1260 PAGE_SIZE << order);
1262 if (want_init_on_free())
1263 kernel_init_free_pages(page, 1 << order);
1265 kernel_poison_pages(page, 1 << order, 0);
1267 * arch_free_page() can make the page's contents inaccessible. s390
1268 * does this. So nothing which can access the page's contents should
1269 * happen after this.
1271 arch_free_page(page, order);
1273 debug_pagealloc_unmap_pages(page, 1 << order);
1275 kasan_free_nondeferred_pages(page, order);
1280 #ifdef CONFIG_DEBUG_VM
1282 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1283 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1284 * moved from pcp lists to free lists.
1286 static bool free_pcp_prepare(struct page *page)
1288 return free_pages_prepare(page, 0, true);
1291 static bool bulkfree_pcp_prepare(struct page *page)
1293 if (debug_pagealloc_enabled_static())
1294 return check_free_page(page);
1300 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1301 * moving from pcp lists to free list in order to reduce overhead. With
1302 * debug_pagealloc enabled, they are checked also immediately when being freed
1305 static bool free_pcp_prepare(struct page *page)
1307 if (debug_pagealloc_enabled_static())
1308 return free_pages_prepare(page, 0, true);
1310 return free_pages_prepare(page, 0, false);
1313 static bool bulkfree_pcp_prepare(struct page *page)
1315 return check_free_page(page);
1317 #endif /* CONFIG_DEBUG_VM */
1319 static inline void prefetch_buddy(struct page *page)
1321 unsigned long pfn = page_to_pfn(page);
1322 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1323 struct page *buddy = page + (buddy_pfn - pfn);
1329 * Frees a number of pages from the PCP lists
1330 * Assumes all pages on list are in same zone, and of same order.
1331 * count is the number of pages to free.
1333 * If the zone was previously in an "all pages pinned" state then look to
1334 * see if this freeing clears that state.
1336 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1337 * pinned" detection logic.
1339 static void free_pcppages_bulk(struct zone *zone, int count,
1340 struct per_cpu_pages *pcp)
1342 int migratetype = 0;
1344 int prefetch_nr = READ_ONCE(pcp->batch);
1345 bool isolated_pageblocks;
1346 struct page *page, *tmp;
1350 * Ensure proper count is passed which otherwise would stuck in the
1351 * below while (list_empty(list)) loop.
1353 count = min(pcp->count, count);
1355 struct list_head *list;
1358 * Remove pages from lists in a round-robin fashion. A
1359 * batch_free count is maintained that is incremented when an
1360 * empty list is encountered. This is so more pages are freed
1361 * off fuller lists instead of spinning excessively around empty
1366 if (++migratetype == MIGRATE_PCPTYPES)
1368 list = &pcp->lists[migratetype];
1369 } while (list_empty(list));
1371 /* This is the only non-empty list. Free them all. */
1372 if (batch_free == MIGRATE_PCPTYPES)
1376 page = list_last_entry(list, struct page, lru);
1377 /* must delete to avoid corrupting pcp list */
1378 list_del(&page->lru);
1381 if (bulkfree_pcp_prepare(page))
1384 list_add_tail(&page->lru, &head);
1387 * We are going to put the page back to the global
1388 * pool, prefetch its buddy to speed up later access
1389 * under zone->lock. It is believed the overhead of
1390 * an additional test and calculating buddy_pfn here
1391 * can be offset by reduced memory latency later. To
1392 * avoid excessive prefetching due to large count, only
1393 * prefetch buddy for the first pcp->batch nr of pages.
1396 prefetch_buddy(page);
1399 } while (--count && --batch_free && !list_empty(list));
1402 spin_lock(&zone->lock);
1403 isolated_pageblocks = has_isolate_pageblock(zone);
1406 * Use safe version since after __free_one_page(),
1407 * page->lru.next will not point to original list.
1409 list_for_each_entry_safe(page, tmp, &head, lru) {
1410 int mt = get_pcppage_migratetype(page);
1411 /* MIGRATE_ISOLATE page should not go to pcplists */
1412 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1413 /* Pageblock could have been isolated meanwhile */
1414 if (unlikely(isolated_pageblocks))
1415 mt = get_pageblock_migratetype(page);
1417 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1418 trace_mm_page_pcpu_drain(page, 0, mt);
1420 spin_unlock(&zone->lock);
1423 static void free_one_page(struct zone *zone,
1424 struct page *page, unsigned long pfn,
1426 int migratetype, fpi_t fpi_flags)
1428 spin_lock(&zone->lock);
1429 if (unlikely(has_isolate_pageblock(zone) ||
1430 is_migrate_isolate(migratetype))) {
1431 migratetype = get_pfnblock_migratetype(page, pfn);
1433 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1434 spin_unlock(&zone->lock);
1437 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1438 unsigned long zone, int nid)
1440 mm_zero_struct_page(page);
1441 set_page_links(page, zone, nid, pfn);
1442 init_page_count(page);
1443 page_mapcount_reset(page);
1444 page_cpupid_reset_last(page);
1445 page_kasan_tag_reset(page);
1447 INIT_LIST_HEAD(&page->lru);
1448 #ifdef WANT_PAGE_VIRTUAL
1449 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1450 if (!is_highmem_idx(zone))
1451 set_page_address(page, __va(pfn << PAGE_SHIFT));
1455 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1456 static void __meminit init_reserved_page(unsigned long pfn)
1461 if (!early_page_uninitialised(pfn))
1464 nid = early_pfn_to_nid(pfn);
1465 pgdat = NODE_DATA(nid);
1467 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1468 struct zone *zone = &pgdat->node_zones[zid];
1470 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1473 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1476 static inline void init_reserved_page(unsigned long pfn)
1479 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1482 * Initialised pages do not have PageReserved set. This function is
1483 * called for each range allocated by the bootmem allocator and
1484 * marks the pages PageReserved. The remaining valid pages are later
1485 * sent to the buddy page allocator.
1487 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1489 unsigned long start_pfn = PFN_DOWN(start);
1490 unsigned long end_pfn = PFN_UP(end);
1492 for (; start_pfn < end_pfn; start_pfn++) {
1493 if (pfn_valid(start_pfn)) {
1494 struct page *page = pfn_to_page(start_pfn);
1496 init_reserved_page(start_pfn);
1498 /* Avoid false-positive PageTail() */
1499 INIT_LIST_HEAD(&page->lru);
1502 * no need for atomic set_bit because the struct
1503 * page is not visible yet so nobody should
1506 __SetPageReserved(page);
1511 static void __free_pages_ok(struct page *page, unsigned int order,
1514 unsigned long flags;
1516 unsigned long pfn = page_to_pfn(page);
1518 if (!free_pages_prepare(page, order, true))
1521 migratetype = get_pfnblock_migratetype(page, pfn);
1522 local_irq_save(flags);
1523 __count_vm_events(PGFREE, 1 << order);
1524 free_one_page(page_zone(page), page, pfn, order, migratetype,
1526 local_irq_restore(flags);
1529 void __free_pages_core(struct page *page, unsigned int order)
1531 unsigned int nr_pages = 1 << order;
1532 struct page *p = page;
1536 * When initializing the memmap, __init_single_page() sets the refcount
1537 * of all pages to 1 ("allocated"/"not free"). We have to set the
1538 * refcount of all involved pages to 0.
1541 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1543 __ClearPageReserved(p);
1544 set_page_count(p, 0);
1546 __ClearPageReserved(p);
1547 set_page_count(p, 0);
1549 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1552 * Bypass PCP and place fresh pages right to the tail, primarily
1553 * relevant for memory onlining.
1555 __free_pages_ok(page, order, FPI_TO_TAIL);
1558 #ifdef CONFIG_NEED_MULTIPLE_NODES
1561 * During memory init memblocks map pfns to nids. The search is expensive and
1562 * this caches recent lookups. The implementation of __early_pfn_to_nid
1563 * treats start/end as pfns.
1565 struct mminit_pfnnid_cache {
1566 unsigned long last_start;
1567 unsigned long last_end;
1571 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1574 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1576 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1577 struct mminit_pfnnid_cache *state)
1579 unsigned long start_pfn, end_pfn;
1582 if (state->last_start <= pfn && pfn < state->last_end)
1583 return state->last_nid;
1585 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1586 if (nid != NUMA_NO_NODE) {
1587 state->last_start = start_pfn;
1588 state->last_end = end_pfn;
1589 state->last_nid = nid;
1595 int __meminit early_pfn_to_nid(unsigned long pfn)
1597 static DEFINE_SPINLOCK(early_pfn_lock);
1600 spin_lock(&early_pfn_lock);
1601 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1603 nid = first_online_node;
1604 spin_unlock(&early_pfn_lock);
1608 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1610 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1613 if (early_page_uninitialised(pfn))
1615 __free_pages_core(page, order);
1619 * Check that the whole (or subset of) a pageblock given by the interval of
1620 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1621 * with the migration of free compaction scanner. The scanners then need to
1622 * use only pfn_valid_within() check for arches that allow holes within
1625 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1627 * It's possible on some configurations to have a setup like node0 node1 node0
1628 * i.e. it's possible that all pages within a zones range of pages do not
1629 * belong to a single zone. We assume that a border between node0 and node1
1630 * can occur within a single pageblock, but not a node0 node1 node0
1631 * interleaving within a single pageblock. It is therefore sufficient to check
1632 * the first and last page of a pageblock and avoid checking each individual
1633 * page in a pageblock.
1635 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1636 unsigned long end_pfn, struct zone *zone)
1638 struct page *start_page;
1639 struct page *end_page;
1641 /* end_pfn is one past the range we are checking */
1644 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1647 start_page = pfn_to_online_page(start_pfn);
1651 if (page_zone(start_page) != zone)
1654 end_page = pfn_to_page(end_pfn);
1656 /* This gives a shorter code than deriving page_zone(end_page) */
1657 if (page_zone_id(start_page) != page_zone_id(end_page))
1663 void set_zone_contiguous(struct zone *zone)
1665 unsigned long block_start_pfn = zone->zone_start_pfn;
1666 unsigned long block_end_pfn;
1668 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1669 for (; block_start_pfn < zone_end_pfn(zone);
1670 block_start_pfn = block_end_pfn,
1671 block_end_pfn += pageblock_nr_pages) {
1673 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1675 if (!__pageblock_pfn_to_page(block_start_pfn,
1676 block_end_pfn, zone))
1681 /* We confirm that there is no hole */
1682 zone->contiguous = true;
1685 void clear_zone_contiguous(struct zone *zone)
1687 zone->contiguous = false;
1690 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1691 static void __init deferred_free_range(unsigned long pfn,
1692 unsigned long nr_pages)
1700 page = pfn_to_page(pfn);
1702 /* Free a large naturally-aligned chunk if possible */
1703 if (nr_pages == pageblock_nr_pages &&
1704 (pfn & (pageblock_nr_pages - 1)) == 0) {
1705 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1706 __free_pages_core(page, pageblock_order);
1710 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1711 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1712 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1713 __free_pages_core(page, 0);
1717 /* Completion tracking for deferred_init_memmap() threads */
1718 static atomic_t pgdat_init_n_undone __initdata;
1719 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1721 static inline void __init pgdat_init_report_one_done(void)
1723 if (atomic_dec_and_test(&pgdat_init_n_undone))
1724 complete(&pgdat_init_all_done_comp);
1728 * Returns true if page needs to be initialized or freed to buddy allocator.
1730 * First we check if pfn is valid on architectures where it is possible to have
1731 * holes within pageblock_nr_pages. On systems where it is not possible, this
1732 * function is optimized out.
1734 * Then, we check if a current large page is valid by only checking the validity
1737 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1739 if (!pfn_valid_within(pfn))
1741 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1747 * Free pages to buddy allocator. Try to free aligned pages in
1748 * pageblock_nr_pages sizes.
1750 static void __init deferred_free_pages(unsigned long pfn,
1751 unsigned long end_pfn)
1753 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1754 unsigned long nr_free = 0;
1756 for (; pfn < end_pfn; pfn++) {
1757 if (!deferred_pfn_valid(pfn)) {
1758 deferred_free_range(pfn - nr_free, nr_free);
1760 } else if (!(pfn & nr_pgmask)) {
1761 deferred_free_range(pfn - nr_free, nr_free);
1767 /* Free the last block of pages to allocator */
1768 deferred_free_range(pfn - nr_free, nr_free);
1772 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1773 * by performing it only once every pageblock_nr_pages.
1774 * Return number of pages initialized.
1776 static unsigned long __init deferred_init_pages(struct zone *zone,
1778 unsigned long end_pfn)
1780 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1781 int nid = zone_to_nid(zone);
1782 unsigned long nr_pages = 0;
1783 int zid = zone_idx(zone);
1784 struct page *page = NULL;
1786 for (; pfn < end_pfn; pfn++) {
1787 if (!deferred_pfn_valid(pfn)) {
1790 } else if (!page || !(pfn & nr_pgmask)) {
1791 page = pfn_to_page(pfn);
1795 __init_single_page(page, pfn, zid, nid);
1802 * This function is meant to pre-load the iterator for the zone init.
1803 * Specifically it walks through the ranges until we are caught up to the
1804 * first_init_pfn value and exits there. If we never encounter the value we
1805 * return false indicating there are no valid ranges left.
1808 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1809 unsigned long *spfn, unsigned long *epfn,
1810 unsigned long first_init_pfn)
1815 * Start out by walking through the ranges in this zone that have
1816 * already been initialized. We don't need to do anything with them
1817 * so we just need to flush them out of the system.
1819 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1820 if (*epfn <= first_init_pfn)
1822 if (*spfn < first_init_pfn)
1823 *spfn = first_init_pfn;
1832 * Initialize and free pages. We do it in two loops: first we initialize
1833 * struct page, then free to buddy allocator, because while we are
1834 * freeing pages we can access pages that are ahead (computing buddy
1835 * page in __free_one_page()).
1837 * In order to try and keep some memory in the cache we have the loop
1838 * broken along max page order boundaries. This way we will not cause
1839 * any issues with the buddy page computation.
1841 static unsigned long __init
1842 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1843 unsigned long *end_pfn)
1845 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1846 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1847 unsigned long nr_pages = 0;
1850 /* First we loop through and initialize the page values */
1851 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1854 if (mo_pfn <= *start_pfn)
1857 t = min(mo_pfn, *end_pfn);
1858 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1860 if (mo_pfn < *end_pfn) {
1861 *start_pfn = mo_pfn;
1866 /* Reset values and now loop through freeing pages as needed */
1869 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1875 t = min(mo_pfn, epfn);
1876 deferred_free_pages(spfn, t);
1886 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1889 unsigned long spfn, epfn;
1890 struct zone *zone = arg;
1893 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1896 * Initialize and free pages in MAX_ORDER sized increments so that we
1897 * can avoid introducing any issues with the buddy allocator.
1899 while (spfn < end_pfn) {
1900 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1905 /* An arch may override for more concurrency. */
1907 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1912 /* Initialise remaining memory on a node */
1913 static int __init deferred_init_memmap(void *data)
1915 pg_data_t *pgdat = data;
1916 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1917 unsigned long spfn = 0, epfn = 0;
1918 unsigned long first_init_pfn, flags;
1919 unsigned long start = jiffies;
1921 int zid, max_threads;
1924 /* Bind memory initialisation thread to a local node if possible */
1925 if (!cpumask_empty(cpumask))
1926 set_cpus_allowed_ptr(current, cpumask);
1928 pgdat_resize_lock(pgdat, &flags);
1929 first_init_pfn = pgdat->first_deferred_pfn;
1930 if (first_init_pfn == ULONG_MAX) {
1931 pgdat_resize_unlock(pgdat, &flags);
1932 pgdat_init_report_one_done();
1936 /* Sanity check boundaries */
1937 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1938 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1939 pgdat->first_deferred_pfn = ULONG_MAX;
1942 * Once we unlock here, the zone cannot be grown anymore, thus if an
1943 * interrupt thread must allocate this early in boot, zone must be
1944 * pre-grown prior to start of deferred page initialization.
1946 pgdat_resize_unlock(pgdat, &flags);
1948 /* Only the highest zone is deferred so find it */
1949 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1950 zone = pgdat->node_zones + zid;
1951 if (first_init_pfn < zone_end_pfn(zone))
1955 /* If the zone is empty somebody else may have cleared out the zone */
1956 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1960 max_threads = deferred_page_init_max_threads(cpumask);
1962 while (spfn < epfn) {
1963 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1964 struct padata_mt_job job = {
1965 .thread_fn = deferred_init_memmap_chunk,
1968 .size = epfn_align - spfn,
1969 .align = PAGES_PER_SECTION,
1970 .min_chunk = PAGES_PER_SECTION,
1971 .max_threads = max_threads,
1974 padata_do_multithreaded(&job);
1975 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1979 /* Sanity check that the next zone really is unpopulated */
1980 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1982 pr_info("node %d deferred pages initialised in %ums\n",
1983 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1985 pgdat_init_report_one_done();
1990 * If this zone has deferred pages, try to grow it by initializing enough
1991 * deferred pages to satisfy the allocation specified by order, rounded up to
1992 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1993 * of SECTION_SIZE bytes by initializing struct pages in increments of
1994 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1996 * Return true when zone was grown, otherwise return false. We return true even
1997 * when we grow less than requested, to let the caller decide if there are
1998 * enough pages to satisfy the allocation.
2000 * Note: We use noinline because this function is needed only during boot, and
2001 * it is called from a __ref function _deferred_grow_zone. This way we are
2002 * making sure that it is not inlined into permanent text section.
2004 static noinline bool __init
2005 deferred_grow_zone(struct zone *zone, unsigned int order)
2007 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2008 pg_data_t *pgdat = zone->zone_pgdat;
2009 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2010 unsigned long spfn, epfn, flags;
2011 unsigned long nr_pages = 0;
2014 /* Only the last zone may have deferred pages */
2015 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2018 pgdat_resize_lock(pgdat, &flags);
2021 * If someone grew this zone while we were waiting for spinlock, return
2022 * true, as there might be enough pages already.
2024 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2025 pgdat_resize_unlock(pgdat, &flags);
2029 /* If the zone is empty somebody else may have cleared out the zone */
2030 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2031 first_deferred_pfn)) {
2032 pgdat->first_deferred_pfn = ULONG_MAX;
2033 pgdat_resize_unlock(pgdat, &flags);
2034 /* Retry only once. */
2035 return first_deferred_pfn != ULONG_MAX;
2039 * Initialize and free pages in MAX_ORDER sized increments so
2040 * that we can avoid introducing any issues with the buddy
2043 while (spfn < epfn) {
2044 /* update our first deferred PFN for this section */
2045 first_deferred_pfn = spfn;
2047 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2048 touch_nmi_watchdog();
2050 /* We should only stop along section boundaries */
2051 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2054 /* If our quota has been met we can stop here */
2055 if (nr_pages >= nr_pages_needed)
2059 pgdat->first_deferred_pfn = spfn;
2060 pgdat_resize_unlock(pgdat, &flags);
2062 return nr_pages > 0;
2066 * deferred_grow_zone() is __init, but it is called from
2067 * get_page_from_freelist() during early boot until deferred_pages permanently
2068 * disables this call. This is why we have refdata wrapper to avoid warning,
2069 * and to ensure that the function body gets unloaded.
2072 _deferred_grow_zone(struct zone *zone, unsigned int order)
2074 return deferred_grow_zone(zone, order);
2077 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2079 void __init page_alloc_init_late(void)
2084 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2086 /* There will be num_node_state(N_MEMORY) threads */
2087 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2088 for_each_node_state(nid, N_MEMORY) {
2089 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2092 /* Block until all are initialised */
2093 wait_for_completion(&pgdat_init_all_done_comp);
2096 * The number of managed pages has changed due to the initialisation
2097 * so the pcpu batch and high limits needs to be updated or the limits
2098 * will be artificially small.
2100 for_each_populated_zone(zone)
2101 zone_pcp_update(zone);
2104 * We initialized the rest of the deferred pages. Permanently disable
2105 * on-demand struct page initialization.
2107 static_branch_disable(&deferred_pages);
2109 /* Reinit limits that are based on free pages after the kernel is up */
2110 files_maxfiles_init();
2115 /* Discard memblock private memory */
2118 for_each_node_state(nid, N_MEMORY)
2119 shuffle_free_memory(NODE_DATA(nid));
2121 for_each_populated_zone(zone)
2122 set_zone_contiguous(zone);
2126 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2127 void __init init_cma_reserved_pageblock(struct page *page)
2129 unsigned i = pageblock_nr_pages;
2130 struct page *p = page;
2133 __ClearPageReserved(p);
2134 set_page_count(p, 0);
2137 set_pageblock_migratetype(page, MIGRATE_CMA);
2139 if (pageblock_order >= MAX_ORDER) {
2140 i = pageblock_nr_pages;
2143 set_page_refcounted(p);
2144 __free_pages(p, MAX_ORDER - 1);
2145 p += MAX_ORDER_NR_PAGES;
2146 } while (i -= MAX_ORDER_NR_PAGES);
2148 set_page_refcounted(page);
2149 __free_pages(page, pageblock_order);
2152 adjust_managed_page_count(page, pageblock_nr_pages);
2157 * The order of subdivision here is critical for the IO subsystem.
2158 * Please do not alter this order without good reasons and regression
2159 * testing. Specifically, as large blocks of memory are subdivided,
2160 * the order in which smaller blocks are delivered depends on the order
2161 * they're subdivided in this function. This is the primary factor
2162 * influencing the order in which pages are delivered to the IO
2163 * subsystem according to empirical testing, and this is also justified
2164 * by considering the behavior of a buddy system containing a single
2165 * large block of memory acted on by a series of small allocations.
2166 * This behavior is a critical factor in sglist merging's success.
2170 static inline void expand(struct zone *zone, struct page *page,
2171 int low, int high, int migratetype)
2173 unsigned long size = 1 << high;
2175 while (high > low) {
2178 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2181 * Mark as guard pages (or page), that will allow to
2182 * merge back to allocator when buddy will be freed.
2183 * Corresponding page table entries will not be touched,
2184 * pages will stay not present in virtual address space
2186 if (set_page_guard(zone, &page[size], high, migratetype))
2189 add_to_free_list(&page[size], zone, high, migratetype);
2190 set_buddy_order(&page[size], high);
2194 static void check_new_page_bad(struct page *page)
2196 if (unlikely(page->flags & __PG_HWPOISON)) {
2197 /* Don't complain about hwpoisoned pages */
2198 page_mapcount_reset(page); /* remove PageBuddy */
2203 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2207 * This page is about to be returned from the page allocator
2209 static inline int check_new_page(struct page *page)
2211 if (likely(page_expected_state(page,
2212 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2215 check_new_page_bad(page);
2219 static inline bool free_pages_prezeroed(void)
2221 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2222 page_poisoning_enabled()) || want_init_on_free();
2225 #ifdef CONFIG_DEBUG_VM
2227 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2228 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2229 * also checked when pcp lists are refilled from the free lists.
2231 static inline bool check_pcp_refill(struct page *page)
2233 if (debug_pagealloc_enabled_static())
2234 return check_new_page(page);
2239 static inline bool check_new_pcp(struct page *page)
2241 return check_new_page(page);
2245 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2246 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2247 * enabled, they are also checked when being allocated from the pcp lists.
2249 static inline bool check_pcp_refill(struct page *page)
2251 return check_new_page(page);
2253 static inline bool check_new_pcp(struct page *page)
2255 if (debug_pagealloc_enabled_static())
2256 return check_new_page(page);
2260 #endif /* CONFIG_DEBUG_VM */
2262 static bool check_new_pages(struct page *page, unsigned int order)
2265 for (i = 0; i < (1 << order); i++) {
2266 struct page *p = page + i;
2268 if (unlikely(check_new_page(p)))
2275 inline void post_alloc_hook(struct page *page, unsigned int order,
2278 set_page_private(page, 0);
2279 set_page_refcounted(page);
2281 arch_alloc_page(page, order);
2282 debug_pagealloc_map_pages(page, 1 << order);
2283 kasan_alloc_pages(page, order);
2284 kernel_poison_pages(page, 1 << order, 1);
2285 set_page_owner(page, order, gfp_flags);
2287 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2288 kernel_init_free_pages(page, 1 << order);
2291 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2292 unsigned int alloc_flags)
2294 post_alloc_hook(page, order, gfp_flags);
2296 if (order && (gfp_flags & __GFP_COMP))
2297 prep_compound_page(page, order);
2300 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2301 * allocate the page. The expectation is that the caller is taking
2302 * steps that will free more memory. The caller should avoid the page
2303 * being used for !PFMEMALLOC purposes.
2305 if (alloc_flags & ALLOC_NO_WATERMARKS)
2306 set_page_pfmemalloc(page);
2308 clear_page_pfmemalloc(page);
2312 * Go through the free lists for the given migratetype and remove
2313 * the smallest available page from the freelists
2315 static __always_inline
2316 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2319 unsigned int current_order;
2320 struct free_area *area;
2323 /* Find a page of the appropriate size in the preferred list */
2324 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2325 area = &(zone->free_area[current_order]);
2326 page = get_page_from_free_area(area, migratetype);
2329 del_page_from_free_list(page, zone, current_order);
2330 expand(zone, page, order, current_order, migratetype);
2331 set_pcppage_migratetype(page, migratetype);
2340 * This array describes the order lists are fallen back to when
2341 * the free lists for the desirable migrate type are depleted
2343 static int fallbacks[MIGRATE_TYPES][3] = {
2344 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2345 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2346 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2348 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2350 #ifdef CONFIG_MEMORY_ISOLATION
2351 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2356 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2359 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2362 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2363 unsigned int order) { return NULL; }
2367 * Move the free pages in a range to the freelist tail of the requested type.
2368 * Note that start_page and end_pages are not aligned on a pageblock
2369 * boundary. If alignment is required, use move_freepages_block()
2371 static int move_freepages(struct zone *zone,
2372 struct page *start_page, struct page *end_page,
2373 int migratetype, int *num_movable)
2377 int pages_moved = 0;
2379 for (page = start_page; page <= end_page;) {
2380 if (!pfn_valid_within(page_to_pfn(page))) {
2385 if (!PageBuddy(page)) {
2387 * We assume that pages that could be isolated for
2388 * migration are movable. But we don't actually try
2389 * isolating, as that would be expensive.
2392 (PageLRU(page) || __PageMovable(page)))
2399 /* Make sure we are not inadvertently changing nodes */
2400 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2401 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2403 order = buddy_order(page);
2404 move_to_free_list(page, zone, order, migratetype);
2406 pages_moved += 1 << order;
2412 int move_freepages_block(struct zone *zone, struct page *page,
2413 int migratetype, int *num_movable)
2415 unsigned long start_pfn, end_pfn;
2416 struct page *start_page, *end_page;
2421 start_pfn = page_to_pfn(page);
2422 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2423 start_page = pfn_to_page(start_pfn);
2424 end_page = start_page + pageblock_nr_pages - 1;
2425 end_pfn = start_pfn + pageblock_nr_pages - 1;
2427 /* Do not cross zone boundaries */
2428 if (!zone_spans_pfn(zone, start_pfn))
2430 if (!zone_spans_pfn(zone, end_pfn))
2433 return move_freepages(zone, start_page, end_page, migratetype,
2437 static void change_pageblock_range(struct page *pageblock_page,
2438 int start_order, int migratetype)
2440 int nr_pageblocks = 1 << (start_order - pageblock_order);
2442 while (nr_pageblocks--) {
2443 set_pageblock_migratetype(pageblock_page, migratetype);
2444 pageblock_page += pageblock_nr_pages;
2449 * When we are falling back to another migratetype during allocation, try to
2450 * steal extra free pages from the same pageblocks to satisfy further
2451 * allocations, instead of polluting multiple pageblocks.
2453 * If we are stealing a relatively large buddy page, it is likely there will
2454 * be more free pages in the pageblock, so try to steal them all. For
2455 * reclaimable and unmovable allocations, we steal regardless of page size,
2456 * as fragmentation caused by those allocations polluting movable pageblocks
2457 * is worse than movable allocations stealing from unmovable and reclaimable
2460 static bool can_steal_fallback(unsigned int order, int start_mt)
2463 * Leaving this order check is intended, although there is
2464 * relaxed order check in next check. The reason is that
2465 * we can actually steal whole pageblock if this condition met,
2466 * but, below check doesn't guarantee it and that is just heuristic
2467 * so could be changed anytime.
2469 if (order >= pageblock_order)
2472 if (order >= pageblock_order / 2 ||
2473 start_mt == MIGRATE_RECLAIMABLE ||
2474 start_mt == MIGRATE_UNMOVABLE ||
2475 page_group_by_mobility_disabled)
2481 static inline bool boost_watermark(struct zone *zone)
2483 unsigned long max_boost;
2485 if (!watermark_boost_factor)
2488 * Don't bother in zones that are unlikely to produce results.
2489 * On small machines, including kdump capture kernels running
2490 * in a small area, boosting the watermark can cause an out of
2491 * memory situation immediately.
2493 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2496 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2497 watermark_boost_factor, 10000);
2500 * high watermark may be uninitialised if fragmentation occurs
2501 * very early in boot so do not boost. We do not fall
2502 * through and boost by pageblock_nr_pages as failing
2503 * allocations that early means that reclaim is not going
2504 * to help and it may even be impossible to reclaim the
2505 * boosted watermark resulting in a hang.
2510 max_boost = max(pageblock_nr_pages, max_boost);
2512 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2519 * This function implements actual steal behaviour. If order is large enough,
2520 * we can steal whole pageblock. If not, we first move freepages in this
2521 * pageblock to our migratetype and determine how many already-allocated pages
2522 * are there in the pageblock with a compatible migratetype. If at least half
2523 * of pages are free or compatible, we can change migratetype of the pageblock
2524 * itself, so pages freed in the future will be put on the correct free list.
2526 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2527 unsigned int alloc_flags, int start_type, bool whole_block)
2529 unsigned int current_order = buddy_order(page);
2530 int free_pages, movable_pages, alike_pages;
2533 old_block_type = get_pageblock_migratetype(page);
2536 * This can happen due to races and we want to prevent broken
2537 * highatomic accounting.
2539 if (is_migrate_highatomic(old_block_type))
2542 /* Take ownership for orders >= pageblock_order */
2543 if (current_order >= pageblock_order) {
2544 change_pageblock_range(page, current_order, start_type);
2549 * Boost watermarks to increase reclaim pressure to reduce the
2550 * likelihood of future fallbacks. Wake kswapd now as the node
2551 * may be balanced overall and kswapd will not wake naturally.
2553 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2554 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2556 /* We are not allowed to try stealing from the whole block */
2560 free_pages = move_freepages_block(zone, page, start_type,
2563 * Determine how many pages are compatible with our allocation.
2564 * For movable allocation, it's the number of movable pages which
2565 * we just obtained. For other types it's a bit more tricky.
2567 if (start_type == MIGRATE_MOVABLE) {
2568 alike_pages = movable_pages;
2571 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2572 * to MOVABLE pageblock, consider all non-movable pages as
2573 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2574 * vice versa, be conservative since we can't distinguish the
2575 * exact migratetype of non-movable pages.
2577 if (old_block_type == MIGRATE_MOVABLE)
2578 alike_pages = pageblock_nr_pages
2579 - (free_pages + movable_pages);
2584 /* moving whole block can fail due to zone boundary conditions */
2589 * If a sufficient number of pages in the block are either free or of
2590 * comparable migratability as our allocation, claim the whole block.
2592 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2593 page_group_by_mobility_disabled)
2594 set_pageblock_migratetype(page, start_type);
2599 move_to_free_list(page, zone, current_order, start_type);
2603 * Check whether there is a suitable fallback freepage with requested order.
2604 * If only_stealable is true, this function returns fallback_mt only if
2605 * we can steal other freepages all together. This would help to reduce
2606 * fragmentation due to mixed migratetype pages in one pageblock.
2608 int find_suitable_fallback(struct free_area *area, unsigned int order,
2609 int migratetype, bool only_stealable, bool *can_steal)
2614 if (area->nr_free == 0)
2619 fallback_mt = fallbacks[migratetype][i];
2620 if (fallback_mt == MIGRATE_TYPES)
2623 if (free_area_empty(area, fallback_mt))
2626 if (can_steal_fallback(order, migratetype))
2629 if (!only_stealable)
2640 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2641 * there are no empty page blocks that contain a page with a suitable order
2643 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2644 unsigned int alloc_order)
2647 unsigned long max_managed, flags;
2650 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2651 * Check is race-prone but harmless.
2653 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2654 if (zone->nr_reserved_highatomic >= max_managed)
2657 spin_lock_irqsave(&zone->lock, flags);
2659 /* Recheck the nr_reserved_highatomic limit under the lock */
2660 if (zone->nr_reserved_highatomic >= max_managed)
2664 mt = get_pageblock_migratetype(page);
2665 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2666 && !is_migrate_cma(mt)) {
2667 zone->nr_reserved_highatomic += pageblock_nr_pages;
2668 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2669 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2673 spin_unlock_irqrestore(&zone->lock, flags);
2677 * Used when an allocation is about to fail under memory pressure. This
2678 * potentially hurts the reliability of high-order allocations when under
2679 * intense memory pressure but failed atomic allocations should be easier
2680 * to recover from than an OOM.
2682 * If @force is true, try to unreserve a pageblock even though highatomic
2683 * pageblock is exhausted.
2685 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2688 struct zonelist *zonelist = ac->zonelist;
2689 unsigned long flags;
2696 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2699 * Preserve at least one pageblock unless memory pressure
2702 if (!force && zone->nr_reserved_highatomic <=
2706 spin_lock_irqsave(&zone->lock, flags);
2707 for (order = 0; order < MAX_ORDER; order++) {
2708 struct free_area *area = &(zone->free_area[order]);
2710 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2715 * In page freeing path, migratetype change is racy so
2716 * we can counter several free pages in a pageblock
2717 * in this loop althoug we changed the pageblock type
2718 * from highatomic to ac->migratetype. So we should
2719 * adjust the count once.
2721 if (is_migrate_highatomic_page(page)) {
2723 * It should never happen but changes to
2724 * locking could inadvertently allow a per-cpu
2725 * drain to add pages to MIGRATE_HIGHATOMIC
2726 * while unreserving so be safe and watch for
2729 zone->nr_reserved_highatomic -= min(
2731 zone->nr_reserved_highatomic);
2735 * Convert to ac->migratetype and avoid the normal
2736 * pageblock stealing heuristics. Minimally, the caller
2737 * is doing the work and needs the pages. More
2738 * importantly, if the block was always converted to
2739 * MIGRATE_UNMOVABLE or another type then the number
2740 * of pageblocks that cannot be completely freed
2743 set_pageblock_migratetype(page, ac->migratetype);
2744 ret = move_freepages_block(zone, page, ac->migratetype,
2747 spin_unlock_irqrestore(&zone->lock, flags);
2751 spin_unlock_irqrestore(&zone->lock, flags);
2758 * Try finding a free buddy page on the fallback list and put it on the free
2759 * list of requested migratetype, possibly along with other pages from the same
2760 * block, depending on fragmentation avoidance heuristics. Returns true if
2761 * fallback was found so that __rmqueue_smallest() can grab it.
2763 * The use of signed ints for order and current_order is a deliberate
2764 * deviation from the rest of this file, to make the for loop
2765 * condition simpler.
2767 static __always_inline bool
2768 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2769 unsigned int alloc_flags)
2771 struct free_area *area;
2773 int min_order = order;
2779 * Do not steal pages from freelists belonging to other pageblocks
2780 * i.e. orders < pageblock_order. If there are no local zones free,
2781 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2783 if (alloc_flags & ALLOC_NOFRAGMENT)
2784 min_order = pageblock_order;
2787 * Find the largest available free page in the other list. This roughly
2788 * approximates finding the pageblock with the most free pages, which
2789 * would be too costly to do exactly.
2791 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2793 area = &(zone->free_area[current_order]);
2794 fallback_mt = find_suitable_fallback(area, current_order,
2795 start_migratetype, false, &can_steal);
2796 if (fallback_mt == -1)
2800 * We cannot steal all free pages from the pageblock and the
2801 * requested migratetype is movable. In that case it's better to
2802 * steal and split the smallest available page instead of the
2803 * largest available page, because even if the next movable
2804 * allocation falls back into a different pageblock than this
2805 * one, it won't cause permanent fragmentation.
2807 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2808 && current_order > order)
2817 for (current_order = order; current_order < MAX_ORDER;
2819 area = &(zone->free_area[current_order]);
2820 fallback_mt = find_suitable_fallback(area, current_order,
2821 start_migratetype, false, &can_steal);
2822 if (fallback_mt != -1)
2827 * This should not happen - we already found a suitable fallback
2828 * when looking for the largest page.
2830 VM_BUG_ON(current_order == MAX_ORDER);
2833 page = get_page_from_free_area(area, fallback_mt);
2835 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2838 trace_mm_page_alloc_extfrag(page, order, current_order,
2839 start_migratetype, fallback_mt);
2846 * Do the hard work of removing an element from the buddy allocator.
2847 * Call me with the zone->lock already held.
2849 static __always_inline struct page *
2850 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2851 unsigned int alloc_flags)
2857 * Balance movable allocations between regular and CMA areas by
2858 * allocating from CMA when over half of the zone's free memory
2859 * is in the CMA area.
2861 if (alloc_flags & ALLOC_CMA &&
2862 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2863 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2864 page = __rmqueue_cma_fallback(zone, order);
2870 page = __rmqueue_smallest(zone, order, migratetype);
2871 if (unlikely(!page)) {
2872 if (alloc_flags & ALLOC_CMA)
2873 page = __rmqueue_cma_fallback(zone, order);
2875 if (!page && __rmqueue_fallback(zone, order, migratetype,
2880 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2885 * Obtain a specified number of elements from the buddy allocator, all under
2886 * a single hold of the lock, for efficiency. Add them to the supplied list.
2887 * Returns the number of new pages which were placed at *list.
2889 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2890 unsigned long count, struct list_head *list,
2891 int migratetype, unsigned int alloc_flags)
2895 spin_lock(&zone->lock);
2896 for (i = 0; i < count; ++i) {
2897 struct page *page = __rmqueue(zone, order, migratetype,
2899 if (unlikely(page == NULL))
2902 if (unlikely(check_pcp_refill(page)))
2906 * Split buddy pages returned by expand() are received here in
2907 * physical page order. The page is added to the tail of
2908 * caller's list. From the callers perspective, the linked list
2909 * is ordered by page number under some conditions. This is
2910 * useful for IO devices that can forward direction from the
2911 * head, thus also in the physical page order. This is useful
2912 * for IO devices that can merge IO requests if the physical
2913 * pages are ordered properly.
2915 list_add_tail(&page->lru, list);
2917 if (is_migrate_cma(get_pcppage_migratetype(page)))
2918 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2923 * i pages were removed from the buddy list even if some leak due
2924 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2925 * on i. Do not confuse with 'alloced' which is the number of
2926 * pages added to the pcp list.
2928 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2929 spin_unlock(&zone->lock);
2935 * Called from the vmstat counter updater to drain pagesets of this
2936 * currently executing processor on remote nodes after they have
2939 * Note that this function must be called with the thread pinned to
2940 * a single processor.
2942 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2944 unsigned long flags;
2945 int to_drain, batch;
2947 local_irq_save(flags);
2948 batch = READ_ONCE(pcp->batch);
2949 to_drain = min(pcp->count, batch);
2951 free_pcppages_bulk(zone, to_drain, pcp);
2952 local_irq_restore(flags);
2957 * Drain pcplists of the indicated processor and zone.
2959 * The processor must either be the current processor and the
2960 * thread pinned to the current processor or a processor that
2963 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2965 unsigned long flags;
2966 struct per_cpu_pageset *pset;
2967 struct per_cpu_pages *pcp;
2969 local_irq_save(flags);
2970 pset = per_cpu_ptr(zone->pageset, cpu);
2974 free_pcppages_bulk(zone, pcp->count, pcp);
2975 local_irq_restore(flags);
2979 * Drain pcplists of all zones on the indicated processor.
2981 * The processor must either be the current processor and the
2982 * thread pinned to the current processor or a processor that
2985 static void drain_pages(unsigned int cpu)
2989 for_each_populated_zone(zone) {
2990 drain_pages_zone(cpu, zone);
2995 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2997 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2998 * the single zone's pages.
3000 void drain_local_pages(struct zone *zone)
3002 int cpu = smp_processor_id();
3005 drain_pages_zone(cpu, zone);
3010 static void drain_local_pages_wq(struct work_struct *work)
3012 struct pcpu_drain *drain;
3014 drain = container_of(work, struct pcpu_drain, work);
3017 * drain_all_pages doesn't use proper cpu hotplug protection so
3018 * we can race with cpu offline when the WQ can move this from
3019 * a cpu pinned worker to an unbound one. We can operate on a different
3020 * cpu which is allright but we also have to make sure to not move to
3024 drain_local_pages(drain->zone);
3029 * The implementation of drain_all_pages(), exposing an extra parameter to
3030 * drain on all cpus.
3032 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3033 * not empty. The check for non-emptiness can however race with a free to
3034 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3035 * that need the guarantee that every CPU has drained can disable the
3036 * optimizing racy check.
3038 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3043 * Allocate in the BSS so we wont require allocation in
3044 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3046 static cpumask_t cpus_with_pcps;
3049 * Make sure nobody triggers this path before mm_percpu_wq is fully
3052 if (WARN_ON_ONCE(!mm_percpu_wq))
3056 * Do not drain if one is already in progress unless it's specific to
3057 * a zone. Such callers are primarily CMA and memory hotplug and need
3058 * the drain to be complete when the call returns.
3060 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3063 mutex_lock(&pcpu_drain_mutex);
3067 * We don't care about racing with CPU hotplug event
3068 * as offline notification will cause the notified
3069 * cpu to drain that CPU pcps and on_each_cpu_mask
3070 * disables preemption as part of its processing
3072 for_each_online_cpu(cpu) {
3073 struct per_cpu_pageset *pcp;
3075 bool has_pcps = false;
3077 if (force_all_cpus) {
3079 * The pcp.count check is racy, some callers need a
3080 * guarantee that no cpu is missed.
3084 pcp = per_cpu_ptr(zone->pageset, cpu);
3088 for_each_populated_zone(z) {
3089 pcp = per_cpu_ptr(z->pageset, cpu);
3090 if (pcp->pcp.count) {
3098 cpumask_set_cpu(cpu, &cpus_with_pcps);
3100 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3103 for_each_cpu(cpu, &cpus_with_pcps) {
3104 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3107 INIT_WORK(&drain->work, drain_local_pages_wq);
3108 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3110 for_each_cpu(cpu, &cpus_with_pcps)
3111 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3113 mutex_unlock(&pcpu_drain_mutex);
3117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3119 * When zone parameter is non-NULL, spill just the single zone's pages.
3121 * Note that this can be extremely slow as the draining happens in a workqueue.
3123 void drain_all_pages(struct zone *zone)
3125 __drain_all_pages(zone, false);
3128 #ifdef CONFIG_HIBERNATION
3131 * Touch the watchdog for every WD_PAGE_COUNT pages.
3133 #define WD_PAGE_COUNT (128*1024)
3135 void mark_free_pages(struct zone *zone)
3137 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3138 unsigned long flags;
3139 unsigned int order, t;
3142 if (zone_is_empty(zone))
3145 spin_lock_irqsave(&zone->lock, flags);
3147 max_zone_pfn = zone_end_pfn(zone);
3148 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3149 if (pfn_valid(pfn)) {
3150 page = pfn_to_page(pfn);
3152 if (!--page_count) {
3153 touch_nmi_watchdog();
3154 page_count = WD_PAGE_COUNT;
3157 if (page_zone(page) != zone)
3160 if (!swsusp_page_is_forbidden(page))
3161 swsusp_unset_page_free(page);
3164 for_each_migratetype_order(order, t) {
3165 list_for_each_entry(page,
3166 &zone->free_area[order].free_list[t], lru) {
3169 pfn = page_to_pfn(page);
3170 for (i = 0; i < (1UL << order); i++) {
3171 if (!--page_count) {
3172 touch_nmi_watchdog();
3173 page_count = WD_PAGE_COUNT;
3175 swsusp_set_page_free(pfn_to_page(pfn + i));
3179 spin_unlock_irqrestore(&zone->lock, flags);
3181 #endif /* CONFIG_PM */
3183 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3187 if (!free_pcp_prepare(page))
3190 migratetype = get_pfnblock_migratetype(page, pfn);
3191 set_pcppage_migratetype(page, migratetype);
3195 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3197 struct zone *zone = page_zone(page);
3198 struct per_cpu_pages *pcp;
3201 migratetype = get_pcppage_migratetype(page);
3202 __count_vm_event(PGFREE);
3205 * We only track unmovable, reclaimable and movable on pcp lists.
3206 * Free ISOLATE pages back to the allocator because they are being
3207 * offlined but treat HIGHATOMIC as movable pages so we can get those
3208 * areas back if necessary. Otherwise, we may have to free
3209 * excessively into the page allocator
3211 if (migratetype >= MIGRATE_PCPTYPES) {
3212 if (unlikely(is_migrate_isolate(migratetype))) {
3213 free_one_page(zone, page, pfn, 0, migratetype,
3217 migratetype = MIGRATE_MOVABLE;
3220 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3221 list_add(&page->lru, &pcp->lists[migratetype]);
3223 if (pcp->count >= READ_ONCE(pcp->high))
3224 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3228 * Free a 0-order page
3230 void free_unref_page(struct page *page)
3232 unsigned long flags;
3233 unsigned long pfn = page_to_pfn(page);
3235 if (!free_unref_page_prepare(page, pfn))
3238 local_irq_save(flags);
3239 free_unref_page_commit(page, pfn);
3240 local_irq_restore(flags);
3244 * Free a list of 0-order pages
3246 void free_unref_page_list(struct list_head *list)
3248 struct page *page, *next;
3249 unsigned long flags, pfn;
3250 int batch_count = 0;
3252 /* Prepare pages for freeing */
3253 list_for_each_entry_safe(page, next, list, lru) {
3254 pfn = page_to_pfn(page);
3255 if (!free_unref_page_prepare(page, pfn))
3256 list_del(&page->lru);
3257 set_page_private(page, pfn);
3260 local_irq_save(flags);
3261 list_for_each_entry_safe(page, next, list, lru) {
3262 unsigned long pfn = page_private(page);
3264 set_page_private(page, 0);
3265 trace_mm_page_free_batched(page);
3266 free_unref_page_commit(page, pfn);
3269 * Guard against excessive IRQ disabled times when we get
3270 * a large list of pages to free.
3272 if (++batch_count == SWAP_CLUSTER_MAX) {
3273 local_irq_restore(flags);
3275 local_irq_save(flags);
3278 local_irq_restore(flags);
3282 * split_page takes a non-compound higher-order page, and splits it into
3283 * n (1<<order) sub-pages: page[0..n]
3284 * Each sub-page must be freed individually.
3286 * Note: this is probably too low level an operation for use in drivers.
3287 * Please consult with lkml before using this in your driver.
3289 void split_page(struct page *page, unsigned int order)
3293 VM_BUG_ON_PAGE(PageCompound(page), page);
3294 VM_BUG_ON_PAGE(!page_count(page), page);
3296 for (i = 1; i < (1 << order); i++)
3297 set_page_refcounted(page + i);
3298 split_page_owner(page, 1 << order);
3300 EXPORT_SYMBOL_GPL(split_page);
3302 int __isolate_free_page(struct page *page, unsigned int order)
3304 unsigned long watermark;
3308 BUG_ON(!PageBuddy(page));
3310 zone = page_zone(page);
3311 mt = get_pageblock_migratetype(page);
3313 if (!is_migrate_isolate(mt)) {
3315 * Obey watermarks as if the page was being allocated. We can
3316 * emulate a high-order watermark check with a raised order-0
3317 * watermark, because we already know our high-order page
3320 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3321 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3324 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3327 /* Remove page from free list */
3329 del_page_from_free_list(page, zone, order);
3332 * Set the pageblock if the isolated page is at least half of a
3335 if (order >= pageblock_order - 1) {
3336 struct page *endpage = page + (1 << order) - 1;
3337 for (; page < endpage; page += pageblock_nr_pages) {
3338 int mt = get_pageblock_migratetype(page);
3339 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3340 && !is_migrate_highatomic(mt))
3341 set_pageblock_migratetype(page,
3347 return 1UL << order;
3351 * __putback_isolated_page - Return a now-isolated page back where we got it
3352 * @page: Page that was isolated
3353 * @order: Order of the isolated page
3354 * @mt: The page's pageblock's migratetype
3356 * This function is meant to return a page pulled from the free lists via
3357 * __isolate_free_page back to the free lists they were pulled from.
3359 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3361 struct zone *zone = page_zone(page);
3363 /* zone lock should be held when this function is called */
3364 lockdep_assert_held(&zone->lock);
3366 /* Return isolated page to tail of freelist. */
3367 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3368 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3372 * Update NUMA hit/miss statistics
3374 * Must be called with interrupts disabled.
3376 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3379 enum numa_stat_item local_stat = NUMA_LOCAL;
3381 /* skip numa counters update if numa stats is disabled */
3382 if (!static_branch_likely(&vm_numa_stat_key))
3385 if (zone_to_nid(z) != numa_node_id())
3386 local_stat = NUMA_OTHER;
3388 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3389 __inc_numa_state(z, NUMA_HIT);
3391 __inc_numa_state(z, NUMA_MISS);
3392 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3394 __inc_numa_state(z, local_stat);
3398 /* Remove page from the per-cpu list, caller must protect the list */
3399 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3400 unsigned int alloc_flags,
3401 struct per_cpu_pages *pcp,
3402 struct list_head *list)
3407 if (list_empty(list)) {
3408 pcp->count += rmqueue_bulk(zone, 0,
3409 READ_ONCE(pcp->batch), list,
3410 migratetype, alloc_flags);
3411 if (unlikely(list_empty(list)))
3415 page = list_first_entry(list, struct page, lru);
3416 list_del(&page->lru);
3418 } while (check_new_pcp(page));
3423 /* Lock and remove page from the per-cpu list */
3424 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3425 struct zone *zone, gfp_t gfp_flags,
3426 int migratetype, unsigned int alloc_flags)
3428 struct per_cpu_pages *pcp;
3429 struct list_head *list;
3431 unsigned long flags;
3433 local_irq_save(flags);
3434 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3435 list = &pcp->lists[migratetype];
3436 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3438 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3439 zone_statistics(preferred_zone, zone);
3441 local_irq_restore(flags);
3446 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3449 struct page *rmqueue(struct zone *preferred_zone,
3450 struct zone *zone, unsigned int order,
3451 gfp_t gfp_flags, unsigned int alloc_flags,
3454 unsigned long flags;
3457 if (likely(order == 0)) {
3459 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3460 * we need to skip it when CMA area isn't allowed.
3462 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3463 migratetype != MIGRATE_MOVABLE) {
3464 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3465 migratetype, alloc_flags);
3471 * We most definitely don't want callers attempting to
3472 * allocate greater than order-1 page units with __GFP_NOFAIL.
3474 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3475 spin_lock_irqsave(&zone->lock, flags);
3480 * order-0 request can reach here when the pcplist is skipped
3481 * due to non-CMA allocation context. HIGHATOMIC area is
3482 * reserved for high-order atomic allocation, so order-0
3483 * request should skip it.
3485 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3486 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3488 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3491 page = __rmqueue(zone, order, migratetype, alloc_flags);
3492 } while (page && check_new_pages(page, order));
3493 spin_unlock(&zone->lock);
3496 __mod_zone_freepage_state(zone, -(1 << order),
3497 get_pcppage_migratetype(page));
3499 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3500 zone_statistics(preferred_zone, zone);
3501 local_irq_restore(flags);
3504 /* Separate test+clear to avoid unnecessary atomics */
3505 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3506 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3507 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3510 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3514 local_irq_restore(flags);
3518 #ifdef CONFIG_FAIL_PAGE_ALLOC
3521 struct fault_attr attr;
3523 bool ignore_gfp_highmem;
3524 bool ignore_gfp_reclaim;
3526 } fail_page_alloc = {
3527 .attr = FAULT_ATTR_INITIALIZER,
3528 .ignore_gfp_reclaim = true,
3529 .ignore_gfp_highmem = true,
3533 static int __init setup_fail_page_alloc(char *str)
3535 return setup_fault_attr(&fail_page_alloc.attr, str);
3537 __setup("fail_page_alloc=", setup_fail_page_alloc);
3539 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3541 if (order < fail_page_alloc.min_order)
3543 if (gfp_mask & __GFP_NOFAIL)
3545 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3547 if (fail_page_alloc.ignore_gfp_reclaim &&
3548 (gfp_mask & __GFP_DIRECT_RECLAIM))
3551 return should_fail(&fail_page_alloc.attr, 1 << order);
3554 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3556 static int __init fail_page_alloc_debugfs(void)
3558 umode_t mode = S_IFREG | 0600;
3561 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3562 &fail_page_alloc.attr);
3564 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3565 &fail_page_alloc.ignore_gfp_reclaim);
3566 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3567 &fail_page_alloc.ignore_gfp_highmem);
3568 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3573 late_initcall(fail_page_alloc_debugfs);
3575 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3577 #else /* CONFIG_FAIL_PAGE_ALLOC */
3579 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3584 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3586 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3588 return __should_fail_alloc_page(gfp_mask, order);
3590 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3592 static inline long __zone_watermark_unusable_free(struct zone *z,
3593 unsigned int order, unsigned int alloc_flags)
3595 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3596 long unusable_free = (1 << order) - 1;
3599 * If the caller does not have rights to ALLOC_HARDER then subtract
3600 * the high-atomic reserves. This will over-estimate the size of the
3601 * atomic reserve but it avoids a search.
3603 if (likely(!alloc_harder))
3604 unusable_free += z->nr_reserved_highatomic;
3607 /* If allocation can't use CMA areas don't use free CMA pages */
3608 if (!(alloc_flags & ALLOC_CMA))
3609 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3612 return unusable_free;
3616 * Return true if free base pages are above 'mark'. For high-order checks it
3617 * will return true of the order-0 watermark is reached and there is at least
3618 * one free page of a suitable size. Checking now avoids taking the zone lock
3619 * to check in the allocation paths if no pages are free.
3621 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3622 int highest_zoneidx, unsigned int alloc_flags,
3627 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3629 /* free_pages may go negative - that's OK */
3630 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3632 if (alloc_flags & ALLOC_HIGH)
3635 if (unlikely(alloc_harder)) {
3637 * OOM victims can try even harder than normal ALLOC_HARDER
3638 * users on the grounds that it's definitely going to be in
3639 * the exit path shortly and free memory. Any allocation it
3640 * makes during the free path will be small and short-lived.
3642 if (alloc_flags & ALLOC_OOM)
3649 * Check watermarks for an order-0 allocation request. If these
3650 * are not met, then a high-order request also cannot go ahead
3651 * even if a suitable page happened to be free.
3653 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3656 /* If this is an order-0 request then the watermark is fine */
3660 /* For a high-order request, check at least one suitable page is free */
3661 for (o = order; o < MAX_ORDER; o++) {
3662 struct free_area *area = &z->free_area[o];
3668 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3669 if (!free_area_empty(area, mt))
3674 if ((alloc_flags & ALLOC_CMA) &&
3675 !free_area_empty(area, MIGRATE_CMA)) {
3679 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3685 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3686 int highest_zoneidx, unsigned int alloc_flags)
3688 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3689 zone_page_state(z, NR_FREE_PAGES));
3692 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3693 unsigned long mark, int highest_zoneidx,
3694 unsigned int alloc_flags, gfp_t gfp_mask)
3698 free_pages = zone_page_state(z, NR_FREE_PAGES);
3701 * Fast check for order-0 only. If this fails then the reserves
3702 * need to be calculated.
3707 fast_free = free_pages;
3708 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3709 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3713 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3717 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3718 * when checking the min watermark. The min watermark is the
3719 * point where boosting is ignored so that kswapd is woken up
3720 * when below the low watermark.
3722 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3723 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3724 mark = z->_watermark[WMARK_MIN];
3725 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3726 alloc_flags, free_pages);
3732 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3733 unsigned long mark, int highest_zoneidx)
3735 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3737 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3738 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3740 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3745 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3747 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3748 node_reclaim_distance;
3750 #else /* CONFIG_NUMA */
3751 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3755 #endif /* CONFIG_NUMA */
3758 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3759 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3760 * premature use of a lower zone may cause lowmem pressure problems that
3761 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3762 * probably too small. It only makes sense to spread allocations to avoid
3763 * fragmentation between the Normal and DMA32 zones.
3765 static inline unsigned int
3766 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3768 unsigned int alloc_flags;
3771 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3774 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3776 #ifdef CONFIG_ZONE_DMA32
3780 if (zone_idx(zone) != ZONE_NORMAL)
3784 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3785 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3786 * on UMA that if Normal is populated then so is DMA32.
3788 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3789 if (nr_online_nodes > 1 && !populated_zone(--zone))
3792 alloc_flags |= ALLOC_NOFRAGMENT;
3793 #endif /* CONFIG_ZONE_DMA32 */
3797 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3798 unsigned int alloc_flags)
3801 unsigned int pflags = current->flags;
3803 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3804 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3805 alloc_flags |= ALLOC_CMA;
3812 * get_page_from_freelist goes through the zonelist trying to allocate
3815 static struct page *
3816 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3817 const struct alloc_context *ac)
3821 struct pglist_data *last_pgdat_dirty_limit = NULL;
3826 * Scan zonelist, looking for a zone with enough free.
3827 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3829 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3830 z = ac->preferred_zoneref;
3831 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3836 if (cpusets_enabled() &&
3837 (alloc_flags & ALLOC_CPUSET) &&
3838 !__cpuset_zone_allowed(zone, gfp_mask))
3841 * When allocating a page cache page for writing, we
3842 * want to get it from a node that is within its dirty
3843 * limit, such that no single node holds more than its
3844 * proportional share of globally allowed dirty pages.
3845 * The dirty limits take into account the node's
3846 * lowmem reserves and high watermark so that kswapd
3847 * should be able to balance it without having to
3848 * write pages from its LRU list.
3850 * XXX: For now, allow allocations to potentially
3851 * exceed the per-node dirty limit in the slowpath
3852 * (spread_dirty_pages unset) before going into reclaim,
3853 * which is important when on a NUMA setup the allowed
3854 * nodes are together not big enough to reach the
3855 * global limit. The proper fix for these situations
3856 * will require awareness of nodes in the
3857 * dirty-throttling and the flusher threads.
3859 if (ac->spread_dirty_pages) {
3860 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3863 if (!node_dirty_ok(zone->zone_pgdat)) {
3864 last_pgdat_dirty_limit = zone->zone_pgdat;
3869 if (no_fallback && nr_online_nodes > 1 &&
3870 zone != ac->preferred_zoneref->zone) {
3874 * If moving to a remote node, retry but allow
3875 * fragmenting fallbacks. Locality is more important
3876 * than fragmentation avoidance.
3878 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3879 if (zone_to_nid(zone) != local_nid) {
3880 alloc_flags &= ~ALLOC_NOFRAGMENT;
3885 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3886 if (!zone_watermark_fast(zone, order, mark,
3887 ac->highest_zoneidx, alloc_flags,
3891 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3893 * Watermark failed for this zone, but see if we can
3894 * grow this zone if it contains deferred pages.
3896 if (static_branch_unlikely(&deferred_pages)) {
3897 if (_deferred_grow_zone(zone, order))
3901 /* Checked here to keep the fast path fast */
3902 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3903 if (alloc_flags & ALLOC_NO_WATERMARKS)
3906 if (node_reclaim_mode == 0 ||
3907 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3910 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3912 case NODE_RECLAIM_NOSCAN:
3915 case NODE_RECLAIM_FULL:
3916 /* scanned but unreclaimable */
3919 /* did we reclaim enough */
3920 if (zone_watermark_ok(zone, order, mark,
3921 ac->highest_zoneidx, alloc_flags))
3929 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3930 gfp_mask, alloc_flags, ac->migratetype);
3932 prep_new_page(page, order, gfp_mask, alloc_flags);
3935 * If this is a high-order atomic allocation then check
3936 * if the pageblock should be reserved for the future
3938 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3939 reserve_highatomic_pageblock(page, zone, order);
3943 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3944 /* Try again if zone has deferred pages */
3945 if (static_branch_unlikely(&deferred_pages)) {
3946 if (_deferred_grow_zone(zone, order))
3954 * It's possible on a UMA machine to get through all zones that are
3955 * fragmented. If avoiding fragmentation, reset and try again.
3958 alloc_flags &= ~ALLOC_NOFRAGMENT;
3965 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3967 unsigned int filter = SHOW_MEM_FILTER_NODES;
3970 * This documents exceptions given to allocations in certain
3971 * contexts that are allowed to allocate outside current's set
3974 if (!(gfp_mask & __GFP_NOMEMALLOC))
3975 if (tsk_is_oom_victim(current) ||
3976 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3977 filter &= ~SHOW_MEM_FILTER_NODES;
3978 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3979 filter &= ~SHOW_MEM_FILTER_NODES;
3981 show_mem(filter, nodemask);
3984 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3986 struct va_format vaf;
3988 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3990 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3993 va_start(args, fmt);
3996 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3997 current->comm, &vaf, gfp_mask, &gfp_mask,
3998 nodemask_pr_args(nodemask));
4001 cpuset_print_current_mems_allowed();
4004 warn_alloc_show_mem(gfp_mask, nodemask);
4007 static inline struct page *
4008 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4009 unsigned int alloc_flags,
4010 const struct alloc_context *ac)
4014 page = get_page_from_freelist(gfp_mask, order,
4015 alloc_flags|ALLOC_CPUSET, ac);
4017 * fallback to ignore cpuset restriction if our nodes
4021 page = get_page_from_freelist(gfp_mask, order,
4027 static inline struct page *
4028 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4029 const struct alloc_context *ac, unsigned long *did_some_progress)
4031 struct oom_control oc = {
4032 .zonelist = ac->zonelist,
4033 .nodemask = ac->nodemask,
4035 .gfp_mask = gfp_mask,
4040 *did_some_progress = 0;
4043 * Acquire the oom lock. If that fails, somebody else is
4044 * making progress for us.
4046 if (!mutex_trylock(&oom_lock)) {
4047 *did_some_progress = 1;
4048 schedule_timeout_uninterruptible(1);
4053 * Go through the zonelist yet one more time, keep very high watermark
4054 * here, this is only to catch a parallel oom killing, we must fail if
4055 * we're still under heavy pressure. But make sure that this reclaim
4056 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4057 * allocation which will never fail due to oom_lock already held.
4059 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4060 ~__GFP_DIRECT_RECLAIM, order,
4061 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4065 /* Coredumps can quickly deplete all memory reserves */
4066 if (current->flags & PF_DUMPCORE)
4068 /* The OOM killer will not help higher order allocs */
4069 if (order > PAGE_ALLOC_COSTLY_ORDER)
4072 * We have already exhausted all our reclaim opportunities without any
4073 * success so it is time to admit defeat. We will skip the OOM killer
4074 * because it is very likely that the caller has a more reasonable
4075 * fallback than shooting a random task.
4077 * The OOM killer may not free memory on a specific node.
4079 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4081 /* The OOM killer does not needlessly kill tasks for lowmem */
4082 if (ac->highest_zoneidx < ZONE_NORMAL)
4084 if (pm_suspended_storage())
4087 * XXX: GFP_NOFS allocations should rather fail than rely on
4088 * other request to make a forward progress.
4089 * We are in an unfortunate situation where out_of_memory cannot
4090 * do much for this context but let's try it to at least get
4091 * access to memory reserved if the current task is killed (see
4092 * out_of_memory). Once filesystems are ready to handle allocation
4093 * failures more gracefully we should just bail out here.
4096 /* Exhausted what can be done so it's blame time */
4097 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4098 *did_some_progress = 1;
4101 * Help non-failing allocations by giving them access to memory
4104 if (gfp_mask & __GFP_NOFAIL)
4105 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4106 ALLOC_NO_WATERMARKS, ac);
4109 mutex_unlock(&oom_lock);
4114 * Maximum number of compaction retries wit a progress before OOM
4115 * killer is consider as the only way to move forward.
4117 #define MAX_COMPACT_RETRIES 16
4119 #ifdef CONFIG_COMPACTION
4120 /* Try memory compaction for high-order allocations before reclaim */
4121 static struct page *
4122 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4123 unsigned int alloc_flags, const struct alloc_context *ac,
4124 enum compact_priority prio, enum compact_result *compact_result)
4126 struct page *page = NULL;
4127 unsigned long pflags;
4128 unsigned int noreclaim_flag;
4133 psi_memstall_enter(&pflags);
4134 noreclaim_flag = memalloc_noreclaim_save();
4136 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4139 memalloc_noreclaim_restore(noreclaim_flag);
4140 psi_memstall_leave(&pflags);
4143 * At least in one zone compaction wasn't deferred or skipped, so let's
4144 * count a compaction stall
4146 count_vm_event(COMPACTSTALL);
4148 /* Prep a captured page if available */
4150 prep_new_page(page, order, gfp_mask, alloc_flags);
4152 /* Try get a page from the freelist if available */
4154 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4157 struct zone *zone = page_zone(page);
4159 zone->compact_blockskip_flush = false;
4160 compaction_defer_reset(zone, order, true);
4161 count_vm_event(COMPACTSUCCESS);
4166 * It's bad if compaction run occurs and fails. The most likely reason
4167 * is that pages exist, but not enough to satisfy watermarks.
4169 count_vm_event(COMPACTFAIL);
4177 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4178 enum compact_result compact_result,
4179 enum compact_priority *compact_priority,
4180 int *compaction_retries)
4182 int max_retries = MAX_COMPACT_RETRIES;
4185 int retries = *compaction_retries;
4186 enum compact_priority priority = *compact_priority;
4191 if (compaction_made_progress(compact_result))
4192 (*compaction_retries)++;
4195 * compaction considers all the zone as desperately out of memory
4196 * so it doesn't really make much sense to retry except when the
4197 * failure could be caused by insufficient priority
4199 if (compaction_failed(compact_result))
4200 goto check_priority;
4203 * compaction was skipped because there are not enough order-0 pages
4204 * to work with, so we retry only if it looks like reclaim can help.
4206 if (compaction_needs_reclaim(compact_result)) {
4207 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4212 * make sure the compaction wasn't deferred or didn't bail out early
4213 * due to locks contention before we declare that we should give up.
4214 * But the next retry should use a higher priority if allowed, so
4215 * we don't just keep bailing out endlessly.
4217 if (compaction_withdrawn(compact_result)) {
4218 goto check_priority;
4222 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4223 * costly ones because they are de facto nofail and invoke OOM
4224 * killer to move on while costly can fail and users are ready
4225 * to cope with that. 1/4 retries is rather arbitrary but we
4226 * would need much more detailed feedback from compaction to
4227 * make a better decision.
4229 if (order > PAGE_ALLOC_COSTLY_ORDER)
4231 if (*compaction_retries <= max_retries) {
4237 * Make sure there are attempts at the highest priority if we exhausted
4238 * all retries or failed at the lower priorities.
4241 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4242 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4244 if (*compact_priority > min_priority) {
4245 (*compact_priority)--;
4246 *compaction_retries = 0;
4250 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4254 static inline struct page *
4255 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4256 unsigned int alloc_flags, const struct alloc_context *ac,
4257 enum compact_priority prio, enum compact_result *compact_result)
4259 *compact_result = COMPACT_SKIPPED;
4264 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4265 enum compact_result compact_result,
4266 enum compact_priority *compact_priority,
4267 int *compaction_retries)
4272 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4276 * There are setups with compaction disabled which would prefer to loop
4277 * inside the allocator rather than hit the oom killer prematurely.
4278 * Let's give them a good hope and keep retrying while the order-0
4279 * watermarks are OK.
4281 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4282 ac->highest_zoneidx, ac->nodemask) {
4283 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4284 ac->highest_zoneidx, alloc_flags))
4289 #endif /* CONFIG_COMPACTION */
4291 #ifdef CONFIG_LOCKDEP
4292 static struct lockdep_map __fs_reclaim_map =
4293 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4295 static bool __need_reclaim(gfp_t gfp_mask)
4297 /* no reclaim without waiting on it */
4298 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4301 /* this guy won't enter reclaim */
4302 if (current->flags & PF_MEMALLOC)
4305 if (gfp_mask & __GFP_NOLOCKDEP)
4311 void __fs_reclaim_acquire(void)
4313 lock_map_acquire(&__fs_reclaim_map);
4316 void __fs_reclaim_release(void)
4318 lock_map_release(&__fs_reclaim_map);
4321 void fs_reclaim_acquire(gfp_t gfp_mask)
4323 gfp_mask = current_gfp_context(gfp_mask);
4325 if (__need_reclaim(gfp_mask)) {
4326 if (gfp_mask & __GFP_FS)
4327 __fs_reclaim_acquire();
4329 #ifdef CONFIG_MMU_NOTIFIER
4330 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4331 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4336 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4338 void fs_reclaim_release(gfp_t gfp_mask)
4340 gfp_mask = current_gfp_context(gfp_mask);
4342 if (__need_reclaim(gfp_mask)) {
4343 if (gfp_mask & __GFP_FS)
4344 __fs_reclaim_release();
4347 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4350 /* Perform direct synchronous page reclaim */
4351 static unsigned long
4352 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4353 const struct alloc_context *ac)
4355 unsigned int noreclaim_flag;
4356 unsigned long pflags, progress;
4360 /* We now go into synchronous reclaim */
4361 cpuset_memory_pressure_bump();
4362 psi_memstall_enter(&pflags);
4363 fs_reclaim_acquire(gfp_mask);
4364 noreclaim_flag = memalloc_noreclaim_save();
4366 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4369 memalloc_noreclaim_restore(noreclaim_flag);
4370 fs_reclaim_release(gfp_mask);
4371 psi_memstall_leave(&pflags);
4378 /* The really slow allocator path where we enter direct reclaim */
4379 static inline struct page *
4380 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4381 unsigned int alloc_flags, const struct alloc_context *ac,
4382 unsigned long *did_some_progress)
4384 struct page *page = NULL;
4385 bool drained = false;
4387 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4388 if (unlikely(!(*did_some_progress)))
4392 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4395 * If an allocation failed after direct reclaim, it could be because
4396 * pages are pinned on the per-cpu lists or in high alloc reserves.
4397 * Shrink them and try again
4399 if (!page && !drained) {
4400 unreserve_highatomic_pageblock(ac, false);
4401 drain_all_pages(NULL);
4409 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4410 const struct alloc_context *ac)
4414 pg_data_t *last_pgdat = NULL;
4415 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4417 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4419 if (last_pgdat != zone->zone_pgdat)
4420 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4421 last_pgdat = zone->zone_pgdat;
4425 static inline unsigned int
4426 gfp_to_alloc_flags(gfp_t gfp_mask)
4428 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4431 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4432 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4433 * to save two branches.
4435 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4436 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4439 * The caller may dip into page reserves a bit more if the caller
4440 * cannot run direct reclaim, or if the caller has realtime scheduling
4441 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4442 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4444 alloc_flags |= (__force int)
4445 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4447 if (gfp_mask & __GFP_ATOMIC) {
4449 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4450 * if it can't schedule.
4452 if (!(gfp_mask & __GFP_NOMEMALLOC))
4453 alloc_flags |= ALLOC_HARDER;
4455 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4456 * comment for __cpuset_node_allowed().
4458 alloc_flags &= ~ALLOC_CPUSET;
4459 } else if (unlikely(rt_task(current)) && !in_interrupt())
4460 alloc_flags |= ALLOC_HARDER;
4462 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4467 static bool oom_reserves_allowed(struct task_struct *tsk)
4469 if (!tsk_is_oom_victim(tsk))
4473 * !MMU doesn't have oom reaper so give access to memory reserves
4474 * only to the thread with TIF_MEMDIE set
4476 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4483 * Distinguish requests which really need access to full memory
4484 * reserves from oom victims which can live with a portion of it
4486 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4488 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4490 if (gfp_mask & __GFP_MEMALLOC)
4491 return ALLOC_NO_WATERMARKS;
4492 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4493 return ALLOC_NO_WATERMARKS;
4494 if (!in_interrupt()) {
4495 if (current->flags & PF_MEMALLOC)
4496 return ALLOC_NO_WATERMARKS;
4497 else if (oom_reserves_allowed(current))
4504 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4506 return !!__gfp_pfmemalloc_flags(gfp_mask);
4510 * Checks whether it makes sense to retry the reclaim to make a forward progress
4511 * for the given allocation request.
4513 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4514 * without success, or when we couldn't even meet the watermark if we
4515 * reclaimed all remaining pages on the LRU lists.
4517 * Returns true if a retry is viable or false to enter the oom path.
4520 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4521 struct alloc_context *ac, int alloc_flags,
4522 bool did_some_progress, int *no_progress_loops)
4529 * Costly allocations might have made a progress but this doesn't mean
4530 * their order will become available due to high fragmentation so
4531 * always increment the no progress counter for them
4533 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4534 *no_progress_loops = 0;
4536 (*no_progress_loops)++;
4539 * Make sure we converge to OOM if we cannot make any progress
4540 * several times in the row.
4542 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4543 /* Before OOM, exhaust highatomic_reserve */
4544 return unreserve_highatomic_pageblock(ac, true);
4548 * Keep reclaiming pages while there is a chance this will lead
4549 * somewhere. If none of the target zones can satisfy our allocation
4550 * request even if all reclaimable pages are considered then we are
4551 * screwed and have to go OOM.
4553 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4554 ac->highest_zoneidx, ac->nodemask) {
4555 unsigned long available;
4556 unsigned long reclaimable;
4557 unsigned long min_wmark = min_wmark_pages(zone);
4560 available = reclaimable = zone_reclaimable_pages(zone);
4561 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4564 * Would the allocation succeed if we reclaimed all
4565 * reclaimable pages?
4567 wmark = __zone_watermark_ok(zone, order, min_wmark,
4568 ac->highest_zoneidx, alloc_flags, available);
4569 trace_reclaim_retry_zone(z, order, reclaimable,
4570 available, min_wmark, *no_progress_loops, wmark);
4573 * If we didn't make any progress and have a lot of
4574 * dirty + writeback pages then we should wait for
4575 * an IO to complete to slow down the reclaim and
4576 * prevent from pre mature OOM
4578 if (!did_some_progress) {
4579 unsigned long write_pending;
4581 write_pending = zone_page_state_snapshot(zone,
4582 NR_ZONE_WRITE_PENDING);
4584 if (2 * write_pending > reclaimable) {
4585 congestion_wait(BLK_RW_ASYNC, HZ/10);
4597 * Memory allocation/reclaim might be called from a WQ context and the
4598 * current implementation of the WQ concurrency control doesn't
4599 * recognize that a particular WQ is congested if the worker thread is
4600 * looping without ever sleeping. Therefore we have to do a short sleep
4601 * here rather than calling cond_resched().
4603 if (current->flags & PF_WQ_WORKER)
4604 schedule_timeout_uninterruptible(1);
4611 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4614 * It's possible that cpuset's mems_allowed and the nodemask from
4615 * mempolicy don't intersect. This should be normally dealt with by
4616 * policy_nodemask(), but it's possible to race with cpuset update in
4617 * such a way the check therein was true, and then it became false
4618 * before we got our cpuset_mems_cookie here.
4619 * This assumes that for all allocations, ac->nodemask can come only
4620 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4621 * when it does not intersect with the cpuset restrictions) or the
4622 * caller can deal with a violated nodemask.
4624 if (cpusets_enabled() && ac->nodemask &&
4625 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4626 ac->nodemask = NULL;
4631 * When updating a task's mems_allowed or mempolicy nodemask, it is
4632 * possible to race with parallel threads in such a way that our
4633 * allocation can fail while the mask is being updated. If we are about
4634 * to fail, check if the cpuset changed during allocation and if so,
4637 if (read_mems_allowed_retry(cpuset_mems_cookie))
4643 static inline struct page *
4644 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4645 struct alloc_context *ac)
4647 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4648 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4649 struct page *page = NULL;
4650 unsigned int alloc_flags;
4651 unsigned long did_some_progress;
4652 enum compact_priority compact_priority;
4653 enum compact_result compact_result;
4654 int compaction_retries;
4655 int no_progress_loops;
4656 unsigned int cpuset_mems_cookie;
4660 * We also sanity check to catch abuse of atomic reserves being used by
4661 * callers that are not in atomic context.
4663 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4664 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4665 gfp_mask &= ~__GFP_ATOMIC;
4668 compaction_retries = 0;
4669 no_progress_loops = 0;
4670 compact_priority = DEF_COMPACT_PRIORITY;
4671 cpuset_mems_cookie = read_mems_allowed_begin();
4674 * The fast path uses conservative alloc_flags to succeed only until
4675 * kswapd needs to be woken up, and to avoid the cost of setting up
4676 * alloc_flags precisely. So we do that now.
4678 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4681 * We need to recalculate the starting point for the zonelist iterator
4682 * because we might have used different nodemask in the fast path, or
4683 * there was a cpuset modification and we are retrying - otherwise we
4684 * could end up iterating over non-eligible zones endlessly.
4686 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4687 ac->highest_zoneidx, ac->nodemask);
4688 if (!ac->preferred_zoneref->zone)
4691 if (alloc_flags & ALLOC_KSWAPD)
4692 wake_all_kswapds(order, gfp_mask, ac);
4695 * The adjusted alloc_flags might result in immediate success, so try
4698 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4703 * For costly allocations, try direct compaction first, as it's likely
4704 * that we have enough base pages and don't need to reclaim. For non-
4705 * movable high-order allocations, do that as well, as compaction will
4706 * try prevent permanent fragmentation by migrating from blocks of the
4708 * Don't try this for allocations that are allowed to ignore
4709 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4711 if (can_direct_reclaim &&
4713 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4714 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4715 page = __alloc_pages_direct_compact(gfp_mask, order,
4717 INIT_COMPACT_PRIORITY,
4723 * Checks for costly allocations with __GFP_NORETRY, which
4724 * includes some THP page fault allocations
4726 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4728 * If allocating entire pageblock(s) and compaction
4729 * failed because all zones are below low watermarks
4730 * or is prohibited because it recently failed at this
4731 * order, fail immediately unless the allocator has
4732 * requested compaction and reclaim retry.
4735 * - potentially very expensive because zones are far
4736 * below their low watermarks or this is part of very
4737 * bursty high order allocations,
4738 * - not guaranteed to help because isolate_freepages()
4739 * may not iterate over freed pages as part of its
4741 * - unlikely to make entire pageblocks free on its
4744 if (compact_result == COMPACT_SKIPPED ||
4745 compact_result == COMPACT_DEFERRED)
4749 * Looks like reclaim/compaction is worth trying, but
4750 * sync compaction could be very expensive, so keep
4751 * using async compaction.
4753 compact_priority = INIT_COMPACT_PRIORITY;
4758 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4759 if (alloc_flags & ALLOC_KSWAPD)
4760 wake_all_kswapds(order, gfp_mask, ac);
4762 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4764 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4767 * Reset the nodemask and zonelist iterators if memory policies can be
4768 * ignored. These allocations are high priority and system rather than
4771 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4772 ac->nodemask = NULL;
4773 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4774 ac->highest_zoneidx, ac->nodemask);
4777 /* Attempt with potentially adjusted zonelist and alloc_flags */
4778 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4782 /* Caller is not willing to reclaim, we can't balance anything */
4783 if (!can_direct_reclaim)
4786 /* Avoid recursion of direct reclaim */
4787 if (current->flags & PF_MEMALLOC)
4790 /* Try direct reclaim and then allocating */
4791 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4792 &did_some_progress);
4796 /* Try direct compaction and then allocating */
4797 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4798 compact_priority, &compact_result);
4802 /* Do not loop if specifically requested */
4803 if (gfp_mask & __GFP_NORETRY)
4807 * Do not retry costly high order allocations unless they are
4808 * __GFP_RETRY_MAYFAIL
4810 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4813 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4814 did_some_progress > 0, &no_progress_loops))
4818 * It doesn't make any sense to retry for the compaction if the order-0
4819 * reclaim is not able to make any progress because the current
4820 * implementation of the compaction depends on the sufficient amount
4821 * of free memory (see __compaction_suitable)
4823 if (did_some_progress > 0 &&
4824 should_compact_retry(ac, order, alloc_flags,
4825 compact_result, &compact_priority,
4826 &compaction_retries))
4830 /* Deal with possible cpuset update races before we start OOM killing */
4831 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4834 /* Reclaim has failed us, start killing things */
4835 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4839 /* Avoid allocations with no watermarks from looping endlessly */
4840 if (tsk_is_oom_victim(current) &&
4841 (alloc_flags & ALLOC_OOM ||
4842 (gfp_mask & __GFP_NOMEMALLOC)))
4845 /* Retry as long as the OOM killer is making progress */
4846 if (did_some_progress) {
4847 no_progress_loops = 0;
4852 /* Deal with possible cpuset update races before we fail */
4853 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4857 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4860 if (gfp_mask & __GFP_NOFAIL) {
4862 * All existing users of the __GFP_NOFAIL are blockable, so warn
4863 * of any new users that actually require GFP_NOWAIT
4865 if (WARN_ON_ONCE(!can_direct_reclaim))
4869 * PF_MEMALLOC request from this context is rather bizarre
4870 * because we cannot reclaim anything and only can loop waiting
4871 * for somebody to do a work for us
4873 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4876 * non failing costly orders are a hard requirement which we
4877 * are not prepared for much so let's warn about these users
4878 * so that we can identify them and convert them to something
4881 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4884 * Help non-failing allocations by giving them access to memory
4885 * reserves but do not use ALLOC_NO_WATERMARKS because this
4886 * could deplete whole memory reserves which would just make
4887 * the situation worse
4889 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4897 warn_alloc(gfp_mask, ac->nodemask,
4898 "page allocation failure: order:%u", order);
4903 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4904 int preferred_nid, nodemask_t *nodemask,
4905 struct alloc_context *ac, gfp_t *alloc_mask,
4906 unsigned int *alloc_flags)
4908 ac->highest_zoneidx = gfp_zone(gfp_mask);
4909 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4910 ac->nodemask = nodemask;
4911 ac->migratetype = gfp_migratetype(gfp_mask);
4913 if (cpusets_enabled()) {
4914 *alloc_mask |= __GFP_HARDWALL;
4916 * When we are in the interrupt context, it is irrelevant
4917 * to the current task context. It means that any node ok.
4919 if (!in_interrupt() && !ac->nodemask)
4920 ac->nodemask = &cpuset_current_mems_allowed;
4922 *alloc_flags |= ALLOC_CPUSET;
4925 fs_reclaim_acquire(gfp_mask);
4926 fs_reclaim_release(gfp_mask);
4928 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4930 if (should_fail_alloc_page(gfp_mask, order))
4933 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4935 /* Dirty zone balancing only done in the fast path */
4936 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4939 * The preferred zone is used for statistics but crucially it is
4940 * also used as the starting point for the zonelist iterator. It
4941 * may get reset for allocations that ignore memory policies.
4943 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4944 ac->highest_zoneidx, ac->nodemask);
4950 * This is the 'heart' of the zoned buddy allocator.
4953 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4954 nodemask_t *nodemask)
4957 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4958 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4959 struct alloc_context ac = { };
4962 * There are several places where we assume that the order value is sane
4963 * so bail out early if the request is out of bound.
4965 if (unlikely(order >= MAX_ORDER)) {
4966 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4970 gfp_mask &= gfp_allowed_mask;
4971 alloc_mask = gfp_mask;
4972 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4976 * Forbid the first pass from falling back to types that fragment
4977 * memory until all local zones are considered.
4979 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4981 /* First allocation attempt */
4982 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4987 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4988 * resp. GFP_NOIO which has to be inherited for all allocation requests
4989 * from a particular context which has been marked by
4990 * memalloc_no{fs,io}_{save,restore}.
4992 alloc_mask = current_gfp_context(gfp_mask);
4993 ac.spread_dirty_pages = false;
4996 * Restore the original nodemask if it was potentially replaced with
4997 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4999 ac.nodemask = nodemask;
5001 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5004 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5005 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5006 __free_pages(page, order);
5010 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5014 EXPORT_SYMBOL(__alloc_pages_nodemask);
5017 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5018 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5019 * you need to access high mem.
5021 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5025 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5028 return (unsigned long) page_address(page);
5030 EXPORT_SYMBOL(__get_free_pages);
5032 unsigned long get_zeroed_page(gfp_t gfp_mask)
5034 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5036 EXPORT_SYMBOL(get_zeroed_page);
5038 static inline void free_the_page(struct page *page, unsigned int order)
5040 if (order == 0) /* Via pcp? */
5041 free_unref_page(page);
5043 __free_pages_ok(page, order, FPI_NONE);
5047 * __free_pages - Free pages allocated with alloc_pages().
5048 * @page: The page pointer returned from alloc_pages().
5049 * @order: The order of the allocation.
5051 * This function can free multi-page allocations that are not compound
5052 * pages. It does not check that the @order passed in matches that of
5053 * the allocation, so it is easy to leak memory. Freeing more memory
5054 * than was allocated will probably emit a warning.
5056 * If the last reference to this page is speculative, it will be released
5057 * by put_page() which only frees the first page of a non-compound
5058 * allocation. To prevent the remaining pages from being leaked, we free
5059 * the subsequent pages here. If you want to use the page's reference
5060 * count to decide when to free the allocation, you should allocate a
5061 * compound page, and use put_page() instead of __free_pages().
5063 * Context: May be called in interrupt context or while holding a normal
5064 * spinlock, but not in NMI context or while holding a raw spinlock.
5066 void __free_pages(struct page *page, unsigned int order)
5068 if (put_page_testzero(page))
5069 free_the_page(page, order);
5070 else if (!PageHead(page))
5072 free_the_page(page + (1 << order), order);
5074 EXPORT_SYMBOL(__free_pages);
5076 void free_pages(unsigned long addr, unsigned int order)
5079 VM_BUG_ON(!virt_addr_valid((void *)addr));
5080 __free_pages(virt_to_page((void *)addr), order);
5084 EXPORT_SYMBOL(free_pages);
5088 * An arbitrary-length arbitrary-offset area of memory which resides
5089 * within a 0 or higher order page. Multiple fragments within that page
5090 * are individually refcounted, in the page's reference counter.
5092 * The page_frag functions below provide a simple allocation framework for
5093 * page fragments. This is used by the network stack and network device
5094 * drivers to provide a backing region of memory for use as either an
5095 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5097 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5100 struct page *page = NULL;
5101 gfp_t gfp = gfp_mask;
5103 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5104 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5106 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5107 PAGE_FRAG_CACHE_MAX_ORDER);
5108 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5110 if (unlikely(!page))
5111 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5113 nc->va = page ? page_address(page) : NULL;
5118 void __page_frag_cache_drain(struct page *page, unsigned int count)
5120 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5122 if (page_ref_sub_and_test(page, count))
5123 free_the_page(page, compound_order(page));
5125 EXPORT_SYMBOL(__page_frag_cache_drain);
5127 void *page_frag_alloc(struct page_frag_cache *nc,
5128 unsigned int fragsz, gfp_t gfp_mask)
5130 unsigned int size = PAGE_SIZE;
5134 if (unlikely(!nc->va)) {
5136 page = __page_frag_cache_refill(nc, gfp_mask);
5140 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5141 /* if size can vary use size else just use PAGE_SIZE */
5144 /* Even if we own the page, we do not use atomic_set().
5145 * This would break get_page_unless_zero() users.
5147 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5149 /* reset page count bias and offset to start of new frag */
5150 nc->pfmemalloc = page_is_pfmemalloc(page);
5151 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5155 offset = nc->offset - fragsz;
5156 if (unlikely(offset < 0)) {
5157 page = virt_to_page(nc->va);
5159 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5162 if (unlikely(nc->pfmemalloc)) {
5163 free_the_page(page, compound_order(page));
5167 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5168 /* if size can vary use size else just use PAGE_SIZE */
5171 /* OK, page count is 0, we can safely set it */
5172 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5174 /* reset page count bias and offset to start of new frag */
5175 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5176 offset = size - fragsz;
5180 nc->offset = offset;
5182 return nc->va + offset;
5184 EXPORT_SYMBOL(page_frag_alloc);
5187 * Frees a page fragment allocated out of either a compound or order 0 page.
5189 void page_frag_free(void *addr)
5191 struct page *page = virt_to_head_page(addr);
5193 if (unlikely(put_page_testzero(page)))
5194 free_the_page(page, compound_order(page));
5196 EXPORT_SYMBOL(page_frag_free);
5198 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5202 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5203 unsigned long used = addr + PAGE_ALIGN(size);
5205 split_page(virt_to_page((void *)addr), order);
5206 while (used < alloc_end) {
5211 return (void *)addr;
5215 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5216 * @size: the number of bytes to allocate
5217 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5219 * This function is similar to alloc_pages(), except that it allocates the
5220 * minimum number of pages to satisfy the request. alloc_pages() can only
5221 * allocate memory in power-of-two pages.
5223 * This function is also limited by MAX_ORDER.
5225 * Memory allocated by this function must be released by free_pages_exact().
5227 * Return: pointer to the allocated area or %NULL in case of error.
5229 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5231 unsigned int order = get_order(size);
5234 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5235 gfp_mask &= ~__GFP_COMP;
5237 addr = __get_free_pages(gfp_mask, order);
5238 return make_alloc_exact(addr, order, size);
5240 EXPORT_SYMBOL(alloc_pages_exact);
5243 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5245 * @nid: the preferred node ID where memory should be allocated
5246 * @size: the number of bytes to allocate
5247 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5249 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5252 * Return: pointer to the allocated area or %NULL in case of error.
5254 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5256 unsigned int order = get_order(size);
5259 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5260 gfp_mask &= ~__GFP_COMP;
5262 p = alloc_pages_node(nid, gfp_mask, order);
5265 return make_alloc_exact((unsigned long)page_address(p), order, size);
5269 * free_pages_exact - release memory allocated via alloc_pages_exact()
5270 * @virt: the value returned by alloc_pages_exact.
5271 * @size: size of allocation, same value as passed to alloc_pages_exact().
5273 * Release the memory allocated by a previous call to alloc_pages_exact.
5275 void free_pages_exact(void *virt, size_t size)
5277 unsigned long addr = (unsigned long)virt;
5278 unsigned long end = addr + PAGE_ALIGN(size);
5280 while (addr < end) {
5285 EXPORT_SYMBOL(free_pages_exact);
5288 * nr_free_zone_pages - count number of pages beyond high watermark
5289 * @offset: The zone index of the highest zone
5291 * nr_free_zone_pages() counts the number of pages which are beyond the
5292 * high watermark within all zones at or below a given zone index. For each
5293 * zone, the number of pages is calculated as:
5295 * nr_free_zone_pages = managed_pages - high_pages
5297 * Return: number of pages beyond high watermark.
5299 static unsigned long nr_free_zone_pages(int offset)
5304 /* Just pick one node, since fallback list is circular */
5305 unsigned long sum = 0;
5307 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5309 for_each_zone_zonelist(zone, z, zonelist, offset) {
5310 unsigned long size = zone_managed_pages(zone);
5311 unsigned long high = high_wmark_pages(zone);
5320 * nr_free_buffer_pages - count number of pages beyond high watermark
5322 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5323 * watermark within ZONE_DMA and ZONE_NORMAL.
5325 * Return: number of pages beyond high watermark within ZONE_DMA and
5328 unsigned long nr_free_buffer_pages(void)
5330 return nr_free_zone_pages(gfp_zone(GFP_USER));
5332 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5334 static inline void show_node(struct zone *zone)
5336 if (IS_ENABLED(CONFIG_NUMA))
5337 printk("Node %d ", zone_to_nid(zone));
5340 long si_mem_available(void)
5343 unsigned long pagecache;
5344 unsigned long wmark_low = 0;
5345 unsigned long pages[NR_LRU_LISTS];
5346 unsigned long reclaimable;
5350 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5351 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5354 wmark_low += low_wmark_pages(zone);
5357 * Estimate the amount of memory available for userspace allocations,
5358 * without causing swapping.
5360 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5363 * Not all the page cache can be freed, otherwise the system will
5364 * start swapping. Assume at least half of the page cache, or the
5365 * low watermark worth of cache, needs to stay.
5367 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5368 pagecache -= min(pagecache / 2, wmark_low);
5369 available += pagecache;
5372 * Part of the reclaimable slab and other kernel memory consists of
5373 * items that are in use, and cannot be freed. Cap this estimate at the
5376 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5377 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5378 available += reclaimable - min(reclaimable / 2, wmark_low);
5384 EXPORT_SYMBOL_GPL(si_mem_available);
5386 void si_meminfo(struct sysinfo *val)
5388 val->totalram = totalram_pages();
5389 val->sharedram = global_node_page_state(NR_SHMEM);
5390 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5391 val->bufferram = nr_blockdev_pages();
5392 val->totalhigh = totalhigh_pages();
5393 val->freehigh = nr_free_highpages();
5394 val->mem_unit = PAGE_SIZE;
5397 EXPORT_SYMBOL(si_meminfo);
5400 void si_meminfo_node(struct sysinfo *val, int nid)
5402 int zone_type; /* needs to be signed */
5403 unsigned long managed_pages = 0;
5404 unsigned long managed_highpages = 0;
5405 unsigned long free_highpages = 0;
5406 pg_data_t *pgdat = NODE_DATA(nid);
5408 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5409 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5410 val->totalram = managed_pages;
5411 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5412 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5413 #ifdef CONFIG_HIGHMEM
5414 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5415 struct zone *zone = &pgdat->node_zones[zone_type];
5417 if (is_highmem(zone)) {
5418 managed_highpages += zone_managed_pages(zone);
5419 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5422 val->totalhigh = managed_highpages;
5423 val->freehigh = free_highpages;
5425 val->totalhigh = managed_highpages;
5426 val->freehigh = free_highpages;
5428 val->mem_unit = PAGE_SIZE;
5433 * Determine whether the node should be displayed or not, depending on whether
5434 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5436 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5438 if (!(flags & SHOW_MEM_FILTER_NODES))
5442 * no node mask - aka implicit memory numa policy. Do not bother with
5443 * the synchronization - read_mems_allowed_begin - because we do not
5444 * have to be precise here.
5447 nodemask = &cpuset_current_mems_allowed;
5449 return !node_isset(nid, *nodemask);
5452 #define K(x) ((x) << (PAGE_SHIFT-10))
5454 static void show_migration_types(unsigned char type)
5456 static const char types[MIGRATE_TYPES] = {
5457 [MIGRATE_UNMOVABLE] = 'U',
5458 [MIGRATE_MOVABLE] = 'M',
5459 [MIGRATE_RECLAIMABLE] = 'E',
5460 [MIGRATE_HIGHATOMIC] = 'H',
5462 [MIGRATE_CMA] = 'C',
5464 #ifdef CONFIG_MEMORY_ISOLATION
5465 [MIGRATE_ISOLATE] = 'I',
5468 char tmp[MIGRATE_TYPES + 1];
5472 for (i = 0; i < MIGRATE_TYPES; i++) {
5473 if (type & (1 << i))
5478 printk(KERN_CONT "(%s) ", tmp);
5482 * Show free area list (used inside shift_scroll-lock stuff)
5483 * We also calculate the percentage fragmentation. We do this by counting the
5484 * memory on each free list with the exception of the first item on the list.
5487 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5490 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5492 unsigned long free_pcp = 0;
5497 for_each_populated_zone(zone) {
5498 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5501 for_each_online_cpu(cpu)
5502 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5505 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5506 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5507 " unevictable:%lu dirty:%lu writeback:%lu\n"
5508 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5509 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5510 " free:%lu free_pcp:%lu free_cma:%lu\n",
5511 global_node_page_state(NR_ACTIVE_ANON),
5512 global_node_page_state(NR_INACTIVE_ANON),
5513 global_node_page_state(NR_ISOLATED_ANON),
5514 global_node_page_state(NR_ACTIVE_FILE),
5515 global_node_page_state(NR_INACTIVE_FILE),
5516 global_node_page_state(NR_ISOLATED_FILE),
5517 global_node_page_state(NR_UNEVICTABLE),
5518 global_node_page_state(NR_FILE_DIRTY),
5519 global_node_page_state(NR_WRITEBACK),
5520 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5521 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5522 global_node_page_state(NR_FILE_MAPPED),
5523 global_node_page_state(NR_SHMEM),
5524 global_node_page_state(NR_PAGETABLE),
5525 global_zone_page_state(NR_BOUNCE),
5526 global_zone_page_state(NR_FREE_PAGES),
5528 global_zone_page_state(NR_FREE_CMA_PAGES));
5530 for_each_online_pgdat(pgdat) {
5531 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5535 " active_anon:%lukB"
5536 " inactive_anon:%lukB"
5537 " active_file:%lukB"
5538 " inactive_file:%lukB"
5539 " unevictable:%lukB"
5540 " isolated(anon):%lukB"
5541 " isolated(file):%lukB"
5546 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5548 " shmem_pmdmapped: %lukB"
5551 " writeback_tmp:%lukB"
5552 " kernel_stack:%lukB"
5553 #ifdef CONFIG_SHADOW_CALL_STACK
5554 " shadow_call_stack:%lukB"
5557 " all_unreclaimable? %s"
5560 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5561 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5562 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5563 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5564 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5565 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5566 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5567 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5568 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5569 K(node_page_state(pgdat, NR_WRITEBACK)),
5570 K(node_page_state(pgdat, NR_SHMEM)),
5571 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5572 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5573 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5575 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5577 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5578 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5579 #ifdef CONFIG_SHADOW_CALL_STACK
5580 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5582 K(node_page_state(pgdat, NR_PAGETABLE)),
5583 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5587 for_each_populated_zone(zone) {
5590 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5594 for_each_online_cpu(cpu)
5595 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5604 " reserved_highatomic:%luKB"
5605 " active_anon:%lukB"
5606 " inactive_anon:%lukB"
5607 " active_file:%lukB"
5608 " inactive_file:%lukB"
5609 " unevictable:%lukB"
5610 " writepending:%lukB"
5620 K(zone_page_state(zone, NR_FREE_PAGES)),
5621 K(min_wmark_pages(zone)),
5622 K(low_wmark_pages(zone)),
5623 K(high_wmark_pages(zone)),
5624 K(zone->nr_reserved_highatomic),
5625 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5626 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5627 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5628 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5629 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5630 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5631 K(zone->present_pages),
5632 K(zone_managed_pages(zone)),
5633 K(zone_page_state(zone, NR_MLOCK)),
5634 K(zone_page_state(zone, NR_BOUNCE)),
5636 K(this_cpu_read(zone->pageset->pcp.count)),
5637 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5638 printk("lowmem_reserve[]:");
5639 for (i = 0; i < MAX_NR_ZONES; i++)
5640 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5641 printk(KERN_CONT "\n");
5644 for_each_populated_zone(zone) {
5646 unsigned long nr[MAX_ORDER], flags, total = 0;
5647 unsigned char types[MAX_ORDER];
5649 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5652 printk(KERN_CONT "%s: ", zone->name);
5654 spin_lock_irqsave(&zone->lock, flags);
5655 for (order = 0; order < MAX_ORDER; order++) {
5656 struct free_area *area = &zone->free_area[order];
5659 nr[order] = area->nr_free;
5660 total += nr[order] << order;
5663 for (type = 0; type < MIGRATE_TYPES; type++) {
5664 if (!free_area_empty(area, type))
5665 types[order] |= 1 << type;
5668 spin_unlock_irqrestore(&zone->lock, flags);
5669 for (order = 0; order < MAX_ORDER; order++) {
5670 printk(KERN_CONT "%lu*%lukB ",
5671 nr[order], K(1UL) << order);
5673 show_migration_types(types[order]);
5675 printk(KERN_CONT "= %lukB\n", K(total));
5678 hugetlb_show_meminfo();
5680 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5682 show_swap_cache_info();
5685 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5687 zoneref->zone = zone;
5688 zoneref->zone_idx = zone_idx(zone);
5692 * Builds allocation fallback zone lists.
5694 * Add all populated zones of a node to the zonelist.
5696 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5699 enum zone_type zone_type = MAX_NR_ZONES;
5704 zone = pgdat->node_zones + zone_type;
5705 if (managed_zone(zone)) {
5706 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5707 check_highest_zone(zone_type);
5709 } while (zone_type);
5716 static int __parse_numa_zonelist_order(char *s)
5719 * We used to support different zonlists modes but they turned
5720 * out to be just not useful. Let's keep the warning in place
5721 * if somebody still use the cmd line parameter so that we do
5722 * not fail it silently
5724 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5725 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5731 char numa_zonelist_order[] = "Node";
5734 * sysctl handler for numa_zonelist_order
5736 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5737 void *buffer, size_t *length, loff_t *ppos)
5740 return __parse_numa_zonelist_order(buffer);
5741 return proc_dostring(table, write, buffer, length, ppos);
5745 #define MAX_NODE_LOAD (nr_online_nodes)
5746 static int node_load[MAX_NUMNODES];
5749 * find_next_best_node - find the next node that should appear in a given node's fallback list
5750 * @node: node whose fallback list we're appending
5751 * @used_node_mask: nodemask_t of already used nodes
5753 * We use a number of factors to determine which is the next node that should
5754 * appear on a given node's fallback list. The node should not have appeared
5755 * already in @node's fallback list, and it should be the next closest node
5756 * according to the distance array (which contains arbitrary distance values
5757 * from each node to each node in the system), and should also prefer nodes
5758 * with no CPUs, since presumably they'll have very little allocation pressure
5759 * on them otherwise.
5761 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5763 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5766 int min_val = INT_MAX;
5767 int best_node = NUMA_NO_NODE;
5769 /* Use the local node if we haven't already */
5770 if (!node_isset(node, *used_node_mask)) {
5771 node_set(node, *used_node_mask);
5775 for_each_node_state(n, N_MEMORY) {
5777 /* Don't want a node to appear more than once */
5778 if (node_isset(n, *used_node_mask))
5781 /* Use the distance array to find the distance */
5782 val = node_distance(node, n);
5784 /* Penalize nodes under us ("prefer the next node") */
5787 /* Give preference to headless and unused nodes */
5788 if (!cpumask_empty(cpumask_of_node(n)))
5789 val += PENALTY_FOR_NODE_WITH_CPUS;
5791 /* Slight preference for less loaded node */
5792 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5793 val += node_load[n];
5795 if (val < min_val) {
5802 node_set(best_node, *used_node_mask);
5809 * Build zonelists ordered by node and zones within node.
5810 * This results in maximum locality--normal zone overflows into local
5811 * DMA zone, if any--but risks exhausting DMA zone.
5813 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5816 struct zoneref *zonerefs;
5819 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5821 for (i = 0; i < nr_nodes; i++) {
5824 pg_data_t *node = NODE_DATA(node_order[i]);
5826 nr_zones = build_zonerefs_node(node, zonerefs);
5827 zonerefs += nr_zones;
5829 zonerefs->zone = NULL;
5830 zonerefs->zone_idx = 0;
5834 * Build gfp_thisnode zonelists
5836 static void build_thisnode_zonelists(pg_data_t *pgdat)
5838 struct zoneref *zonerefs;
5841 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5842 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5843 zonerefs += nr_zones;
5844 zonerefs->zone = NULL;
5845 zonerefs->zone_idx = 0;
5849 * Build zonelists ordered by zone and nodes within zones.
5850 * This results in conserving DMA zone[s] until all Normal memory is
5851 * exhausted, but results in overflowing to remote node while memory
5852 * may still exist in local DMA zone.
5855 static void build_zonelists(pg_data_t *pgdat)
5857 static int node_order[MAX_NUMNODES];
5858 int node, load, nr_nodes = 0;
5859 nodemask_t used_mask = NODE_MASK_NONE;
5860 int local_node, prev_node;
5862 /* NUMA-aware ordering of nodes */
5863 local_node = pgdat->node_id;
5864 load = nr_online_nodes;
5865 prev_node = local_node;
5867 memset(node_order, 0, sizeof(node_order));
5868 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5870 * We don't want to pressure a particular node.
5871 * So adding penalty to the first node in same
5872 * distance group to make it round-robin.
5874 if (node_distance(local_node, node) !=
5875 node_distance(local_node, prev_node))
5876 node_load[node] = load;
5878 node_order[nr_nodes++] = node;
5883 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5884 build_thisnode_zonelists(pgdat);
5887 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5889 * Return node id of node used for "local" allocations.
5890 * I.e., first node id of first zone in arg node's generic zonelist.
5891 * Used for initializing percpu 'numa_mem', which is used primarily
5892 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5894 int local_memory_node(int node)
5898 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5899 gfp_zone(GFP_KERNEL),
5901 return zone_to_nid(z->zone);
5905 static void setup_min_unmapped_ratio(void);
5906 static void setup_min_slab_ratio(void);
5907 #else /* CONFIG_NUMA */
5909 static void build_zonelists(pg_data_t *pgdat)
5911 int node, local_node;
5912 struct zoneref *zonerefs;
5915 local_node = pgdat->node_id;
5917 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5918 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5919 zonerefs += nr_zones;
5922 * Now we build the zonelist so that it contains the zones
5923 * of all the other nodes.
5924 * We don't want to pressure a particular node, so when
5925 * building the zones for node N, we make sure that the
5926 * zones coming right after the local ones are those from
5927 * node N+1 (modulo N)
5929 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5930 if (!node_online(node))
5932 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5933 zonerefs += nr_zones;
5935 for (node = 0; node < local_node; node++) {
5936 if (!node_online(node))
5938 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5939 zonerefs += nr_zones;
5942 zonerefs->zone = NULL;
5943 zonerefs->zone_idx = 0;
5946 #endif /* CONFIG_NUMA */
5949 * Boot pageset table. One per cpu which is going to be used for all
5950 * zones and all nodes. The parameters will be set in such a way
5951 * that an item put on a list will immediately be handed over to
5952 * the buddy list. This is safe since pageset manipulation is done
5953 * with interrupts disabled.
5955 * The boot_pagesets must be kept even after bootup is complete for
5956 * unused processors and/or zones. They do play a role for bootstrapping
5957 * hotplugged processors.
5959 * zoneinfo_show() and maybe other functions do
5960 * not check if the processor is online before following the pageset pointer.
5961 * Other parts of the kernel may not check if the zone is available.
5963 static void pageset_init(struct per_cpu_pageset *p);
5964 /* These effectively disable the pcplists in the boot pageset completely */
5965 #define BOOT_PAGESET_HIGH 0
5966 #define BOOT_PAGESET_BATCH 1
5967 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5968 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5970 static void __build_all_zonelists(void *data)
5973 int __maybe_unused cpu;
5974 pg_data_t *self = data;
5975 static DEFINE_SPINLOCK(lock);
5980 memset(node_load, 0, sizeof(node_load));
5984 * This node is hotadded and no memory is yet present. So just
5985 * building zonelists is fine - no need to touch other nodes.
5987 if (self && !node_online(self->node_id)) {
5988 build_zonelists(self);
5990 for_each_online_node(nid) {
5991 pg_data_t *pgdat = NODE_DATA(nid);
5993 build_zonelists(pgdat);
5996 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5998 * We now know the "local memory node" for each node--
5999 * i.e., the node of the first zone in the generic zonelist.
6000 * Set up numa_mem percpu variable for on-line cpus. During
6001 * boot, only the boot cpu should be on-line; we'll init the
6002 * secondary cpus' numa_mem as they come on-line. During
6003 * node/memory hotplug, we'll fixup all on-line cpus.
6005 for_each_online_cpu(cpu)
6006 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6013 static noinline void __init
6014 build_all_zonelists_init(void)
6018 __build_all_zonelists(NULL);
6021 * Initialize the boot_pagesets that are going to be used
6022 * for bootstrapping processors. The real pagesets for
6023 * each zone will be allocated later when the per cpu
6024 * allocator is available.
6026 * boot_pagesets are used also for bootstrapping offline
6027 * cpus if the system is already booted because the pagesets
6028 * are needed to initialize allocators on a specific cpu too.
6029 * F.e. the percpu allocator needs the page allocator which
6030 * needs the percpu allocator in order to allocate its pagesets
6031 * (a chicken-egg dilemma).
6033 for_each_possible_cpu(cpu)
6034 pageset_init(&per_cpu(boot_pageset, cpu));
6036 mminit_verify_zonelist();
6037 cpuset_init_current_mems_allowed();
6041 * unless system_state == SYSTEM_BOOTING.
6043 * __ref due to call of __init annotated helper build_all_zonelists_init
6044 * [protected by SYSTEM_BOOTING].
6046 void __ref build_all_zonelists(pg_data_t *pgdat)
6048 unsigned long vm_total_pages;
6050 if (system_state == SYSTEM_BOOTING) {
6051 build_all_zonelists_init();
6053 __build_all_zonelists(pgdat);
6054 /* cpuset refresh routine should be here */
6056 /* Get the number of free pages beyond high watermark in all zones. */
6057 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6059 * Disable grouping by mobility if the number of pages in the
6060 * system is too low to allow the mechanism to work. It would be
6061 * more accurate, but expensive to check per-zone. This check is
6062 * made on memory-hotadd so a system can start with mobility
6063 * disabled and enable it later
6065 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6066 page_group_by_mobility_disabled = 1;
6068 page_group_by_mobility_disabled = 0;
6070 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6072 page_group_by_mobility_disabled ? "off" : "on",
6075 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6079 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6080 static bool __meminit
6081 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6083 static struct memblock_region *r;
6085 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6086 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6087 for_each_mem_region(r) {
6088 if (*pfn < memblock_region_memory_end_pfn(r))
6092 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6093 memblock_is_mirror(r)) {
6094 *pfn = memblock_region_memory_end_pfn(r);
6102 * Initially all pages are reserved - free ones are freed
6103 * up by memblock_free_all() once the early boot process is
6104 * done. Non-atomic initialization, single-pass.
6106 * All aligned pageblocks are initialized to the specified migratetype
6107 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6108 * zone stats (e.g., nr_isolate_pageblock) are touched.
6110 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6111 unsigned long start_pfn,
6112 enum meminit_context context,
6113 struct vmem_altmap *altmap, int migratetype)
6115 unsigned long pfn, end_pfn = start_pfn + size;
6118 if (highest_memmap_pfn < end_pfn - 1)
6119 highest_memmap_pfn = end_pfn - 1;
6121 #ifdef CONFIG_ZONE_DEVICE
6123 * Honor reservation requested by the driver for this ZONE_DEVICE
6124 * memory. We limit the total number of pages to initialize to just
6125 * those that might contain the memory mapping. We will defer the
6126 * ZONE_DEVICE page initialization until after we have released
6129 if (zone == ZONE_DEVICE) {
6133 if (start_pfn == altmap->base_pfn)
6134 start_pfn += altmap->reserve;
6135 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6139 for (pfn = start_pfn; pfn < end_pfn; ) {
6141 * There can be holes in boot-time mem_map[]s handed to this
6142 * function. They do not exist on hotplugged memory.
6144 if (context == MEMINIT_EARLY) {
6145 if (overlap_memmap_init(zone, &pfn))
6147 if (defer_init(nid, pfn, end_pfn))
6151 page = pfn_to_page(pfn);
6152 __init_single_page(page, pfn, zone, nid);
6153 if (context == MEMINIT_HOTPLUG)
6154 __SetPageReserved(page);
6157 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6158 * such that unmovable allocations won't be scattered all
6159 * over the place during system boot.
6161 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6162 set_pageblock_migratetype(page, migratetype);
6169 #ifdef CONFIG_ZONE_DEVICE
6170 void __ref memmap_init_zone_device(struct zone *zone,
6171 unsigned long start_pfn,
6172 unsigned long nr_pages,
6173 struct dev_pagemap *pgmap)
6175 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6176 struct pglist_data *pgdat = zone->zone_pgdat;
6177 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6178 unsigned long zone_idx = zone_idx(zone);
6179 unsigned long start = jiffies;
6180 int nid = pgdat->node_id;
6182 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6186 * The call to memmap_init_zone should have already taken care
6187 * of the pages reserved for the memmap, so we can just jump to
6188 * the end of that region and start processing the device pages.
6191 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6192 nr_pages = end_pfn - start_pfn;
6195 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6196 struct page *page = pfn_to_page(pfn);
6198 __init_single_page(page, pfn, zone_idx, nid);
6201 * Mark page reserved as it will need to wait for onlining
6202 * phase for it to be fully associated with a zone.
6204 * We can use the non-atomic __set_bit operation for setting
6205 * the flag as we are still initializing the pages.
6207 __SetPageReserved(page);
6210 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6211 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6212 * ever freed or placed on a driver-private list.
6214 page->pgmap = pgmap;
6215 page->zone_device_data = NULL;
6218 * Mark the block movable so that blocks are reserved for
6219 * movable at startup. This will force kernel allocations
6220 * to reserve their blocks rather than leaking throughout
6221 * the address space during boot when many long-lived
6222 * kernel allocations are made.
6224 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6225 * because this is done early in section_activate()
6227 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6228 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6233 pr_info("%s initialised %lu pages in %ums\n", __func__,
6234 nr_pages, jiffies_to_msecs(jiffies - start));
6238 static void __meminit zone_init_free_lists(struct zone *zone)
6240 unsigned int order, t;
6241 for_each_migratetype_order(order, t) {
6242 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6243 zone->free_area[order].nr_free = 0;
6247 void __meminit __weak memmap_init(unsigned long size, int nid,
6249 unsigned long range_start_pfn)
6251 unsigned long start_pfn, end_pfn;
6252 unsigned long range_end_pfn = range_start_pfn + size;
6255 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6256 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6257 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6259 if (end_pfn > start_pfn) {
6260 size = end_pfn - start_pfn;
6261 memmap_init_zone(size, nid, zone, start_pfn,
6262 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6267 static int zone_batchsize(struct zone *zone)
6273 * The per-cpu-pages pools are set to around 1000th of the
6276 batch = zone_managed_pages(zone) / 1024;
6277 /* But no more than a meg. */
6278 if (batch * PAGE_SIZE > 1024 * 1024)
6279 batch = (1024 * 1024) / PAGE_SIZE;
6280 batch /= 4; /* We effectively *= 4 below */
6285 * Clamp the batch to a 2^n - 1 value. Having a power
6286 * of 2 value was found to be more likely to have
6287 * suboptimal cache aliasing properties in some cases.
6289 * For example if 2 tasks are alternately allocating
6290 * batches of pages, one task can end up with a lot
6291 * of pages of one half of the possible page colors
6292 * and the other with pages of the other colors.
6294 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6299 /* The deferral and batching of frees should be suppressed under NOMMU
6302 * The problem is that NOMMU needs to be able to allocate large chunks
6303 * of contiguous memory as there's no hardware page translation to
6304 * assemble apparent contiguous memory from discontiguous pages.
6306 * Queueing large contiguous runs of pages for batching, however,
6307 * causes the pages to actually be freed in smaller chunks. As there
6308 * can be a significant delay between the individual batches being
6309 * recycled, this leads to the once large chunks of space being
6310 * fragmented and becoming unavailable for high-order allocations.
6317 * pcp->high and pcp->batch values are related and generally batch is lower
6318 * than high. They are also related to pcp->count such that count is lower
6319 * than high, and as soon as it reaches high, the pcplist is flushed.
6321 * However, guaranteeing these relations at all times would require e.g. write
6322 * barriers here but also careful usage of read barriers at the read side, and
6323 * thus be prone to error and bad for performance. Thus the update only prevents
6324 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6325 * can cope with those fields changing asynchronously, and fully trust only the
6326 * pcp->count field on the local CPU with interrupts disabled.
6328 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6329 * outside of boot time (or some other assurance that no concurrent updaters
6332 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6333 unsigned long batch)
6335 WRITE_ONCE(pcp->batch, batch);
6336 WRITE_ONCE(pcp->high, high);
6339 static void pageset_init(struct per_cpu_pageset *p)
6341 struct per_cpu_pages *pcp;
6344 memset(p, 0, sizeof(*p));
6347 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6348 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6351 * Set batch and high values safe for a boot pageset. A true percpu
6352 * pageset's initialization will update them subsequently. Here we don't
6353 * need to be as careful as pageset_update() as nobody can access the
6356 pcp->high = BOOT_PAGESET_HIGH;
6357 pcp->batch = BOOT_PAGESET_BATCH;
6360 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6361 unsigned long batch)
6363 struct per_cpu_pageset *p;
6366 for_each_possible_cpu(cpu) {
6367 p = per_cpu_ptr(zone->pageset, cpu);
6368 pageset_update(&p->pcp, high, batch);
6373 * Calculate and set new high and batch values for all per-cpu pagesets of a
6374 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6376 static void zone_set_pageset_high_and_batch(struct zone *zone)
6378 unsigned long new_high, new_batch;
6380 if (percpu_pagelist_fraction) {
6381 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6382 new_batch = max(1UL, new_high / 4);
6383 if ((new_high / 4) > (PAGE_SHIFT * 8))
6384 new_batch = PAGE_SHIFT * 8;
6386 new_batch = zone_batchsize(zone);
6387 new_high = 6 * new_batch;
6388 new_batch = max(1UL, 1 * new_batch);
6391 if (zone->pageset_high == new_high &&
6392 zone->pageset_batch == new_batch)
6395 zone->pageset_high = new_high;
6396 zone->pageset_batch = new_batch;
6398 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6401 void __meminit setup_zone_pageset(struct zone *zone)
6403 struct per_cpu_pageset *p;
6406 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6407 for_each_possible_cpu(cpu) {
6408 p = per_cpu_ptr(zone->pageset, cpu);
6412 zone_set_pageset_high_and_batch(zone);
6416 * Allocate per cpu pagesets and initialize them.
6417 * Before this call only boot pagesets were available.
6419 void __init setup_per_cpu_pageset(void)
6421 struct pglist_data *pgdat;
6423 int __maybe_unused cpu;
6425 for_each_populated_zone(zone)
6426 setup_zone_pageset(zone);
6430 * Unpopulated zones continue using the boot pagesets.
6431 * The numa stats for these pagesets need to be reset.
6432 * Otherwise, they will end up skewing the stats of
6433 * the nodes these zones are associated with.
6435 for_each_possible_cpu(cpu) {
6436 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6437 memset(pcp->vm_numa_stat_diff, 0,
6438 sizeof(pcp->vm_numa_stat_diff));
6442 for_each_online_pgdat(pgdat)
6443 pgdat->per_cpu_nodestats =
6444 alloc_percpu(struct per_cpu_nodestat);
6447 static __meminit void zone_pcp_init(struct zone *zone)
6450 * per cpu subsystem is not up at this point. The following code
6451 * relies on the ability of the linker to provide the
6452 * offset of a (static) per cpu variable into the per cpu area.
6454 zone->pageset = &boot_pageset;
6455 zone->pageset_high = BOOT_PAGESET_HIGH;
6456 zone->pageset_batch = BOOT_PAGESET_BATCH;
6458 if (populated_zone(zone))
6459 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6460 zone->name, zone->present_pages,
6461 zone_batchsize(zone));
6464 void __meminit init_currently_empty_zone(struct zone *zone,
6465 unsigned long zone_start_pfn,
6468 struct pglist_data *pgdat = zone->zone_pgdat;
6469 int zone_idx = zone_idx(zone) + 1;
6471 if (zone_idx > pgdat->nr_zones)
6472 pgdat->nr_zones = zone_idx;
6474 zone->zone_start_pfn = zone_start_pfn;
6476 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6477 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6479 (unsigned long)zone_idx(zone),
6480 zone_start_pfn, (zone_start_pfn + size));
6482 zone_init_free_lists(zone);
6483 zone->initialized = 1;
6487 * get_pfn_range_for_nid - Return the start and end page frames for a node
6488 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6489 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6490 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6492 * It returns the start and end page frame of a node based on information
6493 * provided by memblock_set_node(). If called for a node
6494 * with no available memory, a warning is printed and the start and end
6497 void __init get_pfn_range_for_nid(unsigned int nid,
6498 unsigned long *start_pfn, unsigned long *end_pfn)
6500 unsigned long this_start_pfn, this_end_pfn;
6506 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6507 *start_pfn = min(*start_pfn, this_start_pfn);
6508 *end_pfn = max(*end_pfn, this_end_pfn);
6511 if (*start_pfn == -1UL)
6516 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6517 * assumption is made that zones within a node are ordered in monotonic
6518 * increasing memory addresses so that the "highest" populated zone is used
6520 static void __init find_usable_zone_for_movable(void)
6523 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6524 if (zone_index == ZONE_MOVABLE)
6527 if (arch_zone_highest_possible_pfn[zone_index] >
6528 arch_zone_lowest_possible_pfn[zone_index])
6532 VM_BUG_ON(zone_index == -1);
6533 movable_zone = zone_index;
6537 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6538 * because it is sized independent of architecture. Unlike the other zones,
6539 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6540 * in each node depending on the size of each node and how evenly kernelcore
6541 * is distributed. This helper function adjusts the zone ranges
6542 * provided by the architecture for a given node by using the end of the
6543 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6544 * zones within a node are in order of monotonic increases memory addresses
6546 static void __init adjust_zone_range_for_zone_movable(int nid,
6547 unsigned long zone_type,
6548 unsigned long node_start_pfn,
6549 unsigned long node_end_pfn,
6550 unsigned long *zone_start_pfn,
6551 unsigned long *zone_end_pfn)
6553 /* Only adjust if ZONE_MOVABLE is on this node */
6554 if (zone_movable_pfn[nid]) {
6555 /* Size ZONE_MOVABLE */
6556 if (zone_type == ZONE_MOVABLE) {
6557 *zone_start_pfn = zone_movable_pfn[nid];
6558 *zone_end_pfn = min(node_end_pfn,
6559 arch_zone_highest_possible_pfn[movable_zone]);
6561 /* Adjust for ZONE_MOVABLE starting within this range */
6562 } else if (!mirrored_kernelcore &&
6563 *zone_start_pfn < zone_movable_pfn[nid] &&
6564 *zone_end_pfn > zone_movable_pfn[nid]) {
6565 *zone_end_pfn = zone_movable_pfn[nid];
6567 /* Check if this whole range is within ZONE_MOVABLE */
6568 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6569 *zone_start_pfn = *zone_end_pfn;
6574 * Return the number of pages a zone spans in a node, including holes
6575 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6577 static unsigned long __init zone_spanned_pages_in_node(int nid,
6578 unsigned long zone_type,
6579 unsigned long node_start_pfn,
6580 unsigned long node_end_pfn,
6581 unsigned long *zone_start_pfn,
6582 unsigned long *zone_end_pfn)
6584 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6585 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6586 /* When hotadd a new node from cpu_up(), the node should be empty */
6587 if (!node_start_pfn && !node_end_pfn)
6590 /* Get the start and end of the zone */
6591 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6592 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6593 adjust_zone_range_for_zone_movable(nid, zone_type,
6594 node_start_pfn, node_end_pfn,
6595 zone_start_pfn, zone_end_pfn);
6597 /* Check that this node has pages within the zone's required range */
6598 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6601 /* Move the zone boundaries inside the node if necessary */
6602 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6603 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6605 /* Return the spanned pages */
6606 return *zone_end_pfn - *zone_start_pfn;
6610 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6611 * then all holes in the requested range will be accounted for.
6613 unsigned long __init __absent_pages_in_range(int nid,
6614 unsigned long range_start_pfn,
6615 unsigned long range_end_pfn)
6617 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6618 unsigned long start_pfn, end_pfn;
6621 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6622 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6623 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6624 nr_absent -= end_pfn - start_pfn;
6630 * absent_pages_in_range - Return number of page frames in holes within a range
6631 * @start_pfn: The start PFN to start searching for holes
6632 * @end_pfn: The end PFN to stop searching for holes
6634 * Return: the number of pages frames in memory holes within a range.
6636 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6637 unsigned long end_pfn)
6639 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6642 /* Return the number of page frames in holes in a zone on a node */
6643 static unsigned long __init zone_absent_pages_in_node(int nid,
6644 unsigned long zone_type,
6645 unsigned long node_start_pfn,
6646 unsigned long node_end_pfn)
6648 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6649 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6650 unsigned long zone_start_pfn, zone_end_pfn;
6651 unsigned long nr_absent;
6653 /* When hotadd a new node from cpu_up(), the node should be empty */
6654 if (!node_start_pfn && !node_end_pfn)
6657 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6658 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6660 adjust_zone_range_for_zone_movable(nid, zone_type,
6661 node_start_pfn, node_end_pfn,
6662 &zone_start_pfn, &zone_end_pfn);
6663 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6666 * ZONE_MOVABLE handling.
6667 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6670 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6671 unsigned long start_pfn, end_pfn;
6672 struct memblock_region *r;
6674 for_each_mem_region(r) {
6675 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6676 zone_start_pfn, zone_end_pfn);
6677 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6678 zone_start_pfn, zone_end_pfn);
6680 if (zone_type == ZONE_MOVABLE &&
6681 memblock_is_mirror(r))
6682 nr_absent += end_pfn - start_pfn;
6684 if (zone_type == ZONE_NORMAL &&
6685 !memblock_is_mirror(r))
6686 nr_absent += end_pfn - start_pfn;
6693 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6694 unsigned long node_start_pfn,
6695 unsigned long node_end_pfn)
6697 unsigned long realtotalpages = 0, totalpages = 0;
6700 for (i = 0; i < MAX_NR_ZONES; i++) {
6701 struct zone *zone = pgdat->node_zones + i;
6702 unsigned long zone_start_pfn, zone_end_pfn;
6703 unsigned long spanned, absent;
6704 unsigned long size, real_size;
6706 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6711 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6716 real_size = size - absent;
6719 zone->zone_start_pfn = zone_start_pfn;
6721 zone->zone_start_pfn = 0;
6722 zone->spanned_pages = size;
6723 zone->present_pages = real_size;
6726 realtotalpages += real_size;
6729 pgdat->node_spanned_pages = totalpages;
6730 pgdat->node_present_pages = realtotalpages;
6731 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6735 #ifndef CONFIG_SPARSEMEM
6737 * Calculate the size of the zone->blockflags rounded to an unsigned long
6738 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6739 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6740 * round what is now in bits to nearest long in bits, then return it in
6743 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6745 unsigned long usemapsize;
6747 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6748 usemapsize = roundup(zonesize, pageblock_nr_pages);
6749 usemapsize = usemapsize >> pageblock_order;
6750 usemapsize *= NR_PAGEBLOCK_BITS;
6751 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6753 return usemapsize / 8;
6756 static void __ref setup_usemap(struct pglist_data *pgdat,
6758 unsigned long zone_start_pfn,
6759 unsigned long zonesize)
6761 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6762 zone->pageblock_flags = NULL;
6764 zone->pageblock_flags =
6765 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6767 if (!zone->pageblock_flags)
6768 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6769 usemapsize, zone->name, pgdat->node_id);
6773 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6774 unsigned long zone_start_pfn, unsigned long zonesize) {}
6775 #endif /* CONFIG_SPARSEMEM */
6777 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6779 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6780 void __init set_pageblock_order(void)
6784 /* Check that pageblock_nr_pages has not already been setup */
6785 if (pageblock_order)
6788 if (HPAGE_SHIFT > PAGE_SHIFT)
6789 order = HUGETLB_PAGE_ORDER;
6791 order = MAX_ORDER - 1;
6794 * Assume the largest contiguous order of interest is a huge page.
6795 * This value may be variable depending on boot parameters on IA64 and
6798 pageblock_order = order;
6800 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6803 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6804 * is unused as pageblock_order is set at compile-time. See
6805 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6808 void __init set_pageblock_order(void)
6812 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6814 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6815 unsigned long present_pages)
6817 unsigned long pages = spanned_pages;
6820 * Provide a more accurate estimation if there are holes within
6821 * the zone and SPARSEMEM is in use. If there are holes within the
6822 * zone, each populated memory region may cost us one or two extra
6823 * memmap pages due to alignment because memmap pages for each
6824 * populated regions may not be naturally aligned on page boundary.
6825 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6827 if (spanned_pages > present_pages + (present_pages >> 4) &&
6828 IS_ENABLED(CONFIG_SPARSEMEM))
6829 pages = present_pages;
6831 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6834 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6835 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6837 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6839 spin_lock_init(&ds_queue->split_queue_lock);
6840 INIT_LIST_HEAD(&ds_queue->split_queue);
6841 ds_queue->split_queue_len = 0;
6844 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6847 #ifdef CONFIG_COMPACTION
6848 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6850 init_waitqueue_head(&pgdat->kcompactd_wait);
6853 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6856 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6858 pgdat_resize_init(pgdat);
6860 pgdat_init_split_queue(pgdat);
6861 pgdat_init_kcompactd(pgdat);
6863 init_waitqueue_head(&pgdat->kswapd_wait);
6864 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6866 pgdat_page_ext_init(pgdat);
6867 spin_lock_init(&pgdat->lru_lock);
6868 lruvec_init(&pgdat->__lruvec);
6871 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6872 unsigned long remaining_pages)
6874 atomic_long_set(&zone->managed_pages, remaining_pages);
6875 zone_set_nid(zone, nid);
6876 zone->name = zone_names[idx];
6877 zone->zone_pgdat = NODE_DATA(nid);
6878 spin_lock_init(&zone->lock);
6879 zone_seqlock_init(zone);
6880 zone_pcp_init(zone);
6884 * Set up the zone data structures
6885 * - init pgdat internals
6886 * - init all zones belonging to this node
6888 * NOTE: this function is only called during memory hotplug
6890 #ifdef CONFIG_MEMORY_HOTPLUG
6891 void __ref free_area_init_core_hotplug(int nid)
6894 pg_data_t *pgdat = NODE_DATA(nid);
6896 pgdat_init_internals(pgdat);
6897 for (z = 0; z < MAX_NR_ZONES; z++)
6898 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6903 * Set up the zone data structures:
6904 * - mark all pages reserved
6905 * - mark all memory queues empty
6906 * - clear the memory bitmaps
6908 * NOTE: pgdat should get zeroed by caller.
6909 * NOTE: this function is only called during early init.
6911 static void __init free_area_init_core(struct pglist_data *pgdat)
6914 int nid = pgdat->node_id;
6916 pgdat_init_internals(pgdat);
6917 pgdat->per_cpu_nodestats = &boot_nodestats;
6919 for (j = 0; j < MAX_NR_ZONES; j++) {
6920 struct zone *zone = pgdat->node_zones + j;
6921 unsigned long size, freesize, memmap_pages;
6922 unsigned long zone_start_pfn = zone->zone_start_pfn;
6924 size = zone->spanned_pages;
6925 freesize = zone->present_pages;
6928 * Adjust freesize so that it accounts for how much memory
6929 * is used by this zone for memmap. This affects the watermark
6930 * and per-cpu initialisations
6932 memmap_pages = calc_memmap_size(size, freesize);
6933 if (!is_highmem_idx(j)) {
6934 if (freesize >= memmap_pages) {
6935 freesize -= memmap_pages;
6938 " %s zone: %lu pages used for memmap\n",
6939 zone_names[j], memmap_pages);
6941 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6942 zone_names[j], memmap_pages, freesize);
6945 /* Account for reserved pages */
6946 if (j == 0 && freesize > dma_reserve) {
6947 freesize -= dma_reserve;
6948 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6949 zone_names[0], dma_reserve);
6952 if (!is_highmem_idx(j))
6953 nr_kernel_pages += freesize;
6954 /* Charge for highmem memmap if there are enough kernel pages */
6955 else if (nr_kernel_pages > memmap_pages * 2)
6956 nr_kernel_pages -= memmap_pages;
6957 nr_all_pages += freesize;
6960 * Set an approximate value for lowmem here, it will be adjusted
6961 * when the bootmem allocator frees pages into the buddy system.
6962 * And all highmem pages will be managed by the buddy system.
6964 zone_init_internals(zone, j, nid, freesize);
6969 set_pageblock_order();
6970 setup_usemap(pgdat, zone, zone_start_pfn, size);
6971 init_currently_empty_zone(zone, zone_start_pfn, size);
6972 memmap_init(size, nid, j, zone_start_pfn);
6976 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6977 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6979 unsigned long __maybe_unused start = 0;
6980 unsigned long __maybe_unused offset = 0;
6982 /* Skip empty nodes */
6983 if (!pgdat->node_spanned_pages)
6986 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6987 offset = pgdat->node_start_pfn - start;
6988 /* ia64 gets its own node_mem_map, before this, without bootmem */
6989 if (!pgdat->node_mem_map) {
6990 unsigned long size, end;
6994 * The zone's endpoints aren't required to be MAX_ORDER
6995 * aligned but the node_mem_map endpoints must be in order
6996 * for the buddy allocator to function correctly.
6998 end = pgdat_end_pfn(pgdat);
6999 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7000 size = (end - start) * sizeof(struct page);
7001 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7004 panic("Failed to allocate %ld bytes for node %d memory map\n",
7005 size, pgdat->node_id);
7006 pgdat->node_mem_map = map + offset;
7008 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7009 __func__, pgdat->node_id, (unsigned long)pgdat,
7010 (unsigned long)pgdat->node_mem_map);
7011 #ifndef CONFIG_NEED_MULTIPLE_NODES
7013 * With no DISCONTIG, the global mem_map is just set as node 0's
7015 if (pgdat == NODE_DATA(0)) {
7016 mem_map = NODE_DATA(0)->node_mem_map;
7017 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7023 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7024 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7026 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7027 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7029 pgdat->first_deferred_pfn = ULONG_MAX;
7032 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7035 static void __init free_area_init_node(int nid)
7037 pg_data_t *pgdat = NODE_DATA(nid);
7038 unsigned long start_pfn = 0;
7039 unsigned long end_pfn = 0;
7041 /* pg_data_t should be reset to zero when it's allocated */
7042 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7044 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7046 pgdat->node_id = nid;
7047 pgdat->node_start_pfn = start_pfn;
7048 pgdat->per_cpu_nodestats = NULL;
7050 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7051 (u64)start_pfn << PAGE_SHIFT,
7052 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7053 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7055 alloc_node_mem_map(pgdat);
7056 pgdat_set_deferred_range(pgdat);
7058 free_area_init_core(pgdat);
7061 void __init free_area_init_memoryless_node(int nid)
7063 free_area_init_node(nid);
7066 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7068 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7069 * PageReserved(). Return the number of struct pages that were initialized.
7071 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7076 for (pfn = spfn; pfn < epfn; pfn++) {
7077 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7078 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7079 + pageblock_nr_pages - 1;
7083 * Use a fake node/zone (0) for now. Some of these pages
7084 * (in memblock.reserved but not in memblock.memory) will
7085 * get re-initialized via reserve_bootmem_region() later.
7087 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7088 __SetPageReserved(pfn_to_page(pfn));
7096 * Only struct pages that are backed by physical memory are zeroed and
7097 * initialized by going through __init_single_page(). But, there are some
7098 * struct pages which are reserved in memblock allocator and their fields
7099 * may be accessed (for example page_to_pfn() on some configuration accesses
7100 * flags). We must explicitly initialize those struct pages.
7102 * This function also addresses a similar issue where struct pages are left
7103 * uninitialized because the physical address range is not covered by
7104 * memblock.memory or memblock.reserved. That could happen when memblock
7105 * layout is manually configured via memmap=, or when the highest physical
7106 * address (max_pfn) does not end on a section boundary.
7108 static void __init init_unavailable_mem(void)
7110 phys_addr_t start, end;
7112 phys_addr_t next = 0;
7115 * Loop through unavailable ranges not covered by memblock.memory.
7118 for_each_mem_range(i, &start, &end) {
7120 pgcnt += init_unavailable_range(PFN_DOWN(next),
7126 * Early sections always have a fully populated memmap for the whole
7127 * section - see pfn_valid(). If the last section has holes at the
7128 * end and that section is marked "online", the memmap will be
7129 * considered initialized. Make sure that memmap has a well defined
7132 pgcnt += init_unavailable_range(PFN_DOWN(next),
7133 round_up(max_pfn, PAGES_PER_SECTION));
7136 * Struct pages that do not have backing memory. This could be because
7137 * firmware is using some of this memory, or for some other reasons.
7140 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7143 static inline void __init init_unavailable_mem(void)
7146 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7148 #if MAX_NUMNODES > 1
7150 * Figure out the number of possible node ids.
7152 void __init setup_nr_node_ids(void)
7154 unsigned int highest;
7156 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7157 nr_node_ids = highest + 1;
7162 * node_map_pfn_alignment - determine the maximum internode alignment
7164 * This function should be called after node map is populated and sorted.
7165 * It calculates the maximum power of two alignment which can distinguish
7168 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7169 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7170 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7171 * shifted, 1GiB is enough and this function will indicate so.
7173 * This is used to test whether pfn -> nid mapping of the chosen memory
7174 * model has fine enough granularity to avoid incorrect mapping for the
7175 * populated node map.
7177 * Return: the determined alignment in pfn's. 0 if there is no alignment
7178 * requirement (single node).
7180 unsigned long __init node_map_pfn_alignment(void)
7182 unsigned long accl_mask = 0, last_end = 0;
7183 unsigned long start, end, mask;
7184 int last_nid = NUMA_NO_NODE;
7187 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7188 if (!start || last_nid < 0 || last_nid == nid) {
7195 * Start with a mask granular enough to pin-point to the
7196 * start pfn and tick off bits one-by-one until it becomes
7197 * too coarse to separate the current node from the last.
7199 mask = ~((1 << __ffs(start)) - 1);
7200 while (mask && last_end <= (start & (mask << 1)))
7203 /* accumulate all internode masks */
7207 /* convert mask to number of pages */
7208 return ~accl_mask + 1;
7212 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7214 * Return: the minimum PFN based on information provided via
7215 * memblock_set_node().
7217 unsigned long __init find_min_pfn_with_active_regions(void)
7219 return PHYS_PFN(memblock_start_of_DRAM());
7223 * early_calculate_totalpages()
7224 * Sum pages in active regions for movable zone.
7225 * Populate N_MEMORY for calculating usable_nodes.
7227 static unsigned long __init early_calculate_totalpages(void)
7229 unsigned long totalpages = 0;
7230 unsigned long start_pfn, end_pfn;
7233 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7234 unsigned long pages = end_pfn - start_pfn;
7236 totalpages += pages;
7238 node_set_state(nid, N_MEMORY);
7244 * Find the PFN the Movable zone begins in each node. Kernel memory
7245 * is spread evenly between nodes as long as the nodes have enough
7246 * memory. When they don't, some nodes will have more kernelcore than
7249 static void __init find_zone_movable_pfns_for_nodes(void)
7252 unsigned long usable_startpfn;
7253 unsigned long kernelcore_node, kernelcore_remaining;
7254 /* save the state before borrow the nodemask */
7255 nodemask_t saved_node_state = node_states[N_MEMORY];
7256 unsigned long totalpages = early_calculate_totalpages();
7257 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7258 struct memblock_region *r;
7260 /* Need to find movable_zone earlier when movable_node is specified. */
7261 find_usable_zone_for_movable();
7264 * If movable_node is specified, ignore kernelcore and movablecore
7267 if (movable_node_is_enabled()) {
7268 for_each_mem_region(r) {
7269 if (!memblock_is_hotpluggable(r))
7272 nid = memblock_get_region_node(r);
7274 usable_startpfn = PFN_DOWN(r->base);
7275 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7276 min(usable_startpfn, zone_movable_pfn[nid]) :
7284 * If kernelcore=mirror is specified, ignore movablecore option
7286 if (mirrored_kernelcore) {
7287 bool mem_below_4gb_not_mirrored = false;
7289 for_each_mem_region(r) {
7290 if (memblock_is_mirror(r))
7293 nid = memblock_get_region_node(r);
7295 usable_startpfn = memblock_region_memory_base_pfn(r);
7297 if (usable_startpfn < 0x100000) {
7298 mem_below_4gb_not_mirrored = true;
7302 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7303 min(usable_startpfn, zone_movable_pfn[nid]) :
7307 if (mem_below_4gb_not_mirrored)
7308 pr_warn("This configuration results in unmirrored kernel memory.\n");
7314 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7315 * amount of necessary memory.
7317 if (required_kernelcore_percent)
7318 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7320 if (required_movablecore_percent)
7321 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7325 * If movablecore= was specified, calculate what size of
7326 * kernelcore that corresponds so that memory usable for
7327 * any allocation type is evenly spread. If both kernelcore
7328 * and movablecore are specified, then the value of kernelcore
7329 * will be used for required_kernelcore if it's greater than
7330 * what movablecore would have allowed.
7332 if (required_movablecore) {
7333 unsigned long corepages;
7336 * Round-up so that ZONE_MOVABLE is at least as large as what
7337 * was requested by the user
7339 required_movablecore =
7340 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7341 required_movablecore = min(totalpages, required_movablecore);
7342 corepages = totalpages - required_movablecore;
7344 required_kernelcore = max(required_kernelcore, corepages);
7348 * If kernelcore was not specified or kernelcore size is larger
7349 * than totalpages, there is no ZONE_MOVABLE.
7351 if (!required_kernelcore || required_kernelcore >= totalpages)
7354 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7355 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7358 /* Spread kernelcore memory as evenly as possible throughout nodes */
7359 kernelcore_node = required_kernelcore / usable_nodes;
7360 for_each_node_state(nid, N_MEMORY) {
7361 unsigned long start_pfn, end_pfn;
7364 * Recalculate kernelcore_node if the division per node
7365 * now exceeds what is necessary to satisfy the requested
7366 * amount of memory for the kernel
7368 if (required_kernelcore < kernelcore_node)
7369 kernelcore_node = required_kernelcore / usable_nodes;
7372 * As the map is walked, we track how much memory is usable
7373 * by the kernel using kernelcore_remaining. When it is
7374 * 0, the rest of the node is usable by ZONE_MOVABLE
7376 kernelcore_remaining = kernelcore_node;
7378 /* Go through each range of PFNs within this node */
7379 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7380 unsigned long size_pages;
7382 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7383 if (start_pfn >= end_pfn)
7386 /* Account for what is only usable for kernelcore */
7387 if (start_pfn < usable_startpfn) {
7388 unsigned long kernel_pages;
7389 kernel_pages = min(end_pfn, usable_startpfn)
7392 kernelcore_remaining -= min(kernel_pages,
7393 kernelcore_remaining);
7394 required_kernelcore -= min(kernel_pages,
7395 required_kernelcore);
7397 /* Continue if range is now fully accounted */
7398 if (end_pfn <= usable_startpfn) {
7401 * Push zone_movable_pfn to the end so
7402 * that if we have to rebalance
7403 * kernelcore across nodes, we will
7404 * not double account here
7406 zone_movable_pfn[nid] = end_pfn;
7409 start_pfn = usable_startpfn;
7413 * The usable PFN range for ZONE_MOVABLE is from
7414 * start_pfn->end_pfn. Calculate size_pages as the
7415 * number of pages used as kernelcore
7417 size_pages = end_pfn - start_pfn;
7418 if (size_pages > kernelcore_remaining)
7419 size_pages = kernelcore_remaining;
7420 zone_movable_pfn[nid] = start_pfn + size_pages;
7423 * Some kernelcore has been met, update counts and
7424 * break if the kernelcore for this node has been
7427 required_kernelcore -= min(required_kernelcore,
7429 kernelcore_remaining -= size_pages;
7430 if (!kernelcore_remaining)
7436 * If there is still required_kernelcore, we do another pass with one
7437 * less node in the count. This will push zone_movable_pfn[nid] further
7438 * along on the nodes that still have memory until kernelcore is
7442 if (usable_nodes && required_kernelcore > usable_nodes)
7446 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7447 for (nid = 0; nid < MAX_NUMNODES; nid++)
7448 zone_movable_pfn[nid] =
7449 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7452 /* restore the node_state */
7453 node_states[N_MEMORY] = saved_node_state;
7456 /* Any regular or high memory on that node ? */
7457 static void check_for_memory(pg_data_t *pgdat, int nid)
7459 enum zone_type zone_type;
7461 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7462 struct zone *zone = &pgdat->node_zones[zone_type];
7463 if (populated_zone(zone)) {
7464 if (IS_ENABLED(CONFIG_HIGHMEM))
7465 node_set_state(nid, N_HIGH_MEMORY);
7466 if (zone_type <= ZONE_NORMAL)
7467 node_set_state(nid, N_NORMAL_MEMORY);
7474 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7475 * such cases we allow max_zone_pfn sorted in the descending order
7477 bool __weak arch_has_descending_max_zone_pfns(void)
7483 * free_area_init - Initialise all pg_data_t and zone data
7484 * @max_zone_pfn: an array of max PFNs for each zone
7486 * This will call free_area_init_node() for each active node in the system.
7487 * Using the page ranges provided by memblock_set_node(), the size of each
7488 * zone in each node and their holes is calculated. If the maximum PFN
7489 * between two adjacent zones match, it is assumed that the zone is empty.
7490 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7491 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7492 * starts where the previous one ended. For example, ZONE_DMA32 starts
7493 * at arch_max_dma_pfn.
7495 void __init free_area_init(unsigned long *max_zone_pfn)
7497 unsigned long start_pfn, end_pfn;
7501 /* Record where the zone boundaries are */
7502 memset(arch_zone_lowest_possible_pfn, 0,
7503 sizeof(arch_zone_lowest_possible_pfn));
7504 memset(arch_zone_highest_possible_pfn, 0,
7505 sizeof(arch_zone_highest_possible_pfn));
7507 start_pfn = find_min_pfn_with_active_regions();
7508 descending = arch_has_descending_max_zone_pfns();
7510 for (i = 0; i < MAX_NR_ZONES; i++) {
7512 zone = MAX_NR_ZONES - i - 1;
7516 if (zone == ZONE_MOVABLE)
7519 end_pfn = max(max_zone_pfn[zone], start_pfn);
7520 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7521 arch_zone_highest_possible_pfn[zone] = end_pfn;
7523 start_pfn = end_pfn;
7526 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7527 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7528 find_zone_movable_pfns_for_nodes();
7530 /* Print out the zone ranges */
7531 pr_info("Zone ranges:\n");
7532 for (i = 0; i < MAX_NR_ZONES; i++) {
7533 if (i == ZONE_MOVABLE)
7535 pr_info(" %-8s ", zone_names[i]);
7536 if (arch_zone_lowest_possible_pfn[i] ==
7537 arch_zone_highest_possible_pfn[i])
7540 pr_cont("[mem %#018Lx-%#018Lx]\n",
7541 (u64)arch_zone_lowest_possible_pfn[i]
7543 ((u64)arch_zone_highest_possible_pfn[i]
7544 << PAGE_SHIFT) - 1);
7547 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7548 pr_info("Movable zone start for each node\n");
7549 for (i = 0; i < MAX_NUMNODES; i++) {
7550 if (zone_movable_pfn[i])
7551 pr_info(" Node %d: %#018Lx\n", i,
7552 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7556 * Print out the early node map, and initialize the
7557 * subsection-map relative to active online memory ranges to
7558 * enable future "sub-section" extensions of the memory map.
7560 pr_info("Early memory node ranges\n");
7561 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7562 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7563 (u64)start_pfn << PAGE_SHIFT,
7564 ((u64)end_pfn << PAGE_SHIFT) - 1);
7565 subsection_map_init(start_pfn, end_pfn - start_pfn);
7568 /* Initialise every node */
7569 mminit_verify_pageflags_layout();
7570 setup_nr_node_ids();
7571 init_unavailable_mem();
7572 for_each_online_node(nid) {
7573 pg_data_t *pgdat = NODE_DATA(nid);
7574 free_area_init_node(nid);
7576 /* Any memory on that node */
7577 if (pgdat->node_present_pages)
7578 node_set_state(nid, N_MEMORY);
7579 check_for_memory(pgdat, nid);
7583 static int __init cmdline_parse_core(char *p, unsigned long *core,
7584 unsigned long *percent)
7586 unsigned long long coremem;
7592 /* Value may be a percentage of total memory, otherwise bytes */
7593 coremem = simple_strtoull(p, &endptr, 0);
7594 if (*endptr == '%') {
7595 /* Paranoid check for percent values greater than 100 */
7596 WARN_ON(coremem > 100);
7600 coremem = memparse(p, &p);
7601 /* Paranoid check that UL is enough for the coremem value */
7602 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7604 *core = coremem >> PAGE_SHIFT;
7611 * kernelcore=size sets the amount of memory for use for allocations that
7612 * cannot be reclaimed or migrated.
7614 static int __init cmdline_parse_kernelcore(char *p)
7616 /* parse kernelcore=mirror */
7617 if (parse_option_str(p, "mirror")) {
7618 mirrored_kernelcore = true;
7622 return cmdline_parse_core(p, &required_kernelcore,
7623 &required_kernelcore_percent);
7627 * movablecore=size sets the amount of memory for use for allocations that
7628 * can be reclaimed or migrated.
7630 static int __init cmdline_parse_movablecore(char *p)
7632 return cmdline_parse_core(p, &required_movablecore,
7633 &required_movablecore_percent);
7636 early_param("kernelcore", cmdline_parse_kernelcore);
7637 early_param("movablecore", cmdline_parse_movablecore);
7639 void adjust_managed_page_count(struct page *page, long count)
7641 atomic_long_add(count, &page_zone(page)->managed_pages);
7642 totalram_pages_add(count);
7643 #ifdef CONFIG_HIGHMEM
7644 if (PageHighMem(page))
7645 totalhigh_pages_add(count);
7648 EXPORT_SYMBOL(adjust_managed_page_count);
7650 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7653 unsigned long pages = 0;
7655 start = (void *)PAGE_ALIGN((unsigned long)start);
7656 end = (void *)((unsigned long)end & PAGE_MASK);
7657 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7658 struct page *page = virt_to_page(pos);
7659 void *direct_map_addr;
7662 * 'direct_map_addr' might be different from 'pos'
7663 * because some architectures' virt_to_page()
7664 * work with aliases. Getting the direct map
7665 * address ensures that we get a _writeable_
7666 * alias for the memset().
7668 direct_map_addr = page_address(page);
7669 if ((unsigned int)poison <= 0xFF)
7670 memset(direct_map_addr, poison, PAGE_SIZE);
7672 free_reserved_page(page);
7676 pr_info("Freeing %s memory: %ldK\n",
7677 s, pages << (PAGE_SHIFT - 10));
7682 #ifdef CONFIG_HIGHMEM
7683 void free_highmem_page(struct page *page)
7685 __free_reserved_page(page);
7686 totalram_pages_inc();
7687 atomic_long_inc(&page_zone(page)->managed_pages);
7688 totalhigh_pages_inc();
7693 void __init mem_init_print_info(const char *str)
7695 unsigned long physpages, codesize, datasize, rosize, bss_size;
7696 unsigned long init_code_size, init_data_size;
7698 physpages = get_num_physpages();
7699 codesize = _etext - _stext;
7700 datasize = _edata - _sdata;
7701 rosize = __end_rodata - __start_rodata;
7702 bss_size = __bss_stop - __bss_start;
7703 init_data_size = __init_end - __init_begin;
7704 init_code_size = _einittext - _sinittext;
7707 * Detect special cases and adjust section sizes accordingly:
7708 * 1) .init.* may be embedded into .data sections
7709 * 2) .init.text.* may be out of [__init_begin, __init_end],
7710 * please refer to arch/tile/kernel/vmlinux.lds.S.
7711 * 3) .rodata.* may be embedded into .text or .data sections.
7713 #define adj_init_size(start, end, size, pos, adj) \
7715 if (start <= pos && pos < end && size > adj) \
7719 adj_init_size(__init_begin, __init_end, init_data_size,
7720 _sinittext, init_code_size);
7721 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7722 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7723 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7724 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7726 #undef adj_init_size
7728 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7729 #ifdef CONFIG_HIGHMEM
7733 nr_free_pages() << (PAGE_SHIFT - 10),
7734 physpages << (PAGE_SHIFT - 10),
7735 codesize >> 10, datasize >> 10, rosize >> 10,
7736 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7737 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7738 totalcma_pages << (PAGE_SHIFT - 10),
7739 #ifdef CONFIG_HIGHMEM
7740 totalhigh_pages() << (PAGE_SHIFT - 10),
7742 str ? ", " : "", str ? str : "");
7746 * set_dma_reserve - set the specified number of pages reserved in the first zone
7747 * @new_dma_reserve: The number of pages to mark reserved
7749 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7750 * In the DMA zone, a significant percentage may be consumed by kernel image
7751 * and other unfreeable allocations which can skew the watermarks badly. This
7752 * function may optionally be used to account for unfreeable pages in the
7753 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7754 * smaller per-cpu batchsize.
7756 void __init set_dma_reserve(unsigned long new_dma_reserve)
7758 dma_reserve = new_dma_reserve;
7761 static int page_alloc_cpu_dead(unsigned int cpu)
7764 lru_add_drain_cpu(cpu);
7768 * Spill the event counters of the dead processor
7769 * into the current processors event counters.
7770 * This artificially elevates the count of the current
7773 vm_events_fold_cpu(cpu);
7776 * Zero the differential counters of the dead processor
7777 * so that the vm statistics are consistent.
7779 * This is only okay since the processor is dead and cannot
7780 * race with what we are doing.
7782 cpu_vm_stats_fold(cpu);
7787 int hashdist = HASHDIST_DEFAULT;
7789 static int __init set_hashdist(char *str)
7793 hashdist = simple_strtoul(str, &str, 0);
7796 __setup("hashdist=", set_hashdist);
7799 void __init page_alloc_init(void)
7804 if (num_node_state(N_MEMORY) == 1)
7808 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7809 "mm/page_alloc:dead", NULL,
7810 page_alloc_cpu_dead);
7815 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7816 * or min_free_kbytes changes.
7818 static void calculate_totalreserve_pages(void)
7820 struct pglist_data *pgdat;
7821 unsigned long reserve_pages = 0;
7822 enum zone_type i, j;
7824 for_each_online_pgdat(pgdat) {
7826 pgdat->totalreserve_pages = 0;
7828 for (i = 0; i < MAX_NR_ZONES; i++) {
7829 struct zone *zone = pgdat->node_zones + i;
7831 unsigned long managed_pages = zone_managed_pages(zone);
7833 /* Find valid and maximum lowmem_reserve in the zone */
7834 for (j = i; j < MAX_NR_ZONES; j++) {
7835 if (zone->lowmem_reserve[j] > max)
7836 max = zone->lowmem_reserve[j];
7839 /* we treat the high watermark as reserved pages. */
7840 max += high_wmark_pages(zone);
7842 if (max > managed_pages)
7843 max = managed_pages;
7845 pgdat->totalreserve_pages += max;
7847 reserve_pages += max;
7850 totalreserve_pages = reserve_pages;
7854 * setup_per_zone_lowmem_reserve - called whenever
7855 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7856 * has a correct pages reserved value, so an adequate number of
7857 * pages are left in the zone after a successful __alloc_pages().
7859 static void setup_per_zone_lowmem_reserve(void)
7861 struct pglist_data *pgdat;
7862 enum zone_type i, j;
7864 for_each_online_pgdat(pgdat) {
7865 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7866 struct zone *zone = &pgdat->node_zones[i];
7867 int ratio = sysctl_lowmem_reserve_ratio[i];
7868 bool clear = !ratio || !zone_managed_pages(zone);
7869 unsigned long managed_pages = 0;
7871 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7873 zone->lowmem_reserve[j] = 0;
7875 struct zone *upper_zone = &pgdat->node_zones[j];
7877 managed_pages += zone_managed_pages(upper_zone);
7878 zone->lowmem_reserve[j] = managed_pages / ratio;
7884 /* update totalreserve_pages */
7885 calculate_totalreserve_pages();
7888 static void __setup_per_zone_wmarks(void)
7890 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7891 unsigned long lowmem_pages = 0;
7893 unsigned long flags;
7895 /* Calculate total number of !ZONE_HIGHMEM pages */
7896 for_each_zone(zone) {
7897 if (!is_highmem(zone))
7898 lowmem_pages += zone_managed_pages(zone);
7901 for_each_zone(zone) {
7904 spin_lock_irqsave(&zone->lock, flags);
7905 tmp = (u64)pages_min * zone_managed_pages(zone);
7906 do_div(tmp, lowmem_pages);
7907 if (is_highmem(zone)) {
7909 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7910 * need highmem pages, so cap pages_min to a small
7913 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7914 * deltas control async page reclaim, and so should
7915 * not be capped for highmem.
7917 unsigned long min_pages;
7919 min_pages = zone_managed_pages(zone) / 1024;
7920 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7921 zone->_watermark[WMARK_MIN] = min_pages;
7924 * If it's a lowmem zone, reserve a number of pages
7925 * proportionate to the zone's size.
7927 zone->_watermark[WMARK_MIN] = tmp;
7931 * Set the kswapd watermarks distance according to the
7932 * scale factor in proportion to available memory, but
7933 * ensure a minimum size on small systems.
7935 tmp = max_t(u64, tmp >> 2,
7936 mult_frac(zone_managed_pages(zone),
7937 watermark_scale_factor, 10000));
7939 zone->watermark_boost = 0;
7940 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7941 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7943 spin_unlock_irqrestore(&zone->lock, flags);
7946 /* update totalreserve_pages */
7947 calculate_totalreserve_pages();
7951 * setup_per_zone_wmarks - called when min_free_kbytes changes
7952 * or when memory is hot-{added|removed}
7954 * Ensures that the watermark[min,low,high] values for each zone are set
7955 * correctly with respect to min_free_kbytes.
7957 void setup_per_zone_wmarks(void)
7959 static DEFINE_SPINLOCK(lock);
7962 __setup_per_zone_wmarks();
7967 * Initialise min_free_kbytes.
7969 * For small machines we want it small (128k min). For large machines
7970 * we want it large (256MB max). But it is not linear, because network
7971 * bandwidth does not increase linearly with machine size. We use
7973 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7974 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7990 int __meminit init_per_zone_wmark_min(void)
7992 unsigned long lowmem_kbytes;
7993 int new_min_free_kbytes;
7995 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7996 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7998 if (new_min_free_kbytes > user_min_free_kbytes) {
7999 min_free_kbytes = new_min_free_kbytes;
8000 if (min_free_kbytes < 128)
8001 min_free_kbytes = 128;
8002 if (min_free_kbytes > 262144)
8003 min_free_kbytes = 262144;
8005 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8006 new_min_free_kbytes, user_min_free_kbytes);
8008 setup_per_zone_wmarks();
8009 refresh_zone_stat_thresholds();
8010 setup_per_zone_lowmem_reserve();
8013 setup_min_unmapped_ratio();
8014 setup_min_slab_ratio();
8017 khugepaged_min_free_kbytes_update();
8021 postcore_initcall(init_per_zone_wmark_min)
8024 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8025 * that we can call two helper functions whenever min_free_kbytes
8028 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8029 void *buffer, size_t *length, loff_t *ppos)
8033 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8038 user_min_free_kbytes = min_free_kbytes;
8039 setup_per_zone_wmarks();
8044 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8045 void *buffer, size_t *length, loff_t *ppos)
8049 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8054 setup_per_zone_wmarks();
8060 static void setup_min_unmapped_ratio(void)
8065 for_each_online_pgdat(pgdat)
8066 pgdat->min_unmapped_pages = 0;
8069 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8070 sysctl_min_unmapped_ratio) / 100;
8074 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8075 void *buffer, size_t *length, loff_t *ppos)
8079 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8083 setup_min_unmapped_ratio();
8088 static void setup_min_slab_ratio(void)
8093 for_each_online_pgdat(pgdat)
8094 pgdat->min_slab_pages = 0;
8097 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8098 sysctl_min_slab_ratio) / 100;
8101 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8102 void *buffer, size_t *length, loff_t *ppos)
8106 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8110 setup_min_slab_ratio();
8117 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8118 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8119 * whenever sysctl_lowmem_reserve_ratio changes.
8121 * The reserve ratio obviously has absolutely no relation with the
8122 * minimum watermarks. The lowmem reserve ratio can only make sense
8123 * if in function of the boot time zone sizes.
8125 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8126 void *buffer, size_t *length, loff_t *ppos)
8130 proc_dointvec_minmax(table, write, buffer, length, ppos);
8132 for (i = 0; i < MAX_NR_ZONES; i++) {
8133 if (sysctl_lowmem_reserve_ratio[i] < 1)
8134 sysctl_lowmem_reserve_ratio[i] = 0;
8137 setup_per_zone_lowmem_reserve();
8142 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8143 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8144 * pagelist can have before it gets flushed back to buddy allocator.
8146 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8147 void *buffer, size_t *length, loff_t *ppos)
8150 int old_percpu_pagelist_fraction;
8153 mutex_lock(&pcp_batch_high_lock);
8154 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8156 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8157 if (!write || ret < 0)
8160 /* Sanity checking to avoid pcp imbalance */
8161 if (percpu_pagelist_fraction &&
8162 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8163 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8169 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8172 for_each_populated_zone(zone)
8173 zone_set_pageset_high_and_batch(zone);
8175 mutex_unlock(&pcp_batch_high_lock);
8179 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8181 * Returns the number of pages that arch has reserved but
8182 * is not known to alloc_large_system_hash().
8184 static unsigned long __init arch_reserved_kernel_pages(void)
8191 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8192 * machines. As memory size is increased the scale is also increased but at
8193 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8194 * quadruples the scale is increased by one, which means the size of hash table
8195 * only doubles, instead of quadrupling as well.
8196 * Because 32-bit systems cannot have large physical memory, where this scaling
8197 * makes sense, it is disabled on such platforms.
8199 #if __BITS_PER_LONG > 32
8200 #define ADAPT_SCALE_BASE (64ul << 30)
8201 #define ADAPT_SCALE_SHIFT 2
8202 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8206 * allocate a large system hash table from bootmem
8207 * - it is assumed that the hash table must contain an exact power-of-2
8208 * quantity of entries
8209 * - limit is the number of hash buckets, not the total allocation size
8211 void *__init alloc_large_system_hash(const char *tablename,
8212 unsigned long bucketsize,
8213 unsigned long numentries,
8216 unsigned int *_hash_shift,
8217 unsigned int *_hash_mask,
8218 unsigned long low_limit,
8219 unsigned long high_limit)
8221 unsigned long long max = high_limit;
8222 unsigned long log2qty, size;
8227 /* allow the kernel cmdline to have a say */
8229 /* round applicable memory size up to nearest megabyte */
8230 numentries = nr_kernel_pages;
8231 numentries -= arch_reserved_kernel_pages();
8233 /* It isn't necessary when PAGE_SIZE >= 1MB */
8234 if (PAGE_SHIFT < 20)
8235 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8237 #if __BITS_PER_LONG > 32
8239 unsigned long adapt;
8241 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8242 adapt <<= ADAPT_SCALE_SHIFT)
8247 /* limit to 1 bucket per 2^scale bytes of low memory */
8248 if (scale > PAGE_SHIFT)
8249 numentries >>= (scale - PAGE_SHIFT);
8251 numentries <<= (PAGE_SHIFT - scale);
8253 /* Make sure we've got at least a 0-order allocation.. */
8254 if (unlikely(flags & HASH_SMALL)) {
8255 /* Makes no sense without HASH_EARLY */
8256 WARN_ON(!(flags & HASH_EARLY));
8257 if (!(numentries >> *_hash_shift)) {
8258 numentries = 1UL << *_hash_shift;
8259 BUG_ON(!numentries);
8261 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8262 numentries = PAGE_SIZE / bucketsize;
8264 numentries = roundup_pow_of_two(numentries);
8266 /* limit allocation size to 1/16 total memory by default */
8268 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8269 do_div(max, bucketsize);
8271 max = min(max, 0x80000000ULL);
8273 if (numentries < low_limit)
8274 numentries = low_limit;
8275 if (numentries > max)
8278 log2qty = ilog2(numentries);
8280 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8283 size = bucketsize << log2qty;
8284 if (flags & HASH_EARLY) {
8285 if (flags & HASH_ZERO)
8286 table = memblock_alloc(size, SMP_CACHE_BYTES);
8288 table = memblock_alloc_raw(size,
8290 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8291 table = __vmalloc(size, gfp_flags);
8295 * If bucketsize is not a power-of-two, we may free
8296 * some pages at the end of hash table which
8297 * alloc_pages_exact() automatically does
8299 table = alloc_pages_exact(size, gfp_flags);
8300 kmemleak_alloc(table, size, 1, gfp_flags);
8302 } while (!table && size > PAGE_SIZE && --log2qty);
8305 panic("Failed to allocate %s hash table\n", tablename);
8307 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8308 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8309 virt ? "vmalloc" : "linear");
8312 *_hash_shift = log2qty;
8314 *_hash_mask = (1 << log2qty) - 1;
8320 * This function checks whether pageblock includes unmovable pages or not.
8322 * PageLRU check without isolation or lru_lock could race so that
8323 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8324 * check without lock_page also may miss some movable non-lru pages at
8325 * race condition. So you can't expect this function should be exact.
8327 * Returns a page without holding a reference. If the caller wants to
8328 * dereference that page (e.g., dumping), it has to make sure that it
8329 * cannot get removed (e.g., via memory unplug) concurrently.
8332 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8333 int migratetype, int flags)
8335 unsigned long iter = 0;
8336 unsigned long pfn = page_to_pfn(page);
8337 unsigned long offset = pfn % pageblock_nr_pages;
8339 if (is_migrate_cma_page(page)) {
8341 * CMA allocations (alloc_contig_range) really need to mark
8342 * isolate CMA pageblocks even when they are not movable in fact
8343 * so consider them movable here.
8345 if (is_migrate_cma(migratetype))
8351 for (; iter < pageblock_nr_pages - offset; iter++) {
8352 if (!pfn_valid_within(pfn + iter))
8355 page = pfn_to_page(pfn + iter);
8358 * Both, bootmem allocations and memory holes are marked
8359 * PG_reserved and are unmovable. We can even have unmovable
8360 * allocations inside ZONE_MOVABLE, for example when
8361 * specifying "movablecore".
8363 if (PageReserved(page))
8367 * If the zone is movable and we have ruled out all reserved
8368 * pages then it should be reasonably safe to assume the rest
8371 if (zone_idx(zone) == ZONE_MOVABLE)
8375 * Hugepages are not in LRU lists, but they're movable.
8376 * THPs are on the LRU, but need to be counted as #small pages.
8377 * We need not scan over tail pages because we don't
8378 * handle each tail page individually in migration.
8380 if (PageHuge(page) || PageTransCompound(page)) {
8381 struct page *head = compound_head(page);
8382 unsigned int skip_pages;
8384 if (PageHuge(page)) {
8385 if (!hugepage_migration_supported(page_hstate(head)))
8387 } else if (!PageLRU(head) && !__PageMovable(head)) {
8391 skip_pages = compound_nr(head) - (page - head);
8392 iter += skip_pages - 1;
8397 * We can't use page_count without pin a page
8398 * because another CPU can free compound page.
8399 * This check already skips compound tails of THP
8400 * because their page->_refcount is zero at all time.
8402 if (!page_ref_count(page)) {
8403 if (PageBuddy(page))
8404 iter += (1 << buddy_order(page)) - 1;
8409 * The HWPoisoned page may be not in buddy system, and
8410 * page_count() is not 0.
8412 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8416 * We treat all PageOffline() pages as movable when offlining
8417 * to give drivers a chance to decrement their reference count
8418 * in MEM_GOING_OFFLINE in order to indicate that these pages
8419 * can be offlined as there are no direct references anymore.
8420 * For actually unmovable PageOffline() where the driver does
8421 * not support this, we will fail later when trying to actually
8422 * move these pages that still have a reference count > 0.
8423 * (false negatives in this function only)
8425 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8428 if (__PageMovable(page) || PageLRU(page))
8432 * If there are RECLAIMABLE pages, we need to check
8433 * it. But now, memory offline itself doesn't call
8434 * shrink_node_slabs() and it still to be fixed.
8441 #ifdef CONFIG_CONTIG_ALLOC
8442 static unsigned long pfn_max_align_down(unsigned long pfn)
8444 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8445 pageblock_nr_pages) - 1);
8448 static unsigned long pfn_max_align_up(unsigned long pfn)
8450 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8451 pageblock_nr_pages));
8454 /* [start, end) must belong to a single zone. */
8455 static int __alloc_contig_migrate_range(struct compact_control *cc,
8456 unsigned long start, unsigned long end)
8458 /* This function is based on compact_zone() from compaction.c. */
8459 unsigned int nr_reclaimed;
8460 unsigned long pfn = start;
8461 unsigned int tries = 0;
8463 struct migration_target_control mtc = {
8464 .nid = zone_to_nid(cc->zone),
8465 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8470 while (pfn < end || !list_empty(&cc->migratepages)) {
8471 if (fatal_signal_pending(current)) {
8476 if (list_empty(&cc->migratepages)) {
8477 cc->nr_migratepages = 0;
8478 pfn = isolate_migratepages_range(cc, pfn, end);
8484 } else if (++tries == 5) {
8485 ret = ret < 0 ? ret : -EBUSY;
8489 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8491 cc->nr_migratepages -= nr_reclaimed;
8493 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8494 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8497 putback_movable_pages(&cc->migratepages);
8504 * alloc_contig_range() -- tries to allocate given range of pages
8505 * @start: start PFN to allocate
8506 * @end: one-past-the-last PFN to allocate
8507 * @migratetype: migratetype of the underlaying pageblocks (either
8508 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8509 * in range must have the same migratetype and it must
8510 * be either of the two.
8511 * @gfp_mask: GFP mask to use during compaction
8513 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8514 * aligned. The PFN range must belong to a single zone.
8516 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8517 * pageblocks in the range. Once isolated, the pageblocks should not
8518 * be modified by others.
8520 * Return: zero on success or negative error code. On success all
8521 * pages which PFN is in [start, end) are allocated for the caller and
8522 * need to be freed with free_contig_range().
8524 int alloc_contig_range(unsigned long start, unsigned long end,
8525 unsigned migratetype, gfp_t gfp_mask)
8527 unsigned long outer_start, outer_end;
8531 struct compact_control cc = {
8532 .nr_migratepages = 0,
8534 .zone = page_zone(pfn_to_page(start)),
8535 .mode = MIGRATE_SYNC,
8536 .ignore_skip_hint = true,
8537 .no_set_skip_hint = true,
8538 .gfp_mask = current_gfp_context(gfp_mask),
8539 .alloc_contig = true,
8541 INIT_LIST_HEAD(&cc.migratepages);
8544 * What we do here is we mark all pageblocks in range as
8545 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8546 * have different sizes, and due to the way page allocator
8547 * work, we align the range to biggest of the two pages so
8548 * that page allocator won't try to merge buddies from
8549 * different pageblocks and change MIGRATE_ISOLATE to some
8550 * other migration type.
8552 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8553 * migrate the pages from an unaligned range (ie. pages that
8554 * we are interested in). This will put all the pages in
8555 * range back to page allocator as MIGRATE_ISOLATE.
8557 * When this is done, we take the pages in range from page
8558 * allocator removing them from the buddy system. This way
8559 * page allocator will never consider using them.
8561 * This lets us mark the pageblocks back as
8562 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8563 * aligned range but not in the unaligned, original range are
8564 * put back to page allocator so that buddy can use them.
8567 ret = start_isolate_page_range(pfn_max_align_down(start),
8568 pfn_max_align_up(end), migratetype, 0);
8572 drain_all_pages(cc.zone);
8575 * In case of -EBUSY, we'd like to know which page causes problem.
8576 * So, just fall through. test_pages_isolated() has a tracepoint
8577 * which will report the busy page.
8579 * It is possible that busy pages could become available before
8580 * the call to test_pages_isolated, and the range will actually be
8581 * allocated. So, if we fall through be sure to clear ret so that
8582 * -EBUSY is not accidentally used or returned to caller.
8584 ret = __alloc_contig_migrate_range(&cc, start, end);
8585 if (ret && ret != -EBUSY)
8590 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8591 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8592 * more, all pages in [start, end) are free in page allocator.
8593 * What we are going to do is to allocate all pages from
8594 * [start, end) (that is remove them from page allocator).
8596 * The only problem is that pages at the beginning and at the
8597 * end of interesting range may be not aligned with pages that
8598 * page allocator holds, ie. they can be part of higher order
8599 * pages. Because of this, we reserve the bigger range and
8600 * once this is done free the pages we are not interested in.
8602 * We don't have to hold zone->lock here because the pages are
8603 * isolated thus they won't get removed from buddy.
8606 lru_add_drain_all();
8609 outer_start = start;
8610 while (!PageBuddy(pfn_to_page(outer_start))) {
8611 if (++order >= MAX_ORDER) {
8612 outer_start = start;
8615 outer_start &= ~0UL << order;
8618 if (outer_start != start) {
8619 order = buddy_order(pfn_to_page(outer_start));
8622 * outer_start page could be small order buddy page and
8623 * it doesn't include start page. Adjust outer_start
8624 * in this case to report failed page properly
8625 * on tracepoint in test_pages_isolated()
8627 if (outer_start + (1UL << order) <= start)
8628 outer_start = start;
8631 /* Make sure the range is really isolated. */
8632 if (test_pages_isolated(outer_start, end, 0)) {
8633 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8634 __func__, outer_start, end);
8639 /* Grab isolated pages from freelists. */
8640 outer_end = isolate_freepages_range(&cc, outer_start, end);
8646 /* Free head and tail (if any) */
8647 if (start != outer_start)
8648 free_contig_range(outer_start, start - outer_start);
8649 if (end != outer_end)
8650 free_contig_range(end, outer_end - end);
8653 undo_isolate_page_range(pfn_max_align_down(start),
8654 pfn_max_align_up(end), migratetype);
8657 EXPORT_SYMBOL(alloc_contig_range);
8659 static int __alloc_contig_pages(unsigned long start_pfn,
8660 unsigned long nr_pages, gfp_t gfp_mask)
8662 unsigned long end_pfn = start_pfn + nr_pages;
8664 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8668 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8669 unsigned long nr_pages)
8671 unsigned long i, end_pfn = start_pfn + nr_pages;
8674 for (i = start_pfn; i < end_pfn; i++) {
8675 page = pfn_to_online_page(i);
8679 if (page_zone(page) != z)
8682 if (PageReserved(page))
8685 if (page_count(page) > 0)
8694 static bool zone_spans_last_pfn(const struct zone *zone,
8695 unsigned long start_pfn, unsigned long nr_pages)
8697 unsigned long last_pfn = start_pfn + nr_pages - 1;
8699 return zone_spans_pfn(zone, last_pfn);
8703 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8704 * @nr_pages: Number of contiguous pages to allocate
8705 * @gfp_mask: GFP mask to limit search and used during compaction
8707 * @nodemask: Mask for other possible nodes
8709 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8710 * on an applicable zonelist to find a contiguous pfn range which can then be
8711 * tried for allocation with alloc_contig_range(). This routine is intended
8712 * for allocation requests which can not be fulfilled with the buddy allocator.
8714 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8715 * power of two then the alignment is guaranteed to be to the given nr_pages
8716 * (e.g. 1GB request would be aligned to 1GB).
8718 * Allocated pages can be freed with free_contig_range() or by manually calling
8719 * __free_page() on each allocated page.
8721 * Return: pointer to contiguous pages on success, or NULL if not successful.
8723 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8724 int nid, nodemask_t *nodemask)
8726 unsigned long ret, pfn, flags;
8727 struct zonelist *zonelist;
8731 zonelist = node_zonelist(nid, gfp_mask);
8732 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8733 gfp_zone(gfp_mask), nodemask) {
8734 spin_lock_irqsave(&zone->lock, flags);
8736 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8737 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8738 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8740 * We release the zone lock here because
8741 * alloc_contig_range() will also lock the zone
8742 * at some point. If there's an allocation
8743 * spinning on this lock, it may win the race
8744 * and cause alloc_contig_range() to fail...
8746 spin_unlock_irqrestore(&zone->lock, flags);
8747 ret = __alloc_contig_pages(pfn, nr_pages,
8750 return pfn_to_page(pfn);
8751 spin_lock_irqsave(&zone->lock, flags);
8755 spin_unlock_irqrestore(&zone->lock, flags);
8759 #endif /* CONFIG_CONTIG_ALLOC */
8761 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8763 unsigned int count = 0;
8765 for (; nr_pages--; pfn++) {
8766 struct page *page = pfn_to_page(pfn);
8768 count += page_count(page) != 1;
8771 WARN(count != 0, "%d pages are still in use!\n", count);
8773 EXPORT_SYMBOL(free_contig_range);
8776 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8777 * page high values need to be recalulated.
8779 void __meminit zone_pcp_update(struct zone *zone)
8781 mutex_lock(&pcp_batch_high_lock);
8782 zone_set_pageset_high_and_batch(zone);
8783 mutex_unlock(&pcp_batch_high_lock);
8787 * Effectively disable pcplists for the zone by setting the high limit to 0
8788 * and draining all cpus. A concurrent page freeing on another CPU that's about
8789 * to put the page on pcplist will either finish before the drain and the page
8790 * will be drained, or observe the new high limit and skip the pcplist.
8792 * Must be paired with a call to zone_pcp_enable().
8794 void zone_pcp_disable(struct zone *zone)
8796 mutex_lock(&pcp_batch_high_lock);
8797 __zone_set_pageset_high_and_batch(zone, 0, 1);
8798 __drain_all_pages(zone, true);
8801 void zone_pcp_enable(struct zone *zone)
8803 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8804 mutex_unlock(&pcp_batch_high_lock);
8807 void zone_pcp_reset(struct zone *zone)
8809 unsigned long flags;
8811 struct per_cpu_pageset *pset;
8813 /* avoid races with drain_pages() */
8814 local_irq_save(flags);
8815 if (zone->pageset != &boot_pageset) {
8816 for_each_online_cpu(cpu) {
8817 pset = per_cpu_ptr(zone->pageset, cpu);
8818 drain_zonestat(zone, pset);
8820 free_percpu(zone->pageset);
8821 zone->pageset = &boot_pageset;
8823 local_irq_restore(flags);
8826 #ifdef CONFIG_MEMORY_HOTREMOVE
8828 * All pages in the range must be in a single zone, must not contain holes,
8829 * must span full sections, and must be isolated before calling this function.
8831 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8833 unsigned long pfn = start_pfn;
8837 unsigned long flags;
8839 offline_mem_sections(pfn, end_pfn);
8840 zone = page_zone(pfn_to_page(pfn));
8841 spin_lock_irqsave(&zone->lock, flags);
8842 while (pfn < end_pfn) {
8843 page = pfn_to_page(pfn);
8845 * The HWPoisoned page may be not in buddy system, and
8846 * page_count() is not 0.
8848 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8853 * At this point all remaining PageOffline() pages have a
8854 * reference count of 0 and can simply be skipped.
8856 if (PageOffline(page)) {
8857 BUG_ON(page_count(page));
8858 BUG_ON(PageBuddy(page));
8863 BUG_ON(page_count(page));
8864 BUG_ON(!PageBuddy(page));
8865 order = buddy_order(page);
8866 del_page_from_free_list(page, zone, order);
8867 pfn += (1 << order);
8869 spin_unlock_irqrestore(&zone->lock, flags);
8873 bool is_free_buddy_page(struct page *page)
8875 struct zone *zone = page_zone(page);
8876 unsigned long pfn = page_to_pfn(page);
8877 unsigned long flags;
8880 spin_lock_irqsave(&zone->lock, flags);
8881 for (order = 0; order < MAX_ORDER; order++) {
8882 struct page *page_head = page - (pfn & ((1 << order) - 1));
8884 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8887 spin_unlock_irqrestore(&zone->lock, flags);
8889 return order < MAX_ORDER;
8892 #ifdef CONFIG_MEMORY_FAILURE
8894 * Break down a higher-order page in sub-pages, and keep our target out of
8897 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8898 struct page *target, int low, int high,
8901 unsigned long size = 1 << high;
8902 struct page *current_buddy, *next_page;
8904 while (high > low) {
8908 if (target >= &page[size]) {
8909 next_page = page + size;
8910 current_buddy = page;
8913 current_buddy = page + size;
8916 if (set_page_guard(zone, current_buddy, high, migratetype))
8919 if (current_buddy != target) {
8920 add_to_free_list(current_buddy, zone, high, migratetype);
8921 set_buddy_order(current_buddy, high);
8928 * Take a page that will be marked as poisoned off the buddy allocator.
8930 bool take_page_off_buddy(struct page *page)
8932 struct zone *zone = page_zone(page);
8933 unsigned long pfn = page_to_pfn(page);
8934 unsigned long flags;
8938 spin_lock_irqsave(&zone->lock, flags);
8939 for (order = 0; order < MAX_ORDER; order++) {
8940 struct page *page_head = page - (pfn & ((1 << order) - 1));
8941 int page_order = buddy_order(page_head);
8943 if (PageBuddy(page_head) && page_order >= order) {
8944 unsigned long pfn_head = page_to_pfn(page_head);
8945 int migratetype = get_pfnblock_migratetype(page_head,
8948 del_page_from_free_list(page_head, zone, page_order);
8949 break_down_buddy_pages(zone, page_head, page, 0,
8950 page_order, migratetype);
8954 if (page_count(page_head) > 0)
8957 spin_unlock_irqrestore(&zone->lock, flags);