2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.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/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.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/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 /* Protect totalram_pages and zone->managed_pages */
126 static DEFINE_SPINLOCK(managed_page_count_lock);
128 unsigned long totalram_pages __read_mostly;
129 unsigned long totalreserve_pages __read_mostly;
130 unsigned long totalcma_pages __read_mostly;
132 int percpu_pagelist_fraction;
133 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
143 static inline int get_pcppage_migratetype(struct page *page)
148 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
150 page->index = migratetype;
153 #ifdef CONFIG_PM_SLEEP
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with system_transition_mutex held
159 * (gfp_allowed_mask also should only be modified with system_transition_mutex
160 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
161 * with that modification).
164 static gfp_t saved_gfp_mask;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&system_transition_mutex));
169 if (saved_gfp_mask) {
170 gfp_allowed_mask = saved_gfp_mask;
175 void pm_restrict_gfp_mask(void)
177 WARN_ON(!mutex_is_locked(&system_transition_mutex));
178 WARN_ON(saved_gfp_mask);
179 saved_gfp_mask = gfp_allowed_mask;
180 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
183 bool pm_suspended_storage(void)
185 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
189 #endif /* CONFIG_PM_SLEEP */
191 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
192 unsigned int pageblock_order __read_mostly;
195 static void __free_pages_ok(struct page *page, unsigned int order);
198 * results with 256, 32 in the lowmem_reserve sysctl:
199 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
200 * 1G machine -> (16M dma, 784M normal, 224M high)
201 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
202 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
203 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
205 * TBD: should special case ZONE_DMA32 machines here - in those we normally
206 * don't need any ZONE_NORMAL reservation
208 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
222 EXPORT_SYMBOL(totalram_pages);
224 static char * const zone_names[MAX_NR_ZONES] = {
225 #ifdef CONFIG_ZONE_DMA
228 #ifdef CONFIG_ZONE_DMA32
232 #ifdef CONFIG_HIGHMEM
236 #ifdef CONFIG_ZONE_DEVICE
241 char * const migratetype_names[MIGRATE_TYPES] = {
249 #ifdef CONFIG_MEMORY_ISOLATION
254 compound_page_dtor * const compound_page_dtors[] = {
257 #ifdef CONFIG_HUGETLB_PAGE
260 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
265 int min_free_kbytes = 1024;
266 int user_min_free_kbytes = -1;
267 int watermark_scale_factor = 10;
269 static unsigned long nr_kernel_pages __meminitdata;
270 static unsigned long nr_all_pages __meminitdata;
271 static unsigned long dma_reserve __meminitdata;
273 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
274 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
276 static unsigned long required_kernelcore __initdata;
277 static unsigned long required_kernelcore_percent __initdata;
278 static unsigned long required_movablecore __initdata;
279 static unsigned long required_movablecore_percent __initdata;
280 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
281 static bool mirrored_kernelcore __meminitdata;
283 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
285 EXPORT_SYMBOL(movable_zone);
286 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
289 int nr_node_ids __read_mostly = MAX_NUMNODES;
290 int nr_online_nodes __read_mostly = 1;
291 EXPORT_SYMBOL(nr_node_ids);
292 EXPORT_SYMBOL(nr_online_nodes);
295 int page_group_by_mobility_disabled __read_mostly;
297 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 /* Returns true if the struct page for the pfn is uninitialised */
299 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
301 int nid = early_pfn_to_nid(pfn);
303 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
310 * Returns false when the remaining initialisation should be deferred until
311 * later in the boot cycle when it can be parallelised.
313 static inline bool update_defer_init(pg_data_t *pgdat,
314 unsigned long pfn, unsigned long zone_end,
315 unsigned long *nr_initialised)
317 /* Always populate low zones for address-constrained allocations */
318 if (zone_end < pgdat_end_pfn(pgdat))
321 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
322 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
323 pgdat->first_deferred_pfn = pfn;
330 static inline bool early_page_uninitialised(unsigned long pfn)
335 static inline bool update_defer_init(pg_data_t *pgdat,
336 unsigned long pfn, unsigned long zone_end,
337 unsigned long *nr_initialised)
343 /* Return a pointer to the bitmap storing bits affecting a block of pages */
344 static inline unsigned long *get_pageblock_bitmap(struct page *page,
347 #ifdef CONFIG_SPARSEMEM
348 return __pfn_to_section(pfn)->pageblock_flags;
350 return page_zone(page)->pageblock_flags;
351 #endif /* CONFIG_SPARSEMEM */
354 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
356 #ifdef CONFIG_SPARSEMEM
357 pfn &= (PAGES_PER_SECTION-1);
358 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
360 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
361 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362 #endif /* CONFIG_SPARSEMEM */
366 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
367 * @page: The page within the block of interest
368 * @pfn: The target page frame number
369 * @end_bitidx: The last bit of interest to retrieve
370 * @mask: mask of bits that the caller is interested in
372 * Return: pageblock_bits flags
374 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
376 unsigned long end_bitidx,
379 unsigned long *bitmap;
380 unsigned long bitidx, word_bitidx;
383 bitmap = get_pageblock_bitmap(page, pfn);
384 bitidx = pfn_to_bitidx(page, pfn);
385 word_bitidx = bitidx / BITS_PER_LONG;
386 bitidx &= (BITS_PER_LONG-1);
388 word = bitmap[word_bitidx];
389 bitidx += end_bitidx;
390 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
393 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
394 unsigned long end_bitidx,
397 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
400 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
402 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
406 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
407 * @page: The page within the block of interest
408 * @flags: The flags to set
409 * @pfn: The target page frame number
410 * @end_bitidx: The last bit of interest
411 * @mask: mask of bits that the caller is interested in
413 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
415 unsigned long end_bitidx,
418 unsigned long *bitmap;
419 unsigned long bitidx, word_bitidx;
420 unsigned long old_word, word;
422 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
424 bitmap = get_pageblock_bitmap(page, pfn);
425 bitidx = pfn_to_bitidx(page, pfn);
426 word_bitidx = bitidx / BITS_PER_LONG;
427 bitidx &= (BITS_PER_LONG-1);
429 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
431 bitidx += end_bitidx;
432 mask <<= (BITS_PER_LONG - bitidx - 1);
433 flags <<= (BITS_PER_LONG - bitidx - 1);
435 word = READ_ONCE(bitmap[word_bitidx]);
437 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
438 if (word == old_word)
444 void set_pageblock_migratetype(struct page *page, int migratetype)
446 if (unlikely(page_group_by_mobility_disabled &&
447 migratetype < MIGRATE_PCPTYPES))
448 migratetype = MIGRATE_UNMOVABLE;
450 set_pageblock_flags_group(page, (unsigned long)migratetype,
451 PB_migrate, PB_migrate_end);
454 #ifdef CONFIG_DEBUG_VM
455 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
459 unsigned long pfn = page_to_pfn(page);
460 unsigned long sp, start_pfn;
463 seq = zone_span_seqbegin(zone);
464 start_pfn = zone->zone_start_pfn;
465 sp = zone->spanned_pages;
466 if (!zone_spans_pfn(zone, pfn))
468 } while (zone_span_seqretry(zone, seq));
471 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
472 pfn, zone_to_nid(zone), zone->name,
473 start_pfn, start_pfn + sp);
478 static int page_is_consistent(struct zone *zone, struct page *page)
480 if (!pfn_valid_within(page_to_pfn(page)))
482 if (zone != page_zone(page))
488 * Temporary debugging check for pages not lying within a given zone.
490 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
492 if (page_outside_zone_boundaries(zone, page))
494 if (!page_is_consistent(zone, page))
500 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
506 static void bad_page(struct page *page, const char *reason,
507 unsigned long bad_flags)
509 static unsigned long resume;
510 static unsigned long nr_shown;
511 static unsigned long nr_unshown;
514 * Allow a burst of 60 reports, then keep quiet for that minute;
515 * or allow a steady drip of one report per second.
517 if (nr_shown == 60) {
518 if (time_before(jiffies, resume)) {
524 "BUG: Bad page state: %lu messages suppressed\n",
531 resume = jiffies + 60 * HZ;
533 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
534 current->comm, page_to_pfn(page));
535 __dump_page(page, reason);
536 bad_flags &= page->flags;
538 pr_alert("bad because of flags: %#lx(%pGp)\n",
539 bad_flags, &bad_flags);
540 dump_page_owner(page);
545 /* Leave bad fields for debug, except PageBuddy could make trouble */
546 page_mapcount_reset(page); /* remove PageBuddy */
547 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
551 * Higher-order pages are called "compound pages". They are structured thusly:
553 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
555 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
556 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
558 * The first tail page's ->compound_dtor holds the offset in array of compound
559 * page destructors. See compound_page_dtors.
561 * The first tail page's ->compound_order holds the order of allocation.
562 * This usage means that zero-order pages may not be compound.
565 void free_compound_page(struct page *page)
567 __free_pages_ok(page, compound_order(page));
570 void prep_compound_page(struct page *page, unsigned int order)
573 int nr_pages = 1 << order;
575 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
576 set_compound_order(page, order);
578 for (i = 1; i < nr_pages; i++) {
579 struct page *p = page + i;
580 set_page_count(p, 0);
581 p->mapping = TAIL_MAPPING;
582 set_compound_head(p, page);
584 atomic_set(compound_mapcount_ptr(page), -1);
587 #ifdef CONFIG_DEBUG_PAGEALLOC
588 unsigned int _debug_guardpage_minorder;
589 bool _debug_pagealloc_enabled __read_mostly
590 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
591 EXPORT_SYMBOL(_debug_pagealloc_enabled);
592 bool _debug_guardpage_enabled __read_mostly;
594 static int __init early_debug_pagealloc(char *buf)
598 return kstrtobool(buf, &_debug_pagealloc_enabled);
600 early_param("debug_pagealloc", early_debug_pagealloc);
602 static bool need_debug_guardpage(void)
604 /* If we don't use debug_pagealloc, we don't need guard page */
605 if (!debug_pagealloc_enabled())
608 if (!debug_guardpage_minorder())
614 static void init_debug_guardpage(void)
616 if (!debug_pagealloc_enabled())
619 if (!debug_guardpage_minorder())
622 _debug_guardpage_enabled = true;
625 struct page_ext_operations debug_guardpage_ops = {
626 .need = need_debug_guardpage,
627 .init = init_debug_guardpage,
630 static int __init debug_guardpage_minorder_setup(char *buf)
634 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
635 pr_err("Bad debug_guardpage_minorder value\n");
638 _debug_guardpage_minorder = res;
639 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
642 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
644 static inline bool set_page_guard(struct zone *zone, struct page *page,
645 unsigned int order, int migratetype)
647 struct page_ext *page_ext;
649 if (!debug_guardpage_enabled())
652 if (order >= debug_guardpage_minorder())
655 page_ext = lookup_page_ext(page);
656 if (unlikely(!page_ext))
659 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
661 INIT_LIST_HEAD(&page->lru);
662 set_page_private(page, order);
663 /* Guard pages are not available for any usage */
664 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
669 static inline void clear_page_guard(struct zone *zone, struct page *page,
670 unsigned int order, int migratetype)
672 struct page_ext *page_ext;
674 if (!debug_guardpage_enabled())
677 page_ext = lookup_page_ext(page);
678 if (unlikely(!page_ext))
681 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
683 set_page_private(page, 0);
684 if (!is_migrate_isolate(migratetype))
685 __mod_zone_freepage_state(zone, (1 << order), migratetype);
688 struct page_ext_operations debug_guardpage_ops;
689 static inline bool set_page_guard(struct zone *zone, struct page *page,
690 unsigned int order, int migratetype) { return false; }
691 static inline void clear_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype) {}
695 static inline void set_page_order(struct page *page, unsigned int order)
697 set_page_private(page, order);
698 __SetPageBuddy(page);
701 static inline void rmv_page_order(struct page *page)
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
708 * This function checks whether a page is free && is the buddy
709 * we can coalesce a page and its buddy if
710 * (a) the buddy is not in a hole (check before calling!) &&
711 * (b) the buddy is in the buddy system &&
712 * (c) a page and its buddy have the same order &&
713 * (d) a page and its buddy are in the same zone.
715 * For recording whether a page is in the buddy system, we set PageBuddy.
716 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
718 * For recording page's order, we use page_private(page).
720 static inline int page_is_buddy(struct page *page, struct page *buddy,
723 if (page_is_guard(buddy) && page_order(buddy) == order) {
724 if (page_zone_id(page) != page_zone_id(buddy))
727 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
732 if (PageBuddy(buddy) && page_order(buddy) == order) {
734 * zone check is done late to avoid uselessly
735 * calculating zone/node ids for pages that could
738 if (page_zone_id(page) != page_zone_id(buddy))
741 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
749 * Freeing function for a buddy system allocator.
751 * The concept of a buddy system is to maintain direct-mapped table
752 * (containing bit values) for memory blocks of various "orders".
753 * The bottom level table contains the map for the smallest allocatable
754 * units of memory (here, pages), and each level above it describes
755 * pairs of units from the levels below, hence, "buddies".
756 * At a high level, all that happens here is marking the table entry
757 * at the bottom level available, and propagating the changes upward
758 * as necessary, plus some accounting needed to play nicely with other
759 * parts of the VM system.
760 * At each level, we keep a list of pages, which are heads of continuous
761 * free pages of length of (1 << order) and marked with PageBuddy.
762 * Page's order is recorded in page_private(page) field.
763 * So when we are allocating or freeing one, we can derive the state of the
764 * other. That is, if we allocate a small block, and both were
765 * free, the remainder of the region must be split into blocks.
766 * If a block is freed, and its buddy is also free, then this
767 * triggers coalescing into a block of larger size.
772 static inline void __free_one_page(struct page *page,
774 struct zone *zone, unsigned int order,
777 unsigned long combined_pfn;
778 unsigned long uninitialized_var(buddy_pfn);
780 unsigned int max_order;
782 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
784 VM_BUG_ON(!zone_is_initialized(zone));
785 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
787 VM_BUG_ON(migratetype == -1);
788 if (likely(!is_migrate_isolate(migratetype)))
789 __mod_zone_freepage_state(zone, 1 << order, migratetype);
791 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
792 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795 while (order < max_order - 1) {
796 buddy_pfn = __find_buddy_pfn(pfn, order);
797 buddy = page + (buddy_pfn - pfn);
799 if (!pfn_valid_within(buddy_pfn))
801 if (!page_is_buddy(page, buddy, order))
804 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
805 * merge with it and move up one order.
807 if (page_is_guard(buddy)) {
808 clear_page_guard(zone, buddy, order, migratetype);
810 list_del(&buddy->lru);
811 zone->free_area[order].nr_free--;
812 rmv_page_order(buddy);
814 combined_pfn = buddy_pfn & pfn;
815 page = page + (combined_pfn - pfn);
819 if (max_order < MAX_ORDER) {
820 /* If we are here, it means order is >= pageblock_order.
821 * We want to prevent merge between freepages on isolate
822 * pageblock and normal pageblock. Without this, pageblock
823 * isolation could cause incorrect freepage or CMA accounting.
825 * We don't want to hit this code for the more frequent
828 if (unlikely(has_isolate_pageblock(zone))) {
831 buddy_pfn = __find_buddy_pfn(pfn, order);
832 buddy = page + (buddy_pfn - pfn);
833 buddy_mt = get_pageblock_migratetype(buddy);
835 if (migratetype != buddy_mt
836 && (is_migrate_isolate(migratetype) ||
837 is_migrate_isolate(buddy_mt)))
841 goto continue_merging;
845 set_page_order(page, order);
848 * If this is not the largest possible page, check if the buddy
849 * of the next-highest order is free. If it is, it's possible
850 * that pages are being freed that will coalesce soon. In case,
851 * that is happening, add the free page to the tail of the list
852 * so it's less likely to be used soon and more likely to be merged
853 * as a higher order page
855 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
856 struct page *higher_page, *higher_buddy;
857 combined_pfn = buddy_pfn & pfn;
858 higher_page = page + (combined_pfn - pfn);
859 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
860 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
861 if (pfn_valid_within(buddy_pfn) &&
862 page_is_buddy(higher_page, higher_buddy, order + 1)) {
863 list_add_tail(&page->lru,
864 &zone->free_area[order].free_list[migratetype]);
869 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
871 zone->free_area[order].nr_free++;
875 * A bad page could be due to a number of fields. Instead of multiple branches,
876 * try and check multiple fields with one check. The caller must do a detailed
877 * check if necessary.
879 static inline bool page_expected_state(struct page *page,
880 unsigned long check_flags)
882 if (unlikely(atomic_read(&page->_mapcount) != -1))
885 if (unlikely((unsigned long)page->mapping |
886 page_ref_count(page) |
888 (unsigned long)page->mem_cgroup |
890 (page->flags & check_flags)))
896 static void free_pages_check_bad(struct page *page)
898 const char *bad_reason;
899 unsigned long bad_flags;
904 if (unlikely(atomic_read(&page->_mapcount) != -1))
905 bad_reason = "nonzero mapcount";
906 if (unlikely(page->mapping != NULL))
907 bad_reason = "non-NULL mapping";
908 if (unlikely(page_ref_count(page) != 0))
909 bad_reason = "nonzero _refcount";
910 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
911 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
912 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
915 if (unlikely(page->mem_cgroup))
916 bad_reason = "page still charged to cgroup";
918 bad_page(page, bad_reason, bad_flags);
921 static inline int free_pages_check(struct page *page)
923 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
926 /* Something has gone sideways, find it */
927 free_pages_check_bad(page);
931 static int free_tail_pages_check(struct page *head_page, struct page *page)
936 * We rely page->lru.next never has bit 0 set, unless the page
937 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
939 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
941 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
945 switch (page - head_page) {
947 /* the first tail page: ->mapping may be compound_mapcount() */
948 if (unlikely(compound_mapcount(page))) {
949 bad_page(page, "nonzero compound_mapcount", 0);
955 * the second tail page: ->mapping is
956 * deferred_list.next -- ignore value.
960 if (page->mapping != TAIL_MAPPING) {
961 bad_page(page, "corrupted mapping in tail page", 0);
966 if (unlikely(!PageTail(page))) {
967 bad_page(page, "PageTail not set", 0);
970 if (unlikely(compound_head(page) != head_page)) {
971 bad_page(page, "compound_head not consistent", 0);
976 page->mapping = NULL;
977 clear_compound_head(page);
981 static __always_inline bool free_pages_prepare(struct page *page,
982 unsigned int order, bool check_free)
986 VM_BUG_ON_PAGE(PageTail(page), page);
988 trace_mm_page_free(page, order);
991 * Check tail pages before head page information is cleared to
992 * avoid checking PageCompound for order-0 pages.
994 if (unlikely(order)) {
995 bool compound = PageCompound(page);
998 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1001 ClearPageDoubleMap(page);
1002 for (i = 1; i < (1 << order); i++) {
1004 bad += free_tail_pages_check(page, page + i);
1005 if (unlikely(free_pages_check(page + i))) {
1009 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 if (PageMappingFlags(page))
1013 page->mapping = NULL;
1014 if (memcg_kmem_enabled() && PageKmemcg(page))
1015 memcg_kmem_uncharge(page, order);
1017 bad += free_pages_check(page);
1021 page_cpupid_reset_last(page);
1022 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 reset_page_owner(page, order);
1025 if (!PageHighMem(page)) {
1026 debug_check_no_locks_freed(page_address(page),
1027 PAGE_SIZE << order);
1028 debug_check_no_obj_freed(page_address(page),
1029 PAGE_SIZE << order);
1031 arch_free_page(page, order);
1032 kernel_poison_pages(page, 1 << order, 0);
1033 kernel_map_pages(page, 1 << order, 0);
1034 kasan_free_pages(page, order);
1039 #ifdef CONFIG_DEBUG_VM
1040 static inline bool free_pcp_prepare(struct page *page)
1042 return free_pages_prepare(page, 0, true);
1045 static inline bool bulkfree_pcp_prepare(struct page *page)
1050 static bool free_pcp_prepare(struct page *page)
1052 return free_pages_prepare(page, 0, false);
1055 static bool bulkfree_pcp_prepare(struct page *page)
1057 return free_pages_check(page);
1059 #endif /* CONFIG_DEBUG_VM */
1061 static inline void prefetch_buddy(struct page *page)
1063 unsigned long pfn = page_to_pfn(page);
1064 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1065 struct page *buddy = page + (buddy_pfn - pfn);
1071 * Frees a number of pages from the PCP lists
1072 * Assumes all pages on list are in same zone, and of same order.
1073 * count is the number of pages to free.
1075 * If the zone was previously in an "all pages pinned" state then look to
1076 * see if this freeing clears that state.
1078 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1079 * pinned" detection logic.
1081 static void free_pcppages_bulk(struct zone *zone, int count,
1082 struct per_cpu_pages *pcp)
1084 int migratetype = 0;
1086 int prefetch_nr = 0;
1087 bool isolated_pageblocks;
1088 struct page *page, *tmp;
1092 struct list_head *list;
1095 * Remove pages from lists in a round-robin fashion. A
1096 * batch_free count is maintained that is incremented when an
1097 * empty list is encountered. This is so more pages are freed
1098 * off fuller lists instead of spinning excessively around empty
1103 if (++migratetype == MIGRATE_PCPTYPES)
1105 list = &pcp->lists[migratetype];
1106 } while (list_empty(list));
1108 /* This is the only non-empty list. Free them all. */
1109 if (batch_free == MIGRATE_PCPTYPES)
1113 page = list_last_entry(list, struct page, lru);
1114 /* must delete to avoid corrupting pcp list */
1115 list_del(&page->lru);
1118 if (bulkfree_pcp_prepare(page))
1121 list_add_tail(&page->lru, &head);
1124 * We are going to put the page back to the global
1125 * pool, prefetch its buddy to speed up later access
1126 * under zone->lock. It is believed the overhead of
1127 * an additional test and calculating buddy_pfn here
1128 * can be offset by reduced memory latency later. To
1129 * avoid excessive prefetching due to large count, only
1130 * prefetch buddy for the first pcp->batch nr of pages.
1132 if (prefetch_nr++ < pcp->batch)
1133 prefetch_buddy(page);
1134 } while (--count && --batch_free && !list_empty(list));
1137 spin_lock(&zone->lock);
1138 isolated_pageblocks = has_isolate_pageblock(zone);
1141 * Use safe version since after __free_one_page(),
1142 * page->lru.next will not point to original list.
1144 list_for_each_entry_safe(page, tmp, &head, lru) {
1145 int mt = get_pcppage_migratetype(page);
1146 /* MIGRATE_ISOLATE page should not go to pcplists */
1147 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1148 /* Pageblock could have been isolated meanwhile */
1149 if (unlikely(isolated_pageblocks))
1150 mt = get_pageblock_migratetype(page);
1152 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1153 trace_mm_page_pcpu_drain(page, 0, mt);
1155 spin_unlock(&zone->lock);
1158 static void free_one_page(struct zone *zone,
1159 struct page *page, unsigned long pfn,
1163 spin_lock(&zone->lock);
1164 if (unlikely(has_isolate_pageblock(zone) ||
1165 is_migrate_isolate(migratetype))) {
1166 migratetype = get_pfnblock_migratetype(page, pfn);
1168 __free_one_page(page, pfn, zone, order, migratetype);
1169 spin_unlock(&zone->lock);
1172 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1173 unsigned long zone, int nid)
1175 mm_zero_struct_page(page);
1176 set_page_links(page, zone, nid, pfn);
1177 init_page_count(page);
1178 page_mapcount_reset(page);
1179 page_cpupid_reset_last(page);
1181 INIT_LIST_HEAD(&page->lru);
1182 #ifdef WANT_PAGE_VIRTUAL
1183 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1184 if (!is_highmem_idx(zone))
1185 set_page_address(page, __va(pfn << PAGE_SHIFT));
1189 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1190 static void __meminit init_reserved_page(unsigned long pfn)
1195 if (!early_page_uninitialised(pfn))
1198 nid = early_pfn_to_nid(pfn);
1199 pgdat = NODE_DATA(nid);
1201 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1202 struct zone *zone = &pgdat->node_zones[zid];
1204 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1207 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1210 static inline void init_reserved_page(unsigned long pfn)
1213 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216 * Initialised pages do not have PageReserved set. This function is
1217 * called for each range allocated by the bootmem allocator and
1218 * marks the pages PageReserved. The remaining valid pages are later
1219 * sent to the buddy page allocator.
1221 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1223 unsigned long start_pfn = PFN_DOWN(start);
1224 unsigned long end_pfn = PFN_UP(end);
1226 for (; start_pfn < end_pfn; start_pfn++) {
1227 if (pfn_valid(start_pfn)) {
1228 struct page *page = pfn_to_page(start_pfn);
1230 init_reserved_page(start_pfn);
1232 /* Avoid false-positive PageTail() */
1233 INIT_LIST_HEAD(&page->lru);
1235 SetPageReserved(page);
1240 static void __free_pages_ok(struct page *page, unsigned int order)
1242 unsigned long flags;
1244 unsigned long pfn = page_to_pfn(page);
1246 if (!free_pages_prepare(page, order, true))
1249 migratetype = get_pfnblock_migratetype(page, pfn);
1250 local_irq_save(flags);
1251 __count_vm_events(PGFREE, 1 << order);
1252 free_one_page(page_zone(page), page, pfn, order, migratetype);
1253 local_irq_restore(flags);
1256 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1258 unsigned int nr_pages = 1 << order;
1259 struct page *p = page;
1263 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1265 __ClearPageReserved(p);
1266 set_page_count(p, 0);
1268 __ClearPageReserved(p);
1269 set_page_count(p, 0);
1271 page_zone(page)->managed_pages += nr_pages;
1272 set_page_refcounted(page);
1273 __free_pages(page, order);
1276 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1277 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1279 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1281 int __meminit early_pfn_to_nid(unsigned long pfn)
1283 static DEFINE_SPINLOCK(early_pfn_lock);
1286 spin_lock(&early_pfn_lock);
1287 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1289 nid = first_online_node;
1290 spin_unlock(&early_pfn_lock);
1296 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1297 static inline bool __meminit __maybe_unused
1298 meminit_pfn_in_nid(unsigned long pfn, int node,
1299 struct mminit_pfnnid_cache *state)
1303 nid = __early_pfn_to_nid(pfn, state);
1304 if (nid >= 0 && nid != node)
1309 /* Only safe to use early in boot when initialisation is single-threaded */
1310 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1312 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1317 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1321 static inline bool __meminit __maybe_unused
1322 meminit_pfn_in_nid(unsigned long pfn, int node,
1323 struct mminit_pfnnid_cache *state)
1330 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1333 if (early_page_uninitialised(pfn))
1335 return __free_pages_boot_core(page, order);
1339 * Check that the whole (or subset of) a pageblock given by the interval of
1340 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1341 * with the migration of free compaction scanner. The scanners then need to
1342 * use only pfn_valid_within() check for arches that allow holes within
1345 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1347 * It's possible on some configurations to have a setup like node0 node1 node0
1348 * i.e. it's possible that all pages within a zones range of pages do not
1349 * belong to a single zone. We assume that a border between node0 and node1
1350 * can occur within a single pageblock, but not a node0 node1 node0
1351 * interleaving within a single pageblock. It is therefore sufficient to check
1352 * the first and last page of a pageblock and avoid checking each individual
1353 * page in a pageblock.
1355 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1356 unsigned long end_pfn, struct zone *zone)
1358 struct page *start_page;
1359 struct page *end_page;
1361 /* end_pfn is one past the range we are checking */
1364 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1367 start_page = pfn_to_online_page(start_pfn);
1371 if (page_zone(start_page) != zone)
1374 end_page = pfn_to_page(end_pfn);
1376 /* This gives a shorter code than deriving page_zone(end_page) */
1377 if (page_zone_id(start_page) != page_zone_id(end_page))
1383 void set_zone_contiguous(struct zone *zone)
1385 unsigned long block_start_pfn = zone->zone_start_pfn;
1386 unsigned long block_end_pfn;
1388 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1389 for (; block_start_pfn < zone_end_pfn(zone);
1390 block_start_pfn = block_end_pfn,
1391 block_end_pfn += pageblock_nr_pages) {
1393 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1395 if (!__pageblock_pfn_to_page(block_start_pfn,
1396 block_end_pfn, zone))
1400 /* We confirm that there is no hole */
1401 zone->contiguous = true;
1404 void clear_zone_contiguous(struct zone *zone)
1406 zone->contiguous = false;
1409 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1410 static void __init deferred_free_range(unsigned long pfn,
1411 unsigned long nr_pages)
1419 page = pfn_to_page(pfn);
1421 /* Free a large naturally-aligned chunk if possible */
1422 if (nr_pages == pageblock_nr_pages &&
1423 (pfn & (pageblock_nr_pages - 1)) == 0) {
1424 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1425 __free_pages_boot_core(page, pageblock_order);
1429 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1430 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1431 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1432 __free_pages_boot_core(page, 0);
1436 /* Completion tracking for deferred_init_memmap() threads */
1437 static atomic_t pgdat_init_n_undone __initdata;
1438 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1440 static inline void __init pgdat_init_report_one_done(void)
1442 if (atomic_dec_and_test(&pgdat_init_n_undone))
1443 complete(&pgdat_init_all_done_comp);
1447 * Returns true if page needs to be initialized or freed to buddy allocator.
1449 * First we check if pfn is valid on architectures where it is possible to have
1450 * holes within pageblock_nr_pages. On systems where it is not possible, this
1451 * function is optimized out.
1453 * Then, we check if a current large page is valid by only checking the validity
1456 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1457 * within a node: a pfn is between start and end of a node, but does not belong
1458 * to this memory node.
1460 static inline bool __init
1461 deferred_pfn_valid(int nid, unsigned long pfn,
1462 struct mminit_pfnnid_cache *nid_init_state)
1464 if (!pfn_valid_within(pfn))
1466 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1468 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1474 * Free pages to buddy allocator. Try to free aligned pages in
1475 * pageblock_nr_pages sizes.
1477 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1478 unsigned long end_pfn)
1480 struct mminit_pfnnid_cache nid_init_state = { };
1481 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1482 unsigned long nr_free = 0;
1484 for (; pfn < end_pfn; pfn++) {
1485 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1486 deferred_free_range(pfn - nr_free, nr_free);
1488 } else if (!(pfn & nr_pgmask)) {
1489 deferred_free_range(pfn - nr_free, nr_free);
1491 touch_nmi_watchdog();
1496 /* Free the last block of pages to allocator */
1497 deferred_free_range(pfn - nr_free, nr_free);
1501 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1502 * by performing it only once every pageblock_nr_pages.
1503 * Return number of pages initialized.
1505 static unsigned long __init deferred_init_pages(int nid, int zid,
1507 unsigned long end_pfn)
1509 struct mminit_pfnnid_cache nid_init_state = { };
1510 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1511 unsigned long nr_pages = 0;
1512 struct page *page = NULL;
1514 for (; pfn < end_pfn; pfn++) {
1515 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1518 } else if (!page || !(pfn & nr_pgmask)) {
1519 page = pfn_to_page(pfn);
1520 touch_nmi_watchdog();
1524 __init_single_page(page, pfn, zid, nid);
1530 /* Initialise remaining memory on a node */
1531 static int __init deferred_init_memmap(void *data)
1533 pg_data_t *pgdat = data;
1534 int nid = pgdat->node_id;
1535 unsigned long start = jiffies;
1536 unsigned long nr_pages = 0;
1537 unsigned long spfn, epfn, first_init_pfn, flags;
1538 phys_addr_t spa, epa;
1541 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1544 /* Bind memory initialisation thread to a local node if possible */
1545 if (!cpumask_empty(cpumask))
1546 set_cpus_allowed_ptr(current, cpumask);
1548 pgdat_resize_lock(pgdat, &flags);
1549 first_init_pfn = pgdat->first_deferred_pfn;
1550 if (first_init_pfn == ULONG_MAX) {
1551 pgdat_resize_unlock(pgdat, &flags);
1552 pgdat_init_report_one_done();
1556 /* Sanity check boundaries */
1557 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1558 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1559 pgdat->first_deferred_pfn = ULONG_MAX;
1561 /* Only the highest zone is deferred so find it */
1562 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1563 zone = pgdat->node_zones + zid;
1564 if (first_init_pfn < zone_end_pfn(zone))
1567 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1570 * Initialize and free pages. We do it in two loops: first we initialize
1571 * struct page, than free to buddy allocator, because while we are
1572 * freeing pages we can access pages that are ahead (computing buddy
1573 * page in __free_one_page()).
1575 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1576 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1577 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1578 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1580 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1581 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1582 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1583 deferred_free_pages(nid, zid, spfn, epfn);
1585 pgdat_resize_unlock(pgdat, &flags);
1587 /* Sanity check that the next zone really is unpopulated */
1588 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1590 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1591 jiffies_to_msecs(jiffies - start));
1593 pgdat_init_report_one_done();
1598 * During boot we initialize deferred pages on-demand, as needed, but once
1599 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1600 * and we can permanently disable that path.
1602 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1605 * If this zone has deferred pages, try to grow it by initializing enough
1606 * deferred pages to satisfy the allocation specified by order, rounded up to
1607 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1608 * of SECTION_SIZE bytes by initializing struct pages in increments of
1609 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1611 * Return true when zone was grown, otherwise return false. We return true even
1612 * when we grow less than requested, to let the caller decide if there are
1613 * enough pages to satisfy the allocation.
1615 * Note: We use noinline because this function is needed only during boot, and
1616 * it is called from a __ref function _deferred_grow_zone. This way we are
1617 * making sure that it is not inlined into permanent text section.
1619 static noinline bool __init
1620 deferred_grow_zone(struct zone *zone, unsigned int order)
1622 int zid = zone_idx(zone);
1623 int nid = zone_to_nid(zone);
1624 pg_data_t *pgdat = NODE_DATA(nid);
1625 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1626 unsigned long nr_pages = 0;
1627 unsigned long first_init_pfn, spfn, epfn, t, flags;
1628 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1629 phys_addr_t spa, epa;
1632 /* Only the last zone may have deferred pages */
1633 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1636 pgdat_resize_lock(pgdat, &flags);
1639 * If deferred pages have been initialized while we were waiting for
1640 * the lock, return true, as the zone was grown. The caller will retry
1641 * this zone. We won't return to this function since the caller also
1642 * has this static branch.
1644 if (!static_branch_unlikely(&deferred_pages)) {
1645 pgdat_resize_unlock(pgdat, &flags);
1650 * If someone grew this zone while we were waiting for spinlock, return
1651 * true, as there might be enough pages already.
1653 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1654 pgdat_resize_unlock(pgdat, &flags);
1658 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1660 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1661 pgdat_resize_unlock(pgdat, &flags);
1665 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1666 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1667 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1669 while (spfn < epfn && nr_pages < nr_pages_needed) {
1670 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1671 first_deferred_pfn = min(t, epfn);
1672 nr_pages += deferred_init_pages(nid, zid, spfn,
1673 first_deferred_pfn);
1674 spfn = first_deferred_pfn;
1677 if (nr_pages >= nr_pages_needed)
1681 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1682 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1683 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1684 deferred_free_pages(nid, zid, spfn, epfn);
1686 if (first_deferred_pfn == epfn)
1689 pgdat->first_deferred_pfn = first_deferred_pfn;
1690 pgdat_resize_unlock(pgdat, &flags);
1692 return nr_pages > 0;
1696 * deferred_grow_zone() is __init, but it is called from
1697 * get_page_from_freelist() during early boot until deferred_pages permanently
1698 * disables this call. This is why we have refdata wrapper to avoid warning,
1699 * and to ensure that the function body gets unloaded.
1702 _deferred_grow_zone(struct zone *zone, unsigned int order)
1704 return deferred_grow_zone(zone, order);
1707 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1709 void __init page_alloc_init_late(void)
1713 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1716 /* There will be num_node_state(N_MEMORY) threads */
1717 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1718 for_each_node_state(nid, N_MEMORY) {
1719 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1722 /* Block until all are initialised */
1723 wait_for_completion(&pgdat_init_all_done_comp);
1726 * We initialized the rest of the deferred pages. Permanently disable
1727 * on-demand struct page initialization.
1729 static_branch_disable(&deferred_pages);
1731 /* Reinit limits that are based on free pages after the kernel is up */
1732 files_maxfiles_init();
1734 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1735 /* Discard memblock private memory */
1739 for_each_populated_zone(zone)
1740 set_zone_contiguous(zone);
1744 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1745 void __init init_cma_reserved_pageblock(struct page *page)
1747 unsigned i = pageblock_nr_pages;
1748 struct page *p = page;
1751 __ClearPageReserved(p);
1752 set_page_count(p, 0);
1755 set_pageblock_migratetype(page, MIGRATE_CMA);
1757 if (pageblock_order >= MAX_ORDER) {
1758 i = pageblock_nr_pages;
1761 set_page_refcounted(p);
1762 __free_pages(p, MAX_ORDER - 1);
1763 p += MAX_ORDER_NR_PAGES;
1764 } while (i -= MAX_ORDER_NR_PAGES);
1766 set_page_refcounted(page);
1767 __free_pages(page, pageblock_order);
1770 adjust_managed_page_count(page, pageblock_nr_pages);
1775 * The order of subdivision here is critical for the IO subsystem.
1776 * Please do not alter this order without good reasons and regression
1777 * testing. Specifically, as large blocks of memory are subdivided,
1778 * the order in which smaller blocks are delivered depends on the order
1779 * they're subdivided in this function. This is the primary factor
1780 * influencing the order in which pages are delivered to the IO
1781 * subsystem according to empirical testing, and this is also justified
1782 * by considering the behavior of a buddy system containing a single
1783 * large block of memory acted on by a series of small allocations.
1784 * This behavior is a critical factor in sglist merging's success.
1788 static inline void expand(struct zone *zone, struct page *page,
1789 int low, int high, struct free_area *area,
1792 unsigned long size = 1 << high;
1794 while (high > low) {
1798 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1801 * Mark as guard pages (or page), that will allow to
1802 * merge back to allocator when buddy will be freed.
1803 * Corresponding page table entries will not be touched,
1804 * pages will stay not present in virtual address space
1806 if (set_page_guard(zone, &page[size], high, migratetype))
1809 list_add(&page[size].lru, &area->free_list[migratetype]);
1811 set_page_order(&page[size], high);
1815 static void check_new_page_bad(struct page *page)
1817 const char *bad_reason = NULL;
1818 unsigned long bad_flags = 0;
1820 if (unlikely(atomic_read(&page->_mapcount) != -1))
1821 bad_reason = "nonzero mapcount";
1822 if (unlikely(page->mapping != NULL))
1823 bad_reason = "non-NULL mapping";
1824 if (unlikely(page_ref_count(page) != 0))
1825 bad_reason = "nonzero _count";
1826 if (unlikely(page->flags & __PG_HWPOISON)) {
1827 bad_reason = "HWPoisoned (hardware-corrupted)";
1828 bad_flags = __PG_HWPOISON;
1829 /* Don't complain about hwpoisoned pages */
1830 page_mapcount_reset(page); /* remove PageBuddy */
1833 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1834 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1835 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1838 if (unlikely(page->mem_cgroup))
1839 bad_reason = "page still charged to cgroup";
1841 bad_page(page, bad_reason, bad_flags);
1845 * This page is about to be returned from the page allocator
1847 static inline int check_new_page(struct page *page)
1849 if (likely(page_expected_state(page,
1850 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1853 check_new_page_bad(page);
1857 static inline bool free_pages_prezeroed(void)
1859 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1860 page_poisoning_enabled();
1863 #ifdef CONFIG_DEBUG_VM
1864 static bool check_pcp_refill(struct page *page)
1869 static bool check_new_pcp(struct page *page)
1871 return check_new_page(page);
1874 static bool check_pcp_refill(struct page *page)
1876 return check_new_page(page);
1878 static bool check_new_pcp(struct page *page)
1882 #endif /* CONFIG_DEBUG_VM */
1884 static bool check_new_pages(struct page *page, unsigned int order)
1887 for (i = 0; i < (1 << order); i++) {
1888 struct page *p = page + i;
1890 if (unlikely(check_new_page(p)))
1897 inline void post_alloc_hook(struct page *page, unsigned int order,
1900 set_page_private(page, 0);
1901 set_page_refcounted(page);
1903 arch_alloc_page(page, order);
1904 kernel_map_pages(page, 1 << order, 1);
1905 kernel_poison_pages(page, 1 << order, 1);
1906 kasan_alloc_pages(page, order);
1907 set_page_owner(page, order, gfp_flags);
1910 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1911 unsigned int alloc_flags)
1915 post_alloc_hook(page, order, gfp_flags);
1917 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1918 for (i = 0; i < (1 << order); i++)
1919 clear_highpage(page + i);
1921 if (order && (gfp_flags & __GFP_COMP))
1922 prep_compound_page(page, order);
1925 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1926 * allocate the page. The expectation is that the caller is taking
1927 * steps that will free more memory. The caller should avoid the page
1928 * being used for !PFMEMALLOC purposes.
1930 if (alloc_flags & ALLOC_NO_WATERMARKS)
1931 set_page_pfmemalloc(page);
1933 clear_page_pfmemalloc(page);
1937 * Go through the free lists for the given migratetype and remove
1938 * the smallest available page from the freelists
1940 static __always_inline
1941 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1944 unsigned int current_order;
1945 struct free_area *area;
1948 /* Find a page of the appropriate size in the preferred list */
1949 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1950 area = &(zone->free_area[current_order]);
1951 page = list_first_entry_or_null(&area->free_list[migratetype],
1955 list_del(&page->lru);
1956 rmv_page_order(page);
1958 expand(zone, page, order, current_order, area, migratetype);
1959 set_pcppage_migratetype(page, migratetype);
1968 * This array describes the order lists are fallen back to when
1969 * the free lists for the desirable migrate type are depleted
1971 static int fallbacks[MIGRATE_TYPES][4] = {
1972 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1973 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1974 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1976 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1978 #ifdef CONFIG_MEMORY_ISOLATION
1979 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1984 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1987 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1990 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1991 unsigned int order) { return NULL; }
1995 * Move the free pages in a range to the free lists of the requested type.
1996 * Note that start_page and end_pages are not aligned on a pageblock
1997 * boundary. If alignment is required, use move_freepages_block()
1999 static int move_freepages(struct zone *zone,
2000 struct page *start_page, struct page *end_page,
2001 int migratetype, int *num_movable)
2005 int pages_moved = 0;
2007 #ifndef CONFIG_HOLES_IN_ZONE
2009 * page_zone is not safe to call in this context when
2010 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2011 * anyway as we check zone boundaries in move_freepages_block().
2012 * Remove at a later date when no bug reports exist related to
2013 * grouping pages by mobility
2015 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2016 pfn_valid(page_to_pfn(end_page)) &&
2017 page_zone(start_page) != page_zone(end_page));
2023 for (page = start_page; page <= end_page;) {
2024 if (!pfn_valid_within(page_to_pfn(page))) {
2029 /* Make sure we are not inadvertently changing nodes */
2030 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2032 if (!PageBuddy(page)) {
2034 * We assume that pages that could be isolated for
2035 * migration are movable. But we don't actually try
2036 * isolating, as that would be expensive.
2039 (PageLRU(page) || __PageMovable(page)))
2046 order = page_order(page);
2047 list_move(&page->lru,
2048 &zone->free_area[order].free_list[migratetype]);
2050 pages_moved += 1 << order;
2056 int move_freepages_block(struct zone *zone, struct page *page,
2057 int migratetype, int *num_movable)
2059 unsigned long start_pfn, end_pfn;
2060 struct page *start_page, *end_page;
2062 start_pfn = page_to_pfn(page);
2063 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2064 start_page = pfn_to_page(start_pfn);
2065 end_page = start_page + pageblock_nr_pages - 1;
2066 end_pfn = start_pfn + pageblock_nr_pages - 1;
2068 /* Do not cross zone boundaries */
2069 if (!zone_spans_pfn(zone, start_pfn))
2071 if (!zone_spans_pfn(zone, end_pfn))
2074 return move_freepages(zone, start_page, end_page, migratetype,
2078 static void change_pageblock_range(struct page *pageblock_page,
2079 int start_order, int migratetype)
2081 int nr_pageblocks = 1 << (start_order - pageblock_order);
2083 while (nr_pageblocks--) {
2084 set_pageblock_migratetype(pageblock_page, migratetype);
2085 pageblock_page += pageblock_nr_pages;
2090 * When we are falling back to another migratetype during allocation, try to
2091 * steal extra free pages from the same pageblocks to satisfy further
2092 * allocations, instead of polluting multiple pageblocks.
2094 * If we are stealing a relatively large buddy page, it is likely there will
2095 * be more free pages in the pageblock, so try to steal them all. For
2096 * reclaimable and unmovable allocations, we steal regardless of page size,
2097 * as fragmentation caused by those allocations polluting movable pageblocks
2098 * is worse than movable allocations stealing from unmovable and reclaimable
2101 static bool can_steal_fallback(unsigned int order, int start_mt)
2104 * Leaving this order check is intended, although there is
2105 * relaxed order check in next check. The reason is that
2106 * we can actually steal whole pageblock if this condition met,
2107 * but, below check doesn't guarantee it and that is just heuristic
2108 * so could be changed anytime.
2110 if (order >= pageblock_order)
2113 if (order >= pageblock_order / 2 ||
2114 start_mt == MIGRATE_RECLAIMABLE ||
2115 start_mt == MIGRATE_UNMOVABLE ||
2116 page_group_by_mobility_disabled)
2123 * This function implements actual steal behaviour. If order is large enough,
2124 * we can steal whole pageblock. If not, we first move freepages in this
2125 * pageblock to our migratetype and determine how many already-allocated pages
2126 * are there in the pageblock with a compatible migratetype. If at least half
2127 * of pages are free or compatible, we can change migratetype of the pageblock
2128 * itself, so pages freed in the future will be put on the correct free list.
2130 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2131 int start_type, bool whole_block)
2133 unsigned int current_order = page_order(page);
2134 struct free_area *area;
2135 int free_pages, movable_pages, alike_pages;
2138 old_block_type = get_pageblock_migratetype(page);
2141 * This can happen due to races and we want to prevent broken
2142 * highatomic accounting.
2144 if (is_migrate_highatomic(old_block_type))
2147 /* Take ownership for orders >= pageblock_order */
2148 if (current_order >= pageblock_order) {
2149 change_pageblock_range(page, current_order, start_type);
2153 /* We are not allowed to try stealing from the whole block */
2157 free_pages = move_freepages_block(zone, page, start_type,
2160 * Determine how many pages are compatible with our allocation.
2161 * For movable allocation, it's the number of movable pages which
2162 * we just obtained. For other types it's a bit more tricky.
2164 if (start_type == MIGRATE_MOVABLE) {
2165 alike_pages = movable_pages;
2168 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2169 * to MOVABLE pageblock, consider all non-movable pages as
2170 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2171 * vice versa, be conservative since we can't distinguish the
2172 * exact migratetype of non-movable pages.
2174 if (old_block_type == MIGRATE_MOVABLE)
2175 alike_pages = pageblock_nr_pages
2176 - (free_pages + movable_pages);
2181 /* moving whole block can fail due to zone boundary conditions */
2186 * If a sufficient number of pages in the block are either free or of
2187 * comparable migratability as our allocation, claim the whole block.
2189 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2190 page_group_by_mobility_disabled)
2191 set_pageblock_migratetype(page, start_type);
2196 area = &zone->free_area[current_order];
2197 list_move(&page->lru, &area->free_list[start_type]);
2201 * Check whether there is a suitable fallback freepage with requested order.
2202 * If only_stealable is true, this function returns fallback_mt only if
2203 * we can steal other freepages all together. This would help to reduce
2204 * fragmentation due to mixed migratetype pages in one pageblock.
2206 int find_suitable_fallback(struct free_area *area, unsigned int order,
2207 int migratetype, bool only_stealable, bool *can_steal)
2212 if (area->nr_free == 0)
2217 fallback_mt = fallbacks[migratetype][i];
2218 if (fallback_mt == MIGRATE_TYPES)
2221 if (list_empty(&area->free_list[fallback_mt]))
2224 if (can_steal_fallback(order, migratetype))
2227 if (!only_stealable)
2238 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2239 * there are no empty page blocks that contain a page with a suitable order
2241 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2242 unsigned int alloc_order)
2245 unsigned long max_managed, flags;
2248 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2249 * Check is race-prone but harmless.
2251 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2252 if (zone->nr_reserved_highatomic >= max_managed)
2255 spin_lock_irqsave(&zone->lock, flags);
2257 /* Recheck the nr_reserved_highatomic limit under the lock */
2258 if (zone->nr_reserved_highatomic >= max_managed)
2262 mt = get_pageblock_migratetype(page);
2263 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2264 && !is_migrate_cma(mt)) {
2265 zone->nr_reserved_highatomic += pageblock_nr_pages;
2266 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2267 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2271 spin_unlock_irqrestore(&zone->lock, flags);
2275 * Used when an allocation is about to fail under memory pressure. This
2276 * potentially hurts the reliability of high-order allocations when under
2277 * intense memory pressure but failed atomic allocations should be easier
2278 * to recover from than an OOM.
2280 * If @force is true, try to unreserve a pageblock even though highatomic
2281 * pageblock is exhausted.
2283 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2286 struct zonelist *zonelist = ac->zonelist;
2287 unsigned long flags;
2294 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2297 * Preserve at least one pageblock unless memory pressure
2300 if (!force && zone->nr_reserved_highatomic <=
2304 spin_lock_irqsave(&zone->lock, flags);
2305 for (order = 0; order < MAX_ORDER; order++) {
2306 struct free_area *area = &(zone->free_area[order]);
2308 page = list_first_entry_or_null(
2309 &area->free_list[MIGRATE_HIGHATOMIC],
2315 * In page freeing path, migratetype change is racy so
2316 * we can counter several free pages in a pageblock
2317 * in this loop althoug we changed the pageblock type
2318 * from highatomic to ac->migratetype. So we should
2319 * adjust the count once.
2321 if (is_migrate_highatomic_page(page)) {
2323 * It should never happen but changes to
2324 * locking could inadvertently allow a per-cpu
2325 * drain to add pages to MIGRATE_HIGHATOMIC
2326 * while unreserving so be safe and watch for
2329 zone->nr_reserved_highatomic -= min(
2331 zone->nr_reserved_highatomic);
2335 * Convert to ac->migratetype and avoid the normal
2336 * pageblock stealing heuristics. Minimally, the caller
2337 * is doing the work and needs the pages. More
2338 * importantly, if the block was always converted to
2339 * MIGRATE_UNMOVABLE or another type then the number
2340 * of pageblocks that cannot be completely freed
2343 set_pageblock_migratetype(page, ac->migratetype);
2344 ret = move_freepages_block(zone, page, ac->migratetype,
2347 spin_unlock_irqrestore(&zone->lock, flags);
2351 spin_unlock_irqrestore(&zone->lock, flags);
2358 * Try finding a free buddy page on the fallback list and put it on the free
2359 * list of requested migratetype, possibly along with other pages from the same
2360 * block, depending on fragmentation avoidance heuristics. Returns true if
2361 * fallback was found so that __rmqueue_smallest() can grab it.
2363 * The use of signed ints for order and current_order is a deliberate
2364 * deviation from the rest of this file, to make the for loop
2365 * condition simpler.
2367 static __always_inline bool
2368 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2370 struct free_area *area;
2377 * Find the largest available free page in the other list. This roughly
2378 * approximates finding the pageblock with the most free pages, which
2379 * would be too costly to do exactly.
2381 for (current_order = MAX_ORDER - 1; current_order >= order;
2383 area = &(zone->free_area[current_order]);
2384 fallback_mt = find_suitable_fallback(area, current_order,
2385 start_migratetype, false, &can_steal);
2386 if (fallback_mt == -1)
2390 * We cannot steal all free pages from the pageblock and the
2391 * requested migratetype is movable. In that case it's better to
2392 * steal and split the smallest available page instead of the
2393 * largest available page, because even if the next movable
2394 * allocation falls back into a different pageblock than this
2395 * one, it won't cause permanent fragmentation.
2397 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2398 && current_order > order)
2407 for (current_order = order; current_order < MAX_ORDER;
2409 area = &(zone->free_area[current_order]);
2410 fallback_mt = find_suitable_fallback(area, current_order,
2411 start_migratetype, false, &can_steal);
2412 if (fallback_mt != -1)
2417 * This should not happen - we already found a suitable fallback
2418 * when looking for the largest page.
2420 VM_BUG_ON(current_order == MAX_ORDER);
2423 page = list_first_entry(&area->free_list[fallback_mt],
2426 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2428 trace_mm_page_alloc_extfrag(page, order, current_order,
2429 start_migratetype, fallback_mt);
2436 * Do the hard work of removing an element from the buddy allocator.
2437 * Call me with the zone->lock already held.
2439 static __always_inline struct page *
2440 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2445 page = __rmqueue_smallest(zone, order, migratetype);
2446 if (unlikely(!page)) {
2447 if (migratetype == MIGRATE_MOVABLE)
2448 page = __rmqueue_cma_fallback(zone, order);
2450 if (!page && __rmqueue_fallback(zone, order, migratetype))
2454 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2459 * Obtain a specified number of elements from the buddy allocator, all under
2460 * a single hold of the lock, for efficiency. Add them to the supplied list.
2461 * Returns the number of new pages which were placed at *list.
2463 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2464 unsigned long count, struct list_head *list,
2469 spin_lock(&zone->lock);
2470 for (i = 0; i < count; ++i) {
2471 struct page *page = __rmqueue(zone, order, migratetype);
2472 if (unlikely(page == NULL))
2475 if (unlikely(check_pcp_refill(page)))
2479 * Split buddy pages returned by expand() are received here in
2480 * physical page order. The page is added to the tail of
2481 * caller's list. From the callers perspective, the linked list
2482 * is ordered by page number under some conditions. This is
2483 * useful for IO devices that can forward direction from the
2484 * head, thus also in the physical page order. This is useful
2485 * for IO devices that can merge IO requests if the physical
2486 * pages are ordered properly.
2488 list_add_tail(&page->lru, list);
2490 if (is_migrate_cma(get_pcppage_migratetype(page)))
2491 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2496 * i pages were removed from the buddy list even if some leak due
2497 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2498 * on i. Do not confuse with 'alloced' which is the number of
2499 * pages added to the pcp list.
2501 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2502 spin_unlock(&zone->lock);
2508 * Called from the vmstat counter updater to drain pagesets of this
2509 * currently executing processor on remote nodes after they have
2512 * Note that this function must be called with the thread pinned to
2513 * a single processor.
2515 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2517 unsigned long flags;
2518 int to_drain, batch;
2520 local_irq_save(flags);
2521 batch = READ_ONCE(pcp->batch);
2522 to_drain = min(pcp->count, batch);
2524 free_pcppages_bulk(zone, to_drain, pcp);
2525 local_irq_restore(flags);
2530 * Drain pcplists of the indicated processor and zone.
2532 * The processor must either be the current processor and the
2533 * thread pinned to the current processor or a processor that
2536 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2538 unsigned long flags;
2539 struct per_cpu_pageset *pset;
2540 struct per_cpu_pages *pcp;
2542 local_irq_save(flags);
2543 pset = per_cpu_ptr(zone->pageset, cpu);
2547 free_pcppages_bulk(zone, pcp->count, pcp);
2548 local_irq_restore(flags);
2552 * Drain pcplists of all zones on the indicated processor.
2554 * The processor must either be the current processor and the
2555 * thread pinned to the current processor or a processor that
2558 static void drain_pages(unsigned int cpu)
2562 for_each_populated_zone(zone) {
2563 drain_pages_zone(cpu, zone);
2568 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2570 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2571 * the single zone's pages.
2573 void drain_local_pages(struct zone *zone)
2575 int cpu = smp_processor_id();
2578 drain_pages_zone(cpu, zone);
2583 static void drain_local_pages_wq(struct work_struct *work)
2586 * drain_all_pages doesn't use proper cpu hotplug protection so
2587 * we can race with cpu offline when the WQ can move this from
2588 * a cpu pinned worker to an unbound one. We can operate on a different
2589 * cpu which is allright but we also have to make sure to not move to
2593 drain_local_pages(NULL);
2598 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2600 * When zone parameter is non-NULL, spill just the single zone's pages.
2602 * Note that this can be extremely slow as the draining happens in a workqueue.
2604 void drain_all_pages(struct zone *zone)
2609 * Allocate in the BSS so we wont require allocation in
2610 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2612 static cpumask_t cpus_with_pcps;
2615 * Make sure nobody triggers this path before mm_percpu_wq is fully
2618 if (WARN_ON_ONCE(!mm_percpu_wq))
2622 * Do not drain if one is already in progress unless it's specific to
2623 * a zone. Such callers are primarily CMA and memory hotplug and need
2624 * the drain to be complete when the call returns.
2626 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2629 mutex_lock(&pcpu_drain_mutex);
2633 * We don't care about racing with CPU hotplug event
2634 * as offline notification will cause the notified
2635 * cpu to drain that CPU pcps and on_each_cpu_mask
2636 * disables preemption as part of its processing
2638 for_each_online_cpu(cpu) {
2639 struct per_cpu_pageset *pcp;
2641 bool has_pcps = false;
2644 pcp = per_cpu_ptr(zone->pageset, cpu);
2648 for_each_populated_zone(z) {
2649 pcp = per_cpu_ptr(z->pageset, cpu);
2650 if (pcp->pcp.count) {
2658 cpumask_set_cpu(cpu, &cpus_with_pcps);
2660 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2663 for_each_cpu(cpu, &cpus_with_pcps) {
2664 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2665 INIT_WORK(work, drain_local_pages_wq);
2666 queue_work_on(cpu, mm_percpu_wq, work);
2668 for_each_cpu(cpu, &cpus_with_pcps)
2669 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2671 mutex_unlock(&pcpu_drain_mutex);
2674 #ifdef CONFIG_HIBERNATION
2677 * Touch the watchdog for every WD_PAGE_COUNT pages.
2679 #define WD_PAGE_COUNT (128*1024)
2681 void mark_free_pages(struct zone *zone)
2683 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2684 unsigned long flags;
2685 unsigned int order, t;
2688 if (zone_is_empty(zone))
2691 spin_lock_irqsave(&zone->lock, flags);
2693 max_zone_pfn = zone_end_pfn(zone);
2694 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2695 if (pfn_valid(pfn)) {
2696 page = pfn_to_page(pfn);
2698 if (!--page_count) {
2699 touch_nmi_watchdog();
2700 page_count = WD_PAGE_COUNT;
2703 if (page_zone(page) != zone)
2706 if (!swsusp_page_is_forbidden(page))
2707 swsusp_unset_page_free(page);
2710 for_each_migratetype_order(order, t) {
2711 list_for_each_entry(page,
2712 &zone->free_area[order].free_list[t], lru) {
2715 pfn = page_to_pfn(page);
2716 for (i = 0; i < (1UL << order); i++) {
2717 if (!--page_count) {
2718 touch_nmi_watchdog();
2719 page_count = WD_PAGE_COUNT;
2721 swsusp_set_page_free(pfn_to_page(pfn + i));
2725 spin_unlock_irqrestore(&zone->lock, flags);
2727 #endif /* CONFIG_PM */
2729 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2733 if (!free_pcp_prepare(page))
2736 migratetype = get_pfnblock_migratetype(page, pfn);
2737 set_pcppage_migratetype(page, migratetype);
2741 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2743 struct zone *zone = page_zone(page);
2744 struct per_cpu_pages *pcp;
2747 migratetype = get_pcppage_migratetype(page);
2748 __count_vm_event(PGFREE);
2751 * We only track unmovable, reclaimable and movable on pcp lists.
2752 * Free ISOLATE pages back to the allocator because they are being
2753 * offlined but treat HIGHATOMIC as movable pages so we can get those
2754 * areas back if necessary. Otherwise, we may have to free
2755 * excessively into the page allocator
2757 if (migratetype >= MIGRATE_PCPTYPES) {
2758 if (unlikely(is_migrate_isolate(migratetype))) {
2759 free_one_page(zone, page, pfn, 0, migratetype);
2762 migratetype = MIGRATE_MOVABLE;
2765 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2766 list_add(&page->lru, &pcp->lists[migratetype]);
2768 if (pcp->count >= pcp->high) {
2769 unsigned long batch = READ_ONCE(pcp->batch);
2770 free_pcppages_bulk(zone, batch, pcp);
2775 * Free a 0-order page
2777 void free_unref_page(struct page *page)
2779 unsigned long flags;
2780 unsigned long pfn = page_to_pfn(page);
2782 if (!free_unref_page_prepare(page, pfn))
2785 local_irq_save(flags);
2786 free_unref_page_commit(page, pfn);
2787 local_irq_restore(flags);
2791 * Free a list of 0-order pages
2793 void free_unref_page_list(struct list_head *list)
2795 struct page *page, *next;
2796 unsigned long flags, pfn;
2797 int batch_count = 0;
2799 /* Prepare pages for freeing */
2800 list_for_each_entry_safe(page, next, list, lru) {
2801 pfn = page_to_pfn(page);
2802 if (!free_unref_page_prepare(page, pfn))
2803 list_del(&page->lru);
2804 set_page_private(page, pfn);
2807 local_irq_save(flags);
2808 list_for_each_entry_safe(page, next, list, lru) {
2809 unsigned long pfn = page_private(page);
2811 set_page_private(page, 0);
2812 trace_mm_page_free_batched(page);
2813 free_unref_page_commit(page, pfn);
2816 * Guard against excessive IRQ disabled times when we get
2817 * a large list of pages to free.
2819 if (++batch_count == SWAP_CLUSTER_MAX) {
2820 local_irq_restore(flags);
2822 local_irq_save(flags);
2825 local_irq_restore(flags);
2829 * split_page takes a non-compound higher-order page, and splits it into
2830 * n (1<<order) sub-pages: page[0..n]
2831 * Each sub-page must be freed individually.
2833 * Note: this is probably too low level an operation for use in drivers.
2834 * Please consult with lkml before using this in your driver.
2836 void split_page(struct page *page, unsigned int order)
2840 VM_BUG_ON_PAGE(PageCompound(page), page);
2841 VM_BUG_ON_PAGE(!page_count(page), page);
2843 for (i = 1; i < (1 << order); i++)
2844 set_page_refcounted(page + i);
2845 split_page_owner(page, order);
2847 EXPORT_SYMBOL_GPL(split_page);
2849 int __isolate_free_page(struct page *page, unsigned int order)
2851 unsigned long watermark;
2855 BUG_ON(!PageBuddy(page));
2857 zone = page_zone(page);
2858 mt = get_pageblock_migratetype(page);
2860 if (!is_migrate_isolate(mt)) {
2862 * Obey watermarks as if the page was being allocated. We can
2863 * emulate a high-order watermark check with a raised order-0
2864 * watermark, because we already know our high-order page
2867 watermark = min_wmark_pages(zone) + (1UL << order);
2868 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2871 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2874 /* Remove page from free list */
2875 list_del(&page->lru);
2876 zone->free_area[order].nr_free--;
2877 rmv_page_order(page);
2880 * Set the pageblock if the isolated page is at least half of a
2883 if (order >= pageblock_order - 1) {
2884 struct page *endpage = page + (1 << order) - 1;
2885 for (; page < endpage; page += pageblock_nr_pages) {
2886 int mt = get_pageblock_migratetype(page);
2887 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2888 && !is_migrate_highatomic(mt))
2889 set_pageblock_migratetype(page,
2895 return 1UL << order;
2899 * Update NUMA hit/miss statistics
2901 * Must be called with interrupts disabled.
2903 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2906 enum numa_stat_item local_stat = NUMA_LOCAL;
2908 /* skip numa counters update if numa stats is disabled */
2909 if (!static_branch_likely(&vm_numa_stat_key))
2912 if (zone_to_nid(z) != numa_node_id())
2913 local_stat = NUMA_OTHER;
2915 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2916 __inc_numa_state(z, NUMA_HIT);
2918 __inc_numa_state(z, NUMA_MISS);
2919 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2921 __inc_numa_state(z, local_stat);
2925 /* Remove page from the per-cpu list, caller must protect the list */
2926 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2927 struct per_cpu_pages *pcp,
2928 struct list_head *list)
2933 if (list_empty(list)) {
2934 pcp->count += rmqueue_bulk(zone, 0,
2937 if (unlikely(list_empty(list)))
2941 page = list_first_entry(list, struct page, lru);
2942 list_del(&page->lru);
2944 } while (check_new_pcp(page));
2949 /* Lock and remove page from the per-cpu list */
2950 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2951 struct zone *zone, unsigned int order,
2952 gfp_t gfp_flags, int migratetype)
2954 struct per_cpu_pages *pcp;
2955 struct list_head *list;
2957 unsigned long flags;
2959 local_irq_save(flags);
2960 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2961 list = &pcp->lists[migratetype];
2962 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2964 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2965 zone_statistics(preferred_zone, zone);
2967 local_irq_restore(flags);
2972 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2975 struct page *rmqueue(struct zone *preferred_zone,
2976 struct zone *zone, unsigned int order,
2977 gfp_t gfp_flags, unsigned int alloc_flags,
2980 unsigned long flags;
2983 if (likely(order == 0)) {
2984 page = rmqueue_pcplist(preferred_zone, zone, order,
2985 gfp_flags, migratetype);
2990 * We most definitely don't want callers attempting to
2991 * allocate greater than order-1 page units with __GFP_NOFAIL.
2993 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2994 spin_lock_irqsave(&zone->lock, flags);
2998 if (alloc_flags & ALLOC_HARDER) {
2999 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3001 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3004 page = __rmqueue(zone, order, migratetype);
3005 } while (page && check_new_pages(page, order));
3006 spin_unlock(&zone->lock);
3009 __mod_zone_freepage_state(zone, -(1 << order),
3010 get_pcppage_migratetype(page));
3012 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3013 zone_statistics(preferred_zone, zone);
3014 local_irq_restore(flags);
3017 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3021 local_irq_restore(flags);
3025 #ifdef CONFIG_FAIL_PAGE_ALLOC
3028 struct fault_attr attr;
3030 bool ignore_gfp_highmem;
3031 bool ignore_gfp_reclaim;
3033 } fail_page_alloc = {
3034 .attr = FAULT_ATTR_INITIALIZER,
3035 .ignore_gfp_reclaim = true,
3036 .ignore_gfp_highmem = true,
3040 static int __init setup_fail_page_alloc(char *str)
3042 return setup_fault_attr(&fail_page_alloc.attr, str);
3044 __setup("fail_page_alloc=", setup_fail_page_alloc);
3046 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3048 if (order < fail_page_alloc.min_order)
3050 if (gfp_mask & __GFP_NOFAIL)
3052 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3054 if (fail_page_alloc.ignore_gfp_reclaim &&
3055 (gfp_mask & __GFP_DIRECT_RECLAIM))
3058 return should_fail(&fail_page_alloc.attr, 1 << order);
3061 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3063 static int __init fail_page_alloc_debugfs(void)
3065 umode_t mode = S_IFREG | 0600;
3068 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3069 &fail_page_alloc.attr);
3071 return PTR_ERR(dir);
3073 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3074 &fail_page_alloc.ignore_gfp_reclaim))
3076 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3077 &fail_page_alloc.ignore_gfp_highmem))
3079 if (!debugfs_create_u32("min-order", mode, dir,
3080 &fail_page_alloc.min_order))
3085 debugfs_remove_recursive(dir);
3090 late_initcall(fail_page_alloc_debugfs);
3092 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3094 #else /* CONFIG_FAIL_PAGE_ALLOC */
3096 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3101 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3104 * Return true if free base pages are above 'mark'. For high-order checks it
3105 * will return true of the order-0 watermark is reached and there is at least
3106 * one free page of a suitable size. Checking now avoids taking the zone lock
3107 * to check in the allocation paths if no pages are free.
3109 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3110 int classzone_idx, unsigned int alloc_flags,
3115 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3117 /* free_pages may go negative - that's OK */
3118 free_pages -= (1 << order) - 1;
3120 if (alloc_flags & ALLOC_HIGH)
3124 * If the caller does not have rights to ALLOC_HARDER then subtract
3125 * the high-atomic reserves. This will over-estimate the size of the
3126 * atomic reserve but it avoids a search.
3128 if (likely(!alloc_harder)) {
3129 free_pages -= z->nr_reserved_highatomic;
3132 * OOM victims can try even harder than normal ALLOC_HARDER
3133 * users on the grounds that it's definitely going to be in
3134 * the exit path shortly and free memory. Any allocation it
3135 * makes during the free path will be small and short-lived.
3137 if (alloc_flags & ALLOC_OOM)
3145 /* If allocation can't use CMA areas don't use free CMA pages */
3146 if (!(alloc_flags & ALLOC_CMA))
3147 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3151 * Check watermarks for an order-0 allocation request. If these
3152 * are not met, then a high-order request also cannot go ahead
3153 * even if a suitable page happened to be free.
3155 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3158 /* If this is an order-0 request then the watermark is fine */
3162 /* For a high-order request, check at least one suitable page is free */
3163 for (o = order; o < MAX_ORDER; o++) {
3164 struct free_area *area = &z->free_area[o];
3170 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3171 if (!list_empty(&area->free_list[mt]))
3176 if ((alloc_flags & ALLOC_CMA) &&
3177 !list_empty(&area->free_list[MIGRATE_CMA])) {
3182 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3188 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3189 int classzone_idx, unsigned int alloc_flags)
3191 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3192 zone_page_state(z, NR_FREE_PAGES));
3195 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3196 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3198 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3202 /* If allocation can't use CMA areas don't use free CMA pages */
3203 if (!(alloc_flags & ALLOC_CMA))
3204 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3208 * Fast check for order-0 only. If this fails then the reserves
3209 * need to be calculated. There is a corner case where the check
3210 * passes but only the high-order atomic reserve are free. If
3211 * the caller is !atomic then it'll uselessly search the free
3212 * list. That corner case is then slower but it is harmless.
3214 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3217 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3221 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3222 unsigned long mark, int classzone_idx)
3224 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3226 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3227 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3229 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3234 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3236 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3239 #else /* CONFIG_NUMA */
3240 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3244 #endif /* CONFIG_NUMA */
3247 * get_page_from_freelist goes through the zonelist trying to allocate
3250 static struct page *
3251 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3252 const struct alloc_context *ac)
3254 struct zoneref *z = ac->preferred_zoneref;
3256 struct pglist_data *last_pgdat_dirty_limit = NULL;
3259 * Scan zonelist, looking for a zone with enough free.
3260 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3262 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3267 if (cpusets_enabled() &&
3268 (alloc_flags & ALLOC_CPUSET) &&
3269 !__cpuset_zone_allowed(zone, gfp_mask))
3272 * When allocating a page cache page for writing, we
3273 * want to get it from a node that is within its dirty
3274 * limit, such that no single node holds more than its
3275 * proportional share of globally allowed dirty pages.
3276 * The dirty limits take into account the node's
3277 * lowmem reserves and high watermark so that kswapd
3278 * should be able to balance it without having to
3279 * write pages from its LRU list.
3281 * XXX: For now, allow allocations to potentially
3282 * exceed the per-node dirty limit in the slowpath
3283 * (spread_dirty_pages unset) before going into reclaim,
3284 * which is important when on a NUMA setup the allowed
3285 * nodes are together not big enough to reach the
3286 * global limit. The proper fix for these situations
3287 * will require awareness of nodes in the
3288 * dirty-throttling and the flusher threads.
3290 if (ac->spread_dirty_pages) {
3291 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3294 if (!node_dirty_ok(zone->zone_pgdat)) {
3295 last_pgdat_dirty_limit = zone->zone_pgdat;
3300 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3301 if (!zone_watermark_fast(zone, order, mark,
3302 ac_classzone_idx(ac), alloc_flags)) {
3305 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3307 * Watermark failed for this zone, but see if we can
3308 * grow this zone if it contains deferred pages.
3310 if (static_branch_unlikely(&deferred_pages)) {
3311 if (_deferred_grow_zone(zone, order))
3315 /* Checked here to keep the fast path fast */
3316 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3317 if (alloc_flags & ALLOC_NO_WATERMARKS)
3320 if (node_reclaim_mode == 0 ||
3321 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3324 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3326 case NODE_RECLAIM_NOSCAN:
3329 case NODE_RECLAIM_FULL:
3330 /* scanned but unreclaimable */
3333 /* did we reclaim enough */
3334 if (zone_watermark_ok(zone, order, mark,
3335 ac_classzone_idx(ac), alloc_flags))
3343 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3344 gfp_mask, alloc_flags, ac->migratetype);
3346 prep_new_page(page, order, gfp_mask, alloc_flags);
3349 * If this is a high-order atomic allocation then check
3350 * if the pageblock should be reserved for the future
3352 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3353 reserve_highatomic_pageblock(page, zone, order);
3357 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3358 /* Try again if zone has deferred pages */
3359 if (static_branch_unlikely(&deferred_pages)) {
3360 if (_deferred_grow_zone(zone, order))
3371 * Large machines with many possible nodes should not always dump per-node
3372 * meminfo in irq context.
3374 static inline bool should_suppress_show_mem(void)
3379 ret = in_interrupt();
3384 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3386 unsigned int filter = SHOW_MEM_FILTER_NODES;
3387 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3389 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3393 * This documents exceptions given to allocations in certain
3394 * contexts that are allowed to allocate outside current's set
3397 if (!(gfp_mask & __GFP_NOMEMALLOC))
3398 if (tsk_is_oom_victim(current) ||
3399 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3400 filter &= ~SHOW_MEM_FILTER_NODES;
3401 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3402 filter &= ~SHOW_MEM_FILTER_NODES;
3404 show_mem(filter, nodemask);
3407 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3409 struct va_format vaf;
3411 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3412 DEFAULT_RATELIMIT_BURST);
3414 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3417 va_start(args, fmt);
3420 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3421 current->comm, &vaf, gfp_mask, &gfp_mask,
3422 nodemask_pr_args(nodemask));
3425 cpuset_print_current_mems_allowed();
3428 warn_alloc_show_mem(gfp_mask, nodemask);
3431 static inline struct page *
3432 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3433 unsigned int alloc_flags,
3434 const struct alloc_context *ac)
3438 page = get_page_from_freelist(gfp_mask, order,
3439 alloc_flags|ALLOC_CPUSET, ac);
3441 * fallback to ignore cpuset restriction if our nodes
3445 page = get_page_from_freelist(gfp_mask, order,
3451 static inline struct page *
3452 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3453 const struct alloc_context *ac, unsigned long *did_some_progress)
3455 struct oom_control oc = {
3456 .zonelist = ac->zonelist,
3457 .nodemask = ac->nodemask,
3459 .gfp_mask = gfp_mask,
3464 *did_some_progress = 0;
3467 * Acquire the oom lock. If that fails, somebody else is
3468 * making progress for us.
3470 if (!mutex_trylock(&oom_lock)) {
3471 *did_some_progress = 1;
3472 schedule_timeout_uninterruptible(1);
3477 * Go through the zonelist yet one more time, keep very high watermark
3478 * here, this is only to catch a parallel oom killing, we must fail if
3479 * we're still under heavy pressure. But make sure that this reclaim
3480 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3481 * allocation which will never fail due to oom_lock already held.
3483 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3484 ~__GFP_DIRECT_RECLAIM, order,
3485 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3489 /* Coredumps can quickly deplete all memory reserves */
3490 if (current->flags & PF_DUMPCORE)
3492 /* The OOM killer will not help higher order allocs */
3493 if (order > PAGE_ALLOC_COSTLY_ORDER)
3496 * We have already exhausted all our reclaim opportunities without any
3497 * success so it is time to admit defeat. We will skip the OOM killer
3498 * because it is very likely that the caller has a more reasonable
3499 * fallback than shooting a random task.
3501 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3503 /* The OOM killer does not needlessly kill tasks for lowmem */
3504 if (ac->high_zoneidx < ZONE_NORMAL)
3506 if (pm_suspended_storage())
3509 * XXX: GFP_NOFS allocations should rather fail than rely on
3510 * other request to make a forward progress.
3511 * We are in an unfortunate situation where out_of_memory cannot
3512 * do much for this context but let's try it to at least get
3513 * access to memory reserved if the current task is killed (see
3514 * out_of_memory). Once filesystems are ready to handle allocation
3515 * failures more gracefully we should just bail out here.
3518 /* The OOM killer may not free memory on a specific node */
3519 if (gfp_mask & __GFP_THISNODE)
3522 /* Exhausted what can be done so it's blame time */
3523 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3524 *did_some_progress = 1;
3527 * Help non-failing allocations by giving them access to memory
3530 if (gfp_mask & __GFP_NOFAIL)
3531 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3532 ALLOC_NO_WATERMARKS, ac);
3535 mutex_unlock(&oom_lock);
3540 * Maximum number of compaction retries wit a progress before OOM
3541 * killer is consider as the only way to move forward.
3543 #define MAX_COMPACT_RETRIES 16
3545 #ifdef CONFIG_COMPACTION
3546 /* Try memory compaction for high-order allocations before reclaim */
3547 static struct page *
3548 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3549 unsigned int alloc_flags, const struct alloc_context *ac,
3550 enum compact_priority prio, enum compact_result *compact_result)
3553 unsigned int noreclaim_flag;
3558 noreclaim_flag = memalloc_noreclaim_save();
3559 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3561 memalloc_noreclaim_restore(noreclaim_flag);
3563 if (*compact_result <= COMPACT_INACTIVE)
3567 * At least in one zone compaction wasn't deferred or skipped, so let's
3568 * count a compaction stall
3570 count_vm_event(COMPACTSTALL);
3572 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3575 struct zone *zone = page_zone(page);
3577 zone->compact_blockskip_flush = false;
3578 compaction_defer_reset(zone, order, true);
3579 count_vm_event(COMPACTSUCCESS);
3584 * It's bad if compaction run occurs and fails. The most likely reason
3585 * is that pages exist, but not enough to satisfy watermarks.
3587 count_vm_event(COMPACTFAIL);
3595 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3596 enum compact_result compact_result,
3597 enum compact_priority *compact_priority,
3598 int *compaction_retries)
3600 int max_retries = MAX_COMPACT_RETRIES;
3603 int retries = *compaction_retries;
3604 enum compact_priority priority = *compact_priority;
3609 if (compaction_made_progress(compact_result))
3610 (*compaction_retries)++;
3613 * compaction considers all the zone as desperately out of memory
3614 * so it doesn't really make much sense to retry except when the
3615 * failure could be caused by insufficient priority
3617 if (compaction_failed(compact_result))
3618 goto check_priority;
3621 * make sure the compaction wasn't deferred or didn't bail out early
3622 * due to locks contention before we declare that we should give up.
3623 * But do not retry if the given zonelist is not suitable for
3626 if (compaction_withdrawn(compact_result)) {
3627 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3632 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3633 * costly ones because they are de facto nofail and invoke OOM
3634 * killer to move on while costly can fail and users are ready
3635 * to cope with that. 1/4 retries is rather arbitrary but we
3636 * would need much more detailed feedback from compaction to
3637 * make a better decision.
3639 if (order > PAGE_ALLOC_COSTLY_ORDER)
3641 if (*compaction_retries <= max_retries) {
3647 * Make sure there are attempts at the highest priority if we exhausted
3648 * all retries or failed at the lower priorities.
3651 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3652 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3654 if (*compact_priority > min_priority) {
3655 (*compact_priority)--;
3656 *compaction_retries = 0;
3660 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3664 static inline struct page *
3665 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3666 unsigned int alloc_flags, const struct alloc_context *ac,
3667 enum compact_priority prio, enum compact_result *compact_result)
3669 *compact_result = COMPACT_SKIPPED;
3674 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3675 enum compact_result compact_result,
3676 enum compact_priority *compact_priority,
3677 int *compaction_retries)
3682 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3686 * There are setups with compaction disabled which would prefer to loop
3687 * inside the allocator rather than hit the oom killer prematurely.
3688 * Let's give them a good hope and keep retrying while the order-0
3689 * watermarks are OK.
3691 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3693 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3694 ac_classzone_idx(ac), alloc_flags))
3699 #endif /* CONFIG_COMPACTION */
3701 #ifdef CONFIG_LOCKDEP
3702 static struct lockdep_map __fs_reclaim_map =
3703 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3705 static bool __need_fs_reclaim(gfp_t gfp_mask)
3707 gfp_mask = current_gfp_context(gfp_mask);
3709 /* no reclaim without waiting on it */
3710 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3713 /* this guy won't enter reclaim */
3714 if (current->flags & PF_MEMALLOC)
3717 /* We're only interested __GFP_FS allocations for now */
3718 if (!(gfp_mask & __GFP_FS))
3721 if (gfp_mask & __GFP_NOLOCKDEP)
3727 void __fs_reclaim_acquire(void)
3729 lock_map_acquire(&__fs_reclaim_map);
3732 void __fs_reclaim_release(void)
3734 lock_map_release(&__fs_reclaim_map);
3737 void fs_reclaim_acquire(gfp_t gfp_mask)
3739 if (__need_fs_reclaim(gfp_mask))
3740 __fs_reclaim_acquire();
3742 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3744 void fs_reclaim_release(gfp_t gfp_mask)
3746 if (__need_fs_reclaim(gfp_mask))
3747 __fs_reclaim_release();
3749 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3752 /* Perform direct synchronous page reclaim */
3754 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3755 const struct alloc_context *ac)
3757 struct reclaim_state reclaim_state;
3759 unsigned int noreclaim_flag;
3763 /* We now go into synchronous reclaim */
3764 cpuset_memory_pressure_bump();
3765 fs_reclaim_acquire(gfp_mask);
3766 noreclaim_flag = memalloc_noreclaim_save();
3767 reclaim_state.reclaimed_slab = 0;
3768 current->reclaim_state = &reclaim_state;
3770 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3773 current->reclaim_state = NULL;
3774 memalloc_noreclaim_restore(noreclaim_flag);
3775 fs_reclaim_release(gfp_mask);
3782 /* The really slow allocator path where we enter direct reclaim */
3783 static inline struct page *
3784 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3785 unsigned int alloc_flags, const struct alloc_context *ac,
3786 unsigned long *did_some_progress)
3788 struct page *page = NULL;
3789 bool drained = false;
3791 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3792 if (unlikely(!(*did_some_progress)))
3796 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3799 * If an allocation failed after direct reclaim, it could be because
3800 * pages are pinned on the per-cpu lists or in high alloc reserves.
3801 * Shrink them them and try again
3803 if (!page && !drained) {
3804 unreserve_highatomic_pageblock(ac, false);
3805 drain_all_pages(NULL);
3813 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3814 const struct alloc_context *ac)
3818 pg_data_t *last_pgdat = NULL;
3819 enum zone_type high_zoneidx = ac->high_zoneidx;
3821 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3823 if (last_pgdat != zone->zone_pgdat)
3824 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3825 last_pgdat = zone->zone_pgdat;
3829 static inline unsigned int
3830 gfp_to_alloc_flags(gfp_t gfp_mask)
3832 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3834 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3835 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3838 * The caller may dip into page reserves a bit more if the caller
3839 * cannot run direct reclaim, or if the caller has realtime scheduling
3840 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3841 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3843 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3845 if (gfp_mask & __GFP_ATOMIC) {
3847 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3848 * if it can't schedule.
3850 if (!(gfp_mask & __GFP_NOMEMALLOC))
3851 alloc_flags |= ALLOC_HARDER;
3853 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3854 * comment for __cpuset_node_allowed().
3856 alloc_flags &= ~ALLOC_CPUSET;
3857 } else if (unlikely(rt_task(current)) && !in_interrupt())
3858 alloc_flags |= ALLOC_HARDER;
3861 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3862 alloc_flags |= ALLOC_CMA;
3867 static bool oom_reserves_allowed(struct task_struct *tsk)
3869 if (!tsk_is_oom_victim(tsk))
3873 * !MMU doesn't have oom reaper so give access to memory reserves
3874 * only to the thread with TIF_MEMDIE set
3876 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3883 * Distinguish requests which really need access to full memory
3884 * reserves from oom victims which can live with a portion of it
3886 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3888 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3890 if (gfp_mask & __GFP_MEMALLOC)
3891 return ALLOC_NO_WATERMARKS;
3892 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3893 return ALLOC_NO_WATERMARKS;
3894 if (!in_interrupt()) {
3895 if (current->flags & PF_MEMALLOC)
3896 return ALLOC_NO_WATERMARKS;
3897 else if (oom_reserves_allowed(current))
3904 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3906 return !!__gfp_pfmemalloc_flags(gfp_mask);
3910 * Checks whether it makes sense to retry the reclaim to make a forward progress
3911 * for the given allocation request.
3913 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3914 * without success, or when we couldn't even meet the watermark if we
3915 * reclaimed all remaining pages on the LRU lists.
3917 * Returns true if a retry is viable or false to enter the oom path.
3920 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3921 struct alloc_context *ac, int alloc_flags,
3922 bool did_some_progress, int *no_progress_loops)
3928 * Costly allocations might have made a progress but this doesn't mean
3929 * their order will become available due to high fragmentation so
3930 * always increment the no progress counter for them
3932 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3933 *no_progress_loops = 0;
3935 (*no_progress_loops)++;
3938 * Make sure we converge to OOM if we cannot make any progress
3939 * several times in the row.
3941 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3942 /* Before OOM, exhaust highatomic_reserve */
3943 return unreserve_highatomic_pageblock(ac, true);
3947 * Keep reclaiming pages while there is a chance this will lead
3948 * somewhere. If none of the target zones can satisfy our allocation
3949 * request even if all reclaimable pages are considered then we are
3950 * screwed and have to go OOM.
3952 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3954 unsigned long available;
3955 unsigned long reclaimable;
3956 unsigned long min_wmark = min_wmark_pages(zone);
3959 available = reclaimable = zone_reclaimable_pages(zone);
3960 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3963 * Would the allocation succeed if we reclaimed all
3964 * reclaimable pages?
3966 wmark = __zone_watermark_ok(zone, order, min_wmark,
3967 ac_classzone_idx(ac), alloc_flags, available);
3968 trace_reclaim_retry_zone(z, order, reclaimable,
3969 available, min_wmark, *no_progress_loops, wmark);
3972 * If we didn't make any progress and have a lot of
3973 * dirty + writeback pages then we should wait for
3974 * an IO to complete to slow down the reclaim and
3975 * prevent from pre mature OOM
3977 if (!did_some_progress) {
3978 unsigned long write_pending;
3980 write_pending = zone_page_state_snapshot(zone,
3981 NR_ZONE_WRITE_PENDING);
3983 if (2 * write_pending > reclaimable) {
3984 congestion_wait(BLK_RW_ASYNC, HZ/10);
3990 * Memory allocation/reclaim might be called from a WQ
3991 * context and the current implementation of the WQ
3992 * concurrency control doesn't recognize that
3993 * a particular WQ is congested if the worker thread is
3994 * looping without ever sleeping. Therefore we have to
3995 * do a short sleep here rather than calling
3998 if (current->flags & PF_WQ_WORKER)
3999 schedule_timeout_uninterruptible(1);
4011 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4014 * It's possible that cpuset's mems_allowed and the nodemask from
4015 * mempolicy don't intersect. This should be normally dealt with by
4016 * policy_nodemask(), but it's possible to race with cpuset update in
4017 * such a way the check therein was true, and then it became false
4018 * before we got our cpuset_mems_cookie here.
4019 * This assumes that for all allocations, ac->nodemask can come only
4020 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4021 * when it does not intersect with the cpuset restrictions) or the
4022 * caller can deal with a violated nodemask.
4024 if (cpusets_enabled() && ac->nodemask &&
4025 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4026 ac->nodemask = NULL;
4031 * When updating a task's mems_allowed or mempolicy nodemask, it is
4032 * possible to race with parallel threads in such a way that our
4033 * allocation can fail while the mask is being updated. If we are about
4034 * to fail, check if the cpuset changed during allocation and if so,
4037 if (read_mems_allowed_retry(cpuset_mems_cookie))
4043 static inline struct page *
4044 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4045 struct alloc_context *ac)
4047 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4048 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4049 struct page *page = NULL;
4050 unsigned int alloc_flags;
4051 unsigned long did_some_progress;
4052 enum compact_priority compact_priority;
4053 enum compact_result compact_result;
4054 int compaction_retries;
4055 int no_progress_loops;
4056 unsigned int cpuset_mems_cookie;
4060 * In the slowpath, we sanity check order to avoid ever trying to
4061 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4062 * be using allocators in order of preference for an area that is
4065 if (order >= MAX_ORDER) {
4066 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4071 * We also sanity check to catch abuse of atomic reserves being used by
4072 * callers that are not in atomic context.
4074 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4075 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4076 gfp_mask &= ~__GFP_ATOMIC;
4079 compaction_retries = 0;
4080 no_progress_loops = 0;
4081 compact_priority = DEF_COMPACT_PRIORITY;
4082 cpuset_mems_cookie = read_mems_allowed_begin();
4085 * The fast path uses conservative alloc_flags to succeed only until
4086 * kswapd needs to be woken up, and to avoid the cost of setting up
4087 * alloc_flags precisely. So we do that now.
4089 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4092 * We need to recalculate the starting point for the zonelist iterator
4093 * because we might have used different nodemask in the fast path, or
4094 * there was a cpuset modification and we are retrying - otherwise we
4095 * could end up iterating over non-eligible zones endlessly.
4097 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4098 ac->high_zoneidx, ac->nodemask);
4099 if (!ac->preferred_zoneref->zone)
4102 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4103 wake_all_kswapds(order, gfp_mask, ac);
4106 * The adjusted alloc_flags might result in immediate success, so try
4109 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4114 * For costly allocations, try direct compaction first, as it's likely
4115 * that we have enough base pages and don't need to reclaim. For non-
4116 * movable high-order allocations, do that as well, as compaction will
4117 * try prevent permanent fragmentation by migrating from blocks of the
4119 * Don't try this for allocations that are allowed to ignore
4120 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4122 if (can_direct_reclaim &&
4124 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4125 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4126 page = __alloc_pages_direct_compact(gfp_mask, order,
4128 INIT_COMPACT_PRIORITY,
4134 * Checks for costly allocations with __GFP_NORETRY, which
4135 * includes THP page fault allocations
4137 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4139 * If compaction is deferred for high-order allocations,
4140 * it is because sync compaction recently failed. If
4141 * this is the case and the caller requested a THP
4142 * allocation, we do not want to heavily disrupt the
4143 * system, so we fail the allocation instead of entering
4146 if (compact_result == COMPACT_DEFERRED)
4150 * Looks like reclaim/compaction is worth trying, but
4151 * sync compaction could be very expensive, so keep
4152 * using async compaction.
4154 compact_priority = INIT_COMPACT_PRIORITY;
4159 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4160 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4161 wake_all_kswapds(order, gfp_mask, ac);
4163 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4165 alloc_flags = reserve_flags;
4168 * Reset the nodemask and zonelist iterators if memory policies can be
4169 * ignored. These allocations are high priority and system rather than
4172 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4173 ac->nodemask = NULL;
4174 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4175 ac->high_zoneidx, ac->nodemask);
4178 /* Attempt with potentially adjusted zonelist and alloc_flags */
4179 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4183 /* Caller is not willing to reclaim, we can't balance anything */
4184 if (!can_direct_reclaim)
4187 /* Avoid recursion of direct reclaim */
4188 if (current->flags & PF_MEMALLOC)
4191 /* Try direct reclaim and then allocating */
4192 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4193 &did_some_progress);
4197 /* Try direct compaction and then allocating */
4198 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4199 compact_priority, &compact_result);
4203 /* Do not loop if specifically requested */
4204 if (gfp_mask & __GFP_NORETRY)
4208 * Do not retry costly high order allocations unless they are
4209 * __GFP_RETRY_MAYFAIL
4211 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4214 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4215 did_some_progress > 0, &no_progress_loops))
4219 * It doesn't make any sense to retry for the compaction if the order-0
4220 * reclaim is not able to make any progress because the current
4221 * implementation of the compaction depends on the sufficient amount
4222 * of free memory (see __compaction_suitable)
4224 if (did_some_progress > 0 &&
4225 should_compact_retry(ac, order, alloc_flags,
4226 compact_result, &compact_priority,
4227 &compaction_retries))
4231 /* Deal with possible cpuset update races before we start OOM killing */
4232 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4235 /* Reclaim has failed us, start killing things */
4236 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4240 /* Avoid allocations with no watermarks from looping endlessly */
4241 if (tsk_is_oom_victim(current) &&
4242 (alloc_flags == ALLOC_OOM ||
4243 (gfp_mask & __GFP_NOMEMALLOC)))
4246 /* Retry as long as the OOM killer is making progress */
4247 if (did_some_progress) {
4248 no_progress_loops = 0;
4253 /* Deal with possible cpuset update races before we fail */
4254 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4258 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4261 if (gfp_mask & __GFP_NOFAIL) {
4263 * All existing users of the __GFP_NOFAIL are blockable, so warn
4264 * of any new users that actually require GFP_NOWAIT
4266 if (WARN_ON_ONCE(!can_direct_reclaim))
4270 * PF_MEMALLOC request from this context is rather bizarre
4271 * because we cannot reclaim anything and only can loop waiting
4272 * for somebody to do a work for us
4274 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4277 * non failing costly orders are a hard requirement which we
4278 * are not prepared for much so let's warn about these users
4279 * so that we can identify them and convert them to something
4282 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4285 * Help non-failing allocations by giving them access to memory
4286 * reserves but do not use ALLOC_NO_WATERMARKS because this
4287 * could deplete whole memory reserves which would just make
4288 * the situation worse
4290 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4298 warn_alloc(gfp_mask, ac->nodemask,
4299 "page allocation failure: order:%u", order);
4304 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4305 int preferred_nid, nodemask_t *nodemask,
4306 struct alloc_context *ac, gfp_t *alloc_mask,
4307 unsigned int *alloc_flags)
4309 ac->high_zoneidx = gfp_zone(gfp_mask);
4310 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4311 ac->nodemask = nodemask;
4312 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4314 if (cpusets_enabled()) {
4315 *alloc_mask |= __GFP_HARDWALL;
4317 ac->nodemask = &cpuset_current_mems_allowed;
4319 *alloc_flags |= ALLOC_CPUSET;
4322 fs_reclaim_acquire(gfp_mask);
4323 fs_reclaim_release(gfp_mask);
4325 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4327 if (should_fail_alloc_page(gfp_mask, order))
4330 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4331 *alloc_flags |= ALLOC_CMA;
4336 /* Determine whether to spread dirty pages and what the first usable zone */
4337 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4339 /* Dirty zone balancing only done in the fast path */
4340 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4343 * The preferred zone is used for statistics but crucially it is
4344 * also used as the starting point for the zonelist iterator. It
4345 * may get reset for allocations that ignore memory policies.
4347 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4348 ac->high_zoneidx, ac->nodemask);
4352 * This is the 'heart' of the zoned buddy allocator.
4355 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4356 nodemask_t *nodemask)
4359 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4360 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4361 struct alloc_context ac = { };
4363 gfp_mask &= gfp_allowed_mask;
4364 alloc_mask = gfp_mask;
4365 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4368 finalise_ac(gfp_mask, &ac);
4370 /* First allocation attempt */
4371 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4376 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4377 * resp. GFP_NOIO which has to be inherited for all allocation requests
4378 * from a particular context which has been marked by
4379 * memalloc_no{fs,io}_{save,restore}.
4381 alloc_mask = current_gfp_context(gfp_mask);
4382 ac.spread_dirty_pages = false;
4385 * Restore the original nodemask if it was potentially replaced with
4386 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4388 if (unlikely(ac.nodemask != nodemask))
4389 ac.nodemask = nodemask;
4391 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4394 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4395 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4396 __free_pages(page, order);
4400 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4404 EXPORT_SYMBOL(__alloc_pages_nodemask);
4407 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4408 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4409 * you need to access high mem.
4411 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4415 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4418 return (unsigned long) page_address(page);
4420 EXPORT_SYMBOL(__get_free_pages);
4422 unsigned long get_zeroed_page(gfp_t gfp_mask)
4424 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4426 EXPORT_SYMBOL(get_zeroed_page);
4428 void __free_pages(struct page *page, unsigned int order)
4430 if (put_page_testzero(page)) {
4432 free_unref_page(page);
4434 __free_pages_ok(page, order);
4438 EXPORT_SYMBOL(__free_pages);
4440 void free_pages(unsigned long addr, unsigned int order)
4443 VM_BUG_ON(!virt_addr_valid((void *)addr));
4444 __free_pages(virt_to_page((void *)addr), order);
4448 EXPORT_SYMBOL(free_pages);
4452 * An arbitrary-length arbitrary-offset area of memory which resides
4453 * within a 0 or higher order page. Multiple fragments within that page
4454 * are individually refcounted, in the page's reference counter.
4456 * The page_frag functions below provide a simple allocation framework for
4457 * page fragments. This is used by the network stack and network device
4458 * drivers to provide a backing region of memory for use as either an
4459 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4461 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4464 struct page *page = NULL;
4465 gfp_t gfp = gfp_mask;
4467 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4468 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4470 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4471 PAGE_FRAG_CACHE_MAX_ORDER);
4472 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4474 if (unlikely(!page))
4475 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4477 nc->va = page ? page_address(page) : NULL;
4482 void __page_frag_cache_drain(struct page *page, unsigned int count)
4484 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4486 if (page_ref_sub_and_test(page, count)) {
4487 unsigned int order = compound_order(page);
4490 free_unref_page(page);
4492 __free_pages_ok(page, order);
4495 EXPORT_SYMBOL(__page_frag_cache_drain);
4497 void *page_frag_alloc(struct page_frag_cache *nc,
4498 unsigned int fragsz, gfp_t gfp_mask)
4500 unsigned int size = PAGE_SIZE;
4504 if (unlikely(!nc->va)) {
4506 page = __page_frag_cache_refill(nc, gfp_mask);
4510 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4511 /* if size can vary use size else just use PAGE_SIZE */
4514 /* Even if we own the page, we do not use atomic_set().
4515 * This would break get_page_unless_zero() users.
4517 page_ref_add(page, size - 1);
4519 /* reset page count bias and offset to start of new frag */
4520 nc->pfmemalloc = page_is_pfmemalloc(page);
4521 nc->pagecnt_bias = size;
4525 offset = nc->offset - fragsz;
4526 if (unlikely(offset < 0)) {
4527 page = virt_to_page(nc->va);
4529 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4532 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4533 /* if size can vary use size else just use PAGE_SIZE */
4536 /* OK, page count is 0, we can safely set it */
4537 set_page_count(page, size);
4539 /* reset page count bias and offset to start of new frag */
4540 nc->pagecnt_bias = size;
4541 offset = size - fragsz;
4545 nc->offset = offset;
4547 return nc->va + offset;
4549 EXPORT_SYMBOL(page_frag_alloc);
4552 * Frees a page fragment allocated out of either a compound or order 0 page.
4554 void page_frag_free(void *addr)
4556 struct page *page = virt_to_head_page(addr);
4558 if (unlikely(put_page_testzero(page)))
4559 __free_pages_ok(page, compound_order(page));
4561 EXPORT_SYMBOL(page_frag_free);
4563 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4567 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4568 unsigned long used = addr + PAGE_ALIGN(size);
4570 split_page(virt_to_page((void *)addr), order);
4571 while (used < alloc_end) {
4576 return (void *)addr;
4580 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4581 * @size: the number of bytes to allocate
4582 * @gfp_mask: GFP flags for the allocation
4584 * This function is similar to alloc_pages(), except that it allocates the
4585 * minimum number of pages to satisfy the request. alloc_pages() can only
4586 * allocate memory in power-of-two pages.
4588 * This function is also limited by MAX_ORDER.
4590 * Memory allocated by this function must be released by free_pages_exact().
4592 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4594 unsigned int order = get_order(size);
4597 addr = __get_free_pages(gfp_mask, order);
4598 return make_alloc_exact(addr, order, size);
4600 EXPORT_SYMBOL(alloc_pages_exact);
4603 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4605 * @nid: the preferred node ID where memory should be allocated
4606 * @size: the number of bytes to allocate
4607 * @gfp_mask: GFP flags for the allocation
4609 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4612 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4614 unsigned int order = get_order(size);
4615 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4618 return make_alloc_exact((unsigned long)page_address(p), order, size);
4622 * free_pages_exact - release memory allocated via alloc_pages_exact()
4623 * @virt: the value returned by alloc_pages_exact.
4624 * @size: size of allocation, same value as passed to alloc_pages_exact().
4626 * Release the memory allocated by a previous call to alloc_pages_exact.
4628 void free_pages_exact(void *virt, size_t size)
4630 unsigned long addr = (unsigned long)virt;
4631 unsigned long end = addr + PAGE_ALIGN(size);
4633 while (addr < end) {
4638 EXPORT_SYMBOL(free_pages_exact);
4641 * nr_free_zone_pages - count number of pages beyond high watermark
4642 * @offset: The zone index of the highest zone
4644 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4645 * high watermark within all zones at or below a given zone index. For each
4646 * zone, the number of pages is calculated as:
4648 * nr_free_zone_pages = managed_pages - high_pages
4650 static unsigned long nr_free_zone_pages(int offset)
4655 /* Just pick one node, since fallback list is circular */
4656 unsigned long sum = 0;
4658 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4660 for_each_zone_zonelist(zone, z, zonelist, offset) {
4661 unsigned long size = zone->managed_pages;
4662 unsigned long high = high_wmark_pages(zone);
4671 * nr_free_buffer_pages - count number of pages beyond high watermark
4673 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4674 * watermark within ZONE_DMA and ZONE_NORMAL.
4676 unsigned long nr_free_buffer_pages(void)
4678 return nr_free_zone_pages(gfp_zone(GFP_USER));
4680 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4683 * nr_free_pagecache_pages - count number of pages beyond high watermark
4685 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4686 * high watermark within all zones.
4688 unsigned long nr_free_pagecache_pages(void)
4690 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4693 static inline void show_node(struct zone *zone)
4695 if (IS_ENABLED(CONFIG_NUMA))
4696 printk("Node %d ", zone_to_nid(zone));
4699 long si_mem_available(void)
4702 unsigned long pagecache;
4703 unsigned long wmark_low = 0;
4704 unsigned long pages[NR_LRU_LISTS];
4708 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4709 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4712 wmark_low += zone->watermark[WMARK_LOW];
4715 * Estimate the amount of memory available for userspace allocations,
4716 * without causing swapping.
4718 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4721 * Not all the page cache can be freed, otherwise the system will
4722 * start swapping. Assume at least half of the page cache, or the
4723 * low watermark worth of cache, needs to stay.
4725 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4726 pagecache -= min(pagecache / 2, wmark_low);
4727 available += pagecache;
4730 * Part of the reclaimable slab consists of items that are in use,
4731 * and cannot be freed. Cap this estimate at the low watermark.
4733 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4734 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4738 * Part of the kernel memory, which can be released under memory
4741 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4748 EXPORT_SYMBOL_GPL(si_mem_available);
4750 void si_meminfo(struct sysinfo *val)
4752 val->totalram = totalram_pages;
4753 val->sharedram = global_node_page_state(NR_SHMEM);
4754 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4755 val->bufferram = nr_blockdev_pages();
4756 val->totalhigh = totalhigh_pages;
4757 val->freehigh = nr_free_highpages();
4758 val->mem_unit = PAGE_SIZE;
4761 EXPORT_SYMBOL(si_meminfo);
4764 void si_meminfo_node(struct sysinfo *val, int nid)
4766 int zone_type; /* needs to be signed */
4767 unsigned long managed_pages = 0;
4768 unsigned long managed_highpages = 0;
4769 unsigned long free_highpages = 0;
4770 pg_data_t *pgdat = NODE_DATA(nid);
4772 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4773 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4774 val->totalram = managed_pages;
4775 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4776 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4777 #ifdef CONFIG_HIGHMEM
4778 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4779 struct zone *zone = &pgdat->node_zones[zone_type];
4781 if (is_highmem(zone)) {
4782 managed_highpages += zone->managed_pages;
4783 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4786 val->totalhigh = managed_highpages;
4787 val->freehigh = free_highpages;
4789 val->totalhigh = managed_highpages;
4790 val->freehigh = free_highpages;
4792 val->mem_unit = PAGE_SIZE;
4797 * Determine whether the node should be displayed or not, depending on whether
4798 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4800 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4802 if (!(flags & SHOW_MEM_FILTER_NODES))
4806 * no node mask - aka implicit memory numa policy. Do not bother with
4807 * the synchronization - read_mems_allowed_begin - because we do not
4808 * have to be precise here.
4811 nodemask = &cpuset_current_mems_allowed;
4813 return !node_isset(nid, *nodemask);
4816 #define K(x) ((x) << (PAGE_SHIFT-10))
4818 static void show_migration_types(unsigned char type)
4820 static const char types[MIGRATE_TYPES] = {
4821 [MIGRATE_UNMOVABLE] = 'U',
4822 [MIGRATE_MOVABLE] = 'M',
4823 [MIGRATE_RECLAIMABLE] = 'E',
4824 [MIGRATE_HIGHATOMIC] = 'H',
4826 [MIGRATE_CMA] = 'C',
4828 #ifdef CONFIG_MEMORY_ISOLATION
4829 [MIGRATE_ISOLATE] = 'I',
4832 char tmp[MIGRATE_TYPES + 1];
4836 for (i = 0; i < MIGRATE_TYPES; i++) {
4837 if (type & (1 << i))
4842 printk(KERN_CONT "(%s) ", tmp);
4846 * Show free area list (used inside shift_scroll-lock stuff)
4847 * We also calculate the percentage fragmentation. We do this by counting the
4848 * memory on each free list with the exception of the first item on the list.
4851 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4854 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4856 unsigned long free_pcp = 0;
4861 for_each_populated_zone(zone) {
4862 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4865 for_each_online_cpu(cpu)
4866 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4869 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4870 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4871 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4872 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4873 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4874 " free:%lu free_pcp:%lu free_cma:%lu\n",
4875 global_node_page_state(NR_ACTIVE_ANON),
4876 global_node_page_state(NR_INACTIVE_ANON),
4877 global_node_page_state(NR_ISOLATED_ANON),
4878 global_node_page_state(NR_ACTIVE_FILE),
4879 global_node_page_state(NR_INACTIVE_FILE),
4880 global_node_page_state(NR_ISOLATED_FILE),
4881 global_node_page_state(NR_UNEVICTABLE),
4882 global_node_page_state(NR_FILE_DIRTY),
4883 global_node_page_state(NR_WRITEBACK),
4884 global_node_page_state(NR_UNSTABLE_NFS),
4885 global_node_page_state(NR_SLAB_RECLAIMABLE),
4886 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4887 global_node_page_state(NR_FILE_MAPPED),
4888 global_node_page_state(NR_SHMEM),
4889 global_zone_page_state(NR_PAGETABLE),
4890 global_zone_page_state(NR_BOUNCE),
4891 global_zone_page_state(NR_FREE_PAGES),
4893 global_zone_page_state(NR_FREE_CMA_PAGES));
4895 for_each_online_pgdat(pgdat) {
4896 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4900 " active_anon:%lukB"
4901 " inactive_anon:%lukB"
4902 " active_file:%lukB"
4903 " inactive_file:%lukB"
4904 " unevictable:%lukB"
4905 " isolated(anon):%lukB"
4906 " isolated(file):%lukB"
4911 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4913 " shmem_pmdmapped: %lukB"
4916 " writeback_tmp:%lukB"
4918 " all_unreclaimable? %s"
4921 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4922 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4923 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4924 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4925 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4926 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4927 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4928 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4929 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4930 K(node_page_state(pgdat, NR_WRITEBACK)),
4931 K(node_page_state(pgdat, NR_SHMEM)),
4932 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4933 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4934 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4936 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4938 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4939 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4940 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4944 for_each_populated_zone(zone) {
4947 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4951 for_each_online_cpu(cpu)
4952 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4961 " active_anon:%lukB"
4962 " inactive_anon:%lukB"
4963 " active_file:%lukB"
4964 " inactive_file:%lukB"
4965 " unevictable:%lukB"
4966 " writepending:%lukB"
4970 " kernel_stack:%lukB"
4978 K(zone_page_state(zone, NR_FREE_PAGES)),
4979 K(min_wmark_pages(zone)),
4980 K(low_wmark_pages(zone)),
4981 K(high_wmark_pages(zone)),
4982 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4983 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4984 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4985 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4986 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4987 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4988 K(zone->present_pages),
4989 K(zone->managed_pages),
4990 K(zone_page_state(zone, NR_MLOCK)),
4991 zone_page_state(zone, NR_KERNEL_STACK_KB),
4992 K(zone_page_state(zone, NR_PAGETABLE)),
4993 K(zone_page_state(zone, NR_BOUNCE)),
4995 K(this_cpu_read(zone->pageset->pcp.count)),
4996 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4997 printk("lowmem_reserve[]:");
4998 for (i = 0; i < MAX_NR_ZONES; i++)
4999 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5000 printk(KERN_CONT "\n");
5003 for_each_populated_zone(zone) {
5005 unsigned long nr[MAX_ORDER], flags, total = 0;
5006 unsigned char types[MAX_ORDER];
5008 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5011 printk(KERN_CONT "%s: ", zone->name);
5013 spin_lock_irqsave(&zone->lock, flags);
5014 for (order = 0; order < MAX_ORDER; order++) {
5015 struct free_area *area = &zone->free_area[order];
5018 nr[order] = area->nr_free;
5019 total += nr[order] << order;
5022 for (type = 0; type < MIGRATE_TYPES; type++) {
5023 if (!list_empty(&area->free_list[type]))
5024 types[order] |= 1 << type;
5027 spin_unlock_irqrestore(&zone->lock, flags);
5028 for (order = 0; order < MAX_ORDER; order++) {
5029 printk(KERN_CONT "%lu*%lukB ",
5030 nr[order], K(1UL) << order);
5032 show_migration_types(types[order]);
5034 printk(KERN_CONT "= %lukB\n", K(total));
5037 hugetlb_show_meminfo();
5039 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5041 show_swap_cache_info();
5044 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5046 zoneref->zone = zone;
5047 zoneref->zone_idx = zone_idx(zone);
5051 * Builds allocation fallback zone lists.
5053 * Add all populated zones of a node to the zonelist.
5055 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5058 enum zone_type zone_type = MAX_NR_ZONES;
5063 zone = pgdat->node_zones + zone_type;
5064 if (managed_zone(zone)) {
5065 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5066 check_highest_zone(zone_type);
5068 } while (zone_type);
5075 static int __parse_numa_zonelist_order(char *s)
5078 * We used to support different zonlists modes but they turned
5079 * out to be just not useful. Let's keep the warning in place
5080 * if somebody still use the cmd line parameter so that we do
5081 * not fail it silently
5083 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5084 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5090 static __init int setup_numa_zonelist_order(char *s)
5095 return __parse_numa_zonelist_order(s);
5097 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5099 char numa_zonelist_order[] = "Node";
5102 * sysctl handler for numa_zonelist_order
5104 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5105 void __user *buffer, size_t *length,
5112 return proc_dostring(table, write, buffer, length, ppos);
5113 str = memdup_user_nul(buffer, 16);
5115 return PTR_ERR(str);
5117 ret = __parse_numa_zonelist_order(str);
5123 #define MAX_NODE_LOAD (nr_online_nodes)
5124 static int node_load[MAX_NUMNODES];
5127 * find_next_best_node - find the next node that should appear in a given node's fallback list
5128 * @node: node whose fallback list we're appending
5129 * @used_node_mask: nodemask_t of already used nodes
5131 * We use a number of factors to determine which is the next node that should
5132 * appear on a given node's fallback list. The node should not have appeared
5133 * already in @node's fallback list, and it should be the next closest node
5134 * according to the distance array (which contains arbitrary distance values
5135 * from each node to each node in the system), and should also prefer nodes
5136 * with no CPUs, since presumably they'll have very little allocation pressure
5137 * on them otherwise.
5138 * It returns -1 if no node is found.
5140 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5143 int min_val = INT_MAX;
5144 int best_node = NUMA_NO_NODE;
5145 const struct cpumask *tmp = cpumask_of_node(0);
5147 /* Use the local node if we haven't already */
5148 if (!node_isset(node, *used_node_mask)) {
5149 node_set(node, *used_node_mask);
5153 for_each_node_state(n, N_MEMORY) {
5155 /* Don't want a node to appear more than once */
5156 if (node_isset(n, *used_node_mask))
5159 /* Use the distance array to find the distance */
5160 val = node_distance(node, n);
5162 /* Penalize nodes under us ("prefer the next node") */
5165 /* Give preference to headless and unused nodes */
5166 tmp = cpumask_of_node(n);
5167 if (!cpumask_empty(tmp))
5168 val += PENALTY_FOR_NODE_WITH_CPUS;
5170 /* Slight preference for less loaded node */
5171 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5172 val += node_load[n];
5174 if (val < min_val) {
5181 node_set(best_node, *used_node_mask);
5188 * Build zonelists ordered by node and zones within node.
5189 * This results in maximum locality--normal zone overflows into local
5190 * DMA zone, if any--but risks exhausting DMA zone.
5192 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5195 struct zoneref *zonerefs;
5198 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5200 for (i = 0; i < nr_nodes; i++) {
5203 pg_data_t *node = NODE_DATA(node_order[i]);
5205 nr_zones = build_zonerefs_node(node, zonerefs);
5206 zonerefs += nr_zones;
5208 zonerefs->zone = NULL;
5209 zonerefs->zone_idx = 0;
5213 * Build gfp_thisnode zonelists
5215 static void build_thisnode_zonelists(pg_data_t *pgdat)
5217 struct zoneref *zonerefs;
5220 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5221 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5222 zonerefs += nr_zones;
5223 zonerefs->zone = NULL;
5224 zonerefs->zone_idx = 0;
5228 * Build zonelists ordered by zone and nodes within zones.
5229 * This results in conserving DMA zone[s] until all Normal memory is
5230 * exhausted, but results in overflowing to remote node while memory
5231 * may still exist in local DMA zone.
5234 static void build_zonelists(pg_data_t *pgdat)
5236 static int node_order[MAX_NUMNODES];
5237 int node, load, nr_nodes = 0;
5238 nodemask_t used_mask;
5239 int local_node, prev_node;
5241 /* NUMA-aware ordering of nodes */
5242 local_node = pgdat->node_id;
5243 load = nr_online_nodes;
5244 prev_node = local_node;
5245 nodes_clear(used_mask);
5247 memset(node_order, 0, sizeof(node_order));
5248 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5250 * We don't want to pressure a particular node.
5251 * So adding penalty to the first node in same
5252 * distance group to make it round-robin.
5254 if (node_distance(local_node, node) !=
5255 node_distance(local_node, prev_node))
5256 node_load[node] = load;
5258 node_order[nr_nodes++] = node;
5263 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5264 build_thisnode_zonelists(pgdat);
5267 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5269 * Return node id of node used for "local" allocations.
5270 * I.e., first node id of first zone in arg node's generic zonelist.
5271 * Used for initializing percpu 'numa_mem', which is used primarily
5272 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5274 int local_memory_node(int node)
5278 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5279 gfp_zone(GFP_KERNEL),
5281 return zone_to_nid(z->zone);
5285 static void setup_min_unmapped_ratio(void);
5286 static void setup_min_slab_ratio(void);
5287 #else /* CONFIG_NUMA */
5289 static void build_zonelists(pg_data_t *pgdat)
5291 int node, local_node;
5292 struct zoneref *zonerefs;
5295 local_node = pgdat->node_id;
5297 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5298 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5299 zonerefs += nr_zones;
5302 * Now we build the zonelist so that it contains the zones
5303 * of all the other nodes.
5304 * We don't want to pressure a particular node, so when
5305 * building the zones for node N, we make sure that the
5306 * zones coming right after the local ones are those from
5307 * node N+1 (modulo N)
5309 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5310 if (!node_online(node))
5312 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5313 zonerefs += nr_zones;
5315 for (node = 0; node < local_node; node++) {
5316 if (!node_online(node))
5318 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5319 zonerefs += nr_zones;
5322 zonerefs->zone = NULL;
5323 zonerefs->zone_idx = 0;
5326 #endif /* CONFIG_NUMA */
5329 * Boot pageset table. One per cpu which is going to be used for all
5330 * zones and all nodes. The parameters will be set in such a way
5331 * that an item put on a list will immediately be handed over to
5332 * the buddy list. This is safe since pageset manipulation is done
5333 * with interrupts disabled.
5335 * The boot_pagesets must be kept even after bootup is complete for
5336 * unused processors and/or zones. They do play a role for bootstrapping
5337 * hotplugged processors.
5339 * zoneinfo_show() and maybe other functions do
5340 * not check if the processor is online before following the pageset pointer.
5341 * Other parts of the kernel may not check if the zone is available.
5343 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5344 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5345 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5347 static void __build_all_zonelists(void *data)
5350 int __maybe_unused cpu;
5351 pg_data_t *self = data;
5352 static DEFINE_SPINLOCK(lock);
5357 memset(node_load, 0, sizeof(node_load));
5361 * This node is hotadded and no memory is yet present. So just
5362 * building zonelists is fine - no need to touch other nodes.
5364 if (self && !node_online(self->node_id)) {
5365 build_zonelists(self);
5367 for_each_online_node(nid) {
5368 pg_data_t *pgdat = NODE_DATA(nid);
5370 build_zonelists(pgdat);
5373 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5375 * We now know the "local memory node" for each node--
5376 * i.e., the node of the first zone in the generic zonelist.
5377 * Set up numa_mem percpu variable for on-line cpus. During
5378 * boot, only the boot cpu should be on-line; we'll init the
5379 * secondary cpus' numa_mem as they come on-line. During
5380 * node/memory hotplug, we'll fixup all on-line cpus.
5382 for_each_online_cpu(cpu)
5383 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5390 static noinline void __init
5391 build_all_zonelists_init(void)
5395 __build_all_zonelists(NULL);
5398 * Initialize the boot_pagesets that are going to be used
5399 * for bootstrapping processors. The real pagesets for
5400 * each zone will be allocated later when the per cpu
5401 * allocator is available.
5403 * boot_pagesets are used also for bootstrapping offline
5404 * cpus if the system is already booted because the pagesets
5405 * are needed to initialize allocators on a specific cpu too.
5406 * F.e. the percpu allocator needs the page allocator which
5407 * needs the percpu allocator in order to allocate its pagesets
5408 * (a chicken-egg dilemma).
5410 for_each_possible_cpu(cpu)
5411 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5413 mminit_verify_zonelist();
5414 cpuset_init_current_mems_allowed();
5418 * unless system_state == SYSTEM_BOOTING.
5420 * __ref due to call of __init annotated helper build_all_zonelists_init
5421 * [protected by SYSTEM_BOOTING].
5423 void __ref build_all_zonelists(pg_data_t *pgdat)
5425 if (system_state == SYSTEM_BOOTING) {
5426 build_all_zonelists_init();
5428 __build_all_zonelists(pgdat);
5429 /* cpuset refresh routine should be here */
5431 vm_total_pages = nr_free_pagecache_pages();
5433 * Disable grouping by mobility if the number of pages in the
5434 * system is too low to allow the mechanism to work. It would be
5435 * more accurate, but expensive to check per-zone. This check is
5436 * made on memory-hotadd so a system can start with mobility
5437 * disabled and enable it later
5439 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5440 page_group_by_mobility_disabled = 1;
5442 page_group_by_mobility_disabled = 0;
5444 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5446 page_group_by_mobility_disabled ? "off" : "on",
5449 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5454 * Initially all pages are reserved - free ones are freed
5455 * up by free_all_bootmem() once the early boot process is
5456 * done. Non-atomic initialization, single-pass.
5458 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5459 unsigned long start_pfn, enum memmap_context context,
5460 struct vmem_altmap *altmap)
5462 unsigned long end_pfn = start_pfn + size;
5463 pg_data_t *pgdat = NODE_DATA(nid);
5465 unsigned long nr_initialised = 0;
5467 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5468 struct memblock_region *r = NULL, *tmp;
5471 if (highest_memmap_pfn < end_pfn - 1)
5472 highest_memmap_pfn = end_pfn - 1;
5475 * Honor reservation requested by the driver for this ZONE_DEVICE
5478 if (altmap && start_pfn == altmap->base_pfn)
5479 start_pfn += altmap->reserve;
5481 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5483 * There can be holes in boot-time mem_map[]s handed to this
5484 * function. They do not exist on hotplugged memory.
5486 if (context != MEMMAP_EARLY)
5489 if (!early_pfn_valid(pfn))
5491 if (!early_pfn_in_nid(pfn, nid))
5493 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5496 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5498 * Check given memblock attribute by firmware which can affect
5499 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5500 * mirrored, it's an overlapped memmap init. skip it.
5502 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5503 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5504 for_each_memblock(memory, tmp)
5505 if (pfn < memblock_region_memory_end_pfn(tmp))
5509 if (pfn >= memblock_region_memory_base_pfn(r) &&
5510 memblock_is_mirror(r)) {
5511 /* already initialized as NORMAL */
5512 pfn = memblock_region_memory_end_pfn(r);
5519 page = pfn_to_page(pfn);
5520 __init_single_page(page, pfn, zone, nid);
5521 if (context == MEMMAP_HOTPLUG)
5522 SetPageReserved(page);
5525 * Mark the block movable so that blocks are reserved for
5526 * movable at startup. This will force kernel allocations
5527 * to reserve their blocks rather than leaking throughout
5528 * the address space during boot when many long-lived
5529 * kernel allocations are made.
5531 * bitmap is created for zone's valid pfn range. but memmap
5532 * can be created for invalid pages (for alignment)
5533 * check here not to call set_pageblock_migratetype() against
5536 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5537 * because this is done early in sparse_add_one_section
5539 if (!(pfn & (pageblock_nr_pages - 1))) {
5540 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5546 static void __meminit zone_init_free_lists(struct zone *zone)
5548 unsigned int order, t;
5549 for_each_migratetype_order(order, t) {
5550 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5551 zone->free_area[order].nr_free = 0;
5555 #ifndef __HAVE_ARCH_MEMMAP_INIT
5556 #define memmap_init(size, nid, zone, start_pfn) \
5557 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5560 static int zone_batchsize(struct zone *zone)
5566 * The per-cpu-pages pools are set to around 1000th of the
5569 batch = zone->managed_pages / 1024;
5570 /* But no more than a meg. */
5571 if (batch * PAGE_SIZE > 1024 * 1024)
5572 batch = (1024 * 1024) / PAGE_SIZE;
5573 batch /= 4; /* We effectively *= 4 below */
5578 * Clamp the batch to a 2^n - 1 value. Having a power
5579 * of 2 value was found to be more likely to have
5580 * suboptimal cache aliasing properties in some cases.
5582 * For example if 2 tasks are alternately allocating
5583 * batches of pages, one task can end up with a lot
5584 * of pages of one half of the possible page colors
5585 * and the other with pages of the other colors.
5587 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5592 /* The deferral and batching of frees should be suppressed under NOMMU
5595 * The problem is that NOMMU needs to be able to allocate large chunks
5596 * of contiguous memory as there's no hardware page translation to
5597 * assemble apparent contiguous memory from discontiguous pages.
5599 * Queueing large contiguous runs of pages for batching, however,
5600 * causes the pages to actually be freed in smaller chunks. As there
5601 * can be a significant delay between the individual batches being
5602 * recycled, this leads to the once large chunks of space being
5603 * fragmented and becoming unavailable for high-order allocations.
5610 * pcp->high and pcp->batch values are related and dependent on one another:
5611 * ->batch must never be higher then ->high.
5612 * The following function updates them in a safe manner without read side
5615 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5616 * those fields changing asynchronously (acording the the above rule).
5618 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5619 * outside of boot time (or some other assurance that no concurrent updaters
5622 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5623 unsigned long batch)
5625 /* start with a fail safe value for batch */
5629 /* Update high, then batch, in order */
5636 /* a companion to pageset_set_high() */
5637 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5639 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5642 static void pageset_init(struct per_cpu_pageset *p)
5644 struct per_cpu_pages *pcp;
5647 memset(p, 0, sizeof(*p));
5651 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5652 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5655 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5658 pageset_set_batch(p, batch);
5662 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5663 * to the value high for the pageset p.
5665 static void pageset_set_high(struct per_cpu_pageset *p,
5668 unsigned long batch = max(1UL, high / 4);
5669 if ((high / 4) > (PAGE_SHIFT * 8))
5670 batch = PAGE_SHIFT * 8;
5672 pageset_update(&p->pcp, high, batch);
5675 static void pageset_set_high_and_batch(struct zone *zone,
5676 struct per_cpu_pageset *pcp)
5678 if (percpu_pagelist_fraction)
5679 pageset_set_high(pcp,
5680 (zone->managed_pages /
5681 percpu_pagelist_fraction));
5683 pageset_set_batch(pcp, zone_batchsize(zone));
5686 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5688 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5691 pageset_set_high_and_batch(zone, pcp);
5694 void __meminit setup_zone_pageset(struct zone *zone)
5697 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5698 for_each_possible_cpu(cpu)
5699 zone_pageset_init(zone, cpu);
5703 * Allocate per cpu pagesets and initialize them.
5704 * Before this call only boot pagesets were available.
5706 void __init setup_per_cpu_pageset(void)
5708 struct pglist_data *pgdat;
5711 for_each_populated_zone(zone)
5712 setup_zone_pageset(zone);
5714 for_each_online_pgdat(pgdat)
5715 pgdat->per_cpu_nodestats =
5716 alloc_percpu(struct per_cpu_nodestat);
5719 static __meminit void zone_pcp_init(struct zone *zone)
5722 * per cpu subsystem is not up at this point. The following code
5723 * relies on the ability of the linker to provide the
5724 * offset of a (static) per cpu variable into the per cpu area.
5726 zone->pageset = &boot_pageset;
5728 if (populated_zone(zone))
5729 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5730 zone->name, zone->present_pages,
5731 zone_batchsize(zone));
5734 void __meminit init_currently_empty_zone(struct zone *zone,
5735 unsigned long zone_start_pfn,
5738 struct pglist_data *pgdat = zone->zone_pgdat;
5740 pgdat->nr_zones = zone_idx(zone) + 1;
5742 zone->zone_start_pfn = zone_start_pfn;
5744 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5745 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5747 (unsigned long)zone_idx(zone),
5748 zone_start_pfn, (zone_start_pfn + size));
5750 zone_init_free_lists(zone);
5751 zone->initialized = 1;
5754 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5755 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5760 int __meminit __early_pfn_to_nid(unsigned long pfn,
5761 struct mminit_pfnnid_cache *state)
5763 unsigned long start_pfn, end_pfn;
5766 if (state->last_start <= pfn && pfn < state->last_end)
5767 return state->last_nid;
5769 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5771 state->last_start = start_pfn;
5772 state->last_end = end_pfn;
5773 state->last_nid = nid;
5778 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5781 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5782 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5783 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5785 * If an architecture guarantees that all ranges registered contain no holes
5786 * and may be freed, this this function may be used instead of calling
5787 * memblock_free_early_nid() manually.
5789 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5791 unsigned long start_pfn, end_pfn;
5794 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5795 start_pfn = min(start_pfn, max_low_pfn);
5796 end_pfn = min(end_pfn, max_low_pfn);
5798 if (start_pfn < end_pfn)
5799 memblock_free_early_nid(PFN_PHYS(start_pfn),
5800 (end_pfn - start_pfn) << PAGE_SHIFT,
5806 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5807 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5809 * If an architecture guarantees that all ranges registered contain no holes and may
5810 * be freed, this function may be used instead of calling memory_present() manually.
5812 void __init sparse_memory_present_with_active_regions(int nid)
5814 unsigned long start_pfn, end_pfn;
5817 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5818 memory_present(this_nid, start_pfn, end_pfn);
5822 * get_pfn_range_for_nid - Return the start and end page frames for a node
5823 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5824 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5825 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5827 * It returns the start and end page frame of a node based on information
5828 * provided by memblock_set_node(). If called for a node
5829 * with no available memory, a warning is printed and the start and end
5832 void __meminit get_pfn_range_for_nid(unsigned int nid,
5833 unsigned long *start_pfn, unsigned long *end_pfn)
5835 unsigned long this_start_pfn, this_end_pfn;
5841 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5842 *start_pfn = min(*start_pfn, this_start_pfn);
5843 *end_pfn = max(*end_pfn, this_end_pfn);
5846 if (*start_pfn == -1UL)
5851 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5852 * assumption is made that zones within a node are ordered in monotonic
5853 * increasing memory addresses so that the "highest" populated zone is used
5855 static void __init find_usable_zone_for_movable(void)
5858 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5859 if (zone_index == ZONE_MOVABLE)
5862 if (arch_zone_highest_possible_pfn[zone_index] >
5863 arch_zone_lowest_possible_pfn[zone_index])
5867 VM_BUG_ON(zone_index == -1);
5868 movable_zone = zone_index;
5872 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5873 * because it is sized independent of architecture. Unlike the other zones,
5874 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5875 * in each node depending on the size of each node and how evenly kernelcore
5876 * is distributed. This helper function adjusts the zone ranges
5877 * provided by the architecture for a given node by using the end of the
5878 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5879 * zones within a node are in order of monotonic increases memory addresses
5881 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5882 unsigned long zone_type,
5883 unsigned long node_start_pfn,
5884 unsigned long node_end_pfn,
5885 unsigned long *zone_start_pfn,
5886 unsigned long *zone_end_pfn)
5888 /* Only adjust if ZONE_MOVABLE is on this node */
5889 if (zone_movable_pfn[nid]) {
5890 /* Size ZONE_MOVABLE */
5891 if (zone_type == ZONE_MOVABLE) {
5892 *zone_start_pfn = zone_movable_pfn[nid];
5893 *zone_end_pfn = min(node_end_pfn,
5894 arch_zone_highest_possible_pfn[movable_zone]);
5896 /* Adjust for ZONE_MOVABLE starting within this range */
5897 } else if (!mirrored_kernelcore &&
5898 *zone_start_pfn < zone_movable_pfn[nid] &&
5899 *zone_end_pfn > zone_movable_pfn[nid]) {
5900 *zone_end_pfn = zone_movable_pfn[nid];
5902 /* Check if this whole range is within ZONE_MOVABLE */
5903 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5904 *zone_start_pfn = *zone_end_pfn;
5909 * Return the number of pages a zone spans in a node, including holes
5910 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5912 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5913 unsigned long zone_type,
5914 unsigned long node_start_pfn,
5915 unsigned long node_end_pfn,
5916 unsigned long *zone_start_pfn,
5917 unsigned long *zone_end_pfn,
5918 unsigned long *ignored)
5920 /* When hotadd a new node from cpu_up(), the node should be empty */
5921 if (!node_start_pfn && !node_end_pfn)
5924 /* Get the start and end of the zone */
5925 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5926 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5927 adjust_zone_range_for_zone_movable(nid, zone_type,
5928 node_start_pfn, node_end_pfn,
5929 zone_start_pfn, zone_end_pfn);
5931 /* Check that this node has pages within the zone's required range */
5932 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5935 /* Move the zone boundaries inside the node if necessary */
5936 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5937 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5939 /* Return the spanned pages */
5940 return *zone_end_pfn - *zone_start_pfn;
5944 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5945 * then all holes in the requested range will be accounted for.
5947 unsigned long __meminit __absent_pages_in_range(int nid,
5948 unsigned long range_start_pfn,
5949 unsigned long range_end_pfn)
5951 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5952 unsigned long start_pfn, end_pfn;
5955 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5956 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5957 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5958 nr_absent -= end_pfn - start_pfn;
5964 * absent_pages_in_range - Return number of page frames in holes within a range
5965 * @start_pfn: The start PFN to start searching for holes
5966 * @end_pfn: The end PFN to stop searching for holes
5968 * It returns the number of pages frames in memory holes within a range.
5970 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5971 unsigned long end_pfn)
5973 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5976 /* Return the number of page frames in holes in a zone on a node */
5977 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5978 unsigned long zone_type,
5979 unsigned long node_start_pfn,
5980 unsigned long node_end_pfn,
5981 unsigned long *ignored)
5983 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5984 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5985 unsigned long zone_start_pfn, zone_end_pfn;
5986 unsigned long nr_absent;
5988 /* When hotadd a new node from cpu_up(), the node should be empty */
5989 if (!node_start_pfn && !node_end_pfn)
5992 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5993 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5995 adjust_zone_range_for_zone_movable(nid, zone_type,
5996 node_start_pfn, node_end_pfn,
5997 &zone_start_pfn, &zone_end_pfn);
5998 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6001 * ZONE_MOVABLE handling.
6002 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6005 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6006 unsigned long start_pfn, end_pfn;
6007 struct memblock_region *r;
6009 for_each_memblock(memory, r) {
6010 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6011 zone_start_pfn, zone_end_pfn);
6012 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6013 zone_start_pfn, zone_end_pfn);
6015 if (zone_type == ZONE_MOVABLE &&
6016 memblock_is_mirror(r))
6017 nr_absent += end_pfn - start_pfn;
6019 if (zone_type == ZONE_NORMAL &&
6020 !memblock_is_mirror(r))
6021 nr_absent += end_pfn - start_pfn;
6028 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6029 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6030 unsigned long zone_type,
6031 unsigned long node_start_pfn,
6032 unsigned long node_end_pfn,
6033 unsigned long *zone_start_pfn,
6034 unsigned long *zone_end_pfn,
6035 unsigned long *zones_size)
6039 *zone_start_pfn = node_start_pfn;
6040 for (zone = 0; zone < zone_type; zone++)
6041 *zone_start_pfn += zones_size[zone];
6043 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6045 return zones_size[zone_type];
6048 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6049 unsigned long zone_type,
6050 unsigned long node_start_pfn,
6051 unsigned long node_end_pfn,
6052 unsigned long *zholes_size)
6057 return zholes_size[zone_type];
6060 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6062 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6063 unsigned long node_start_pfn,
6064 unsigned long node_end_pfn,
6065 unsigned long *zones_size,
6066 unsigned long *zholes_size)
6068 unsigned long realtotalpages = 0, totalpages = 0;
6071 for (i = 0; i < MAX_NR_ZONES; i++) {
6072 struct zone *zone = pgdat->node_zones + i;
6073 unsigned long zone_start_pfn, zone_end_pfn;
6074 unsigned long size, real_size;
6076 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6082 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6083 node_start_pfn, node_end_pfn,
6086 zone->zone_start_pfn = zone_start_pfn;
6088 zone->zone_start_pfn = 0;
6089 zone->spanned_pages = size;
6090 zone->present_pages = real_size;
6093 realtotalpages += real_size;
6096 pgdat->node_spanned_pages = totalpages;
6097 pgdat->node_present_pages = realtotalpages;
6098 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6102 #ifndef CONFIG_SPARSEMEM
6104 * Calculate the size of the zone->blockflags rounded to an unsigned long
6105 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6106 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6107 * round what is now in bits to nearest long in bits, then return it in
6110 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6112 unsigned long usemapsize;
6114 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6115 usemapsize = roundup(zonesize, pageblock_nr_pages);
6116 usemapsize = usemapsize >> pageblock_order;
6117 usemapsize *= NR_PAGEBLOCK_BITS;
6118 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6120 return usemapsize / 8;
6123 static void __ref setup_usemap(struct pglist_data *pgdat,
6125 unsigned long zone_start_pfn,
6126 unsigned long zonesize)
6128 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6129 zone->pageblock_flags = NULL;
6131 zone->pageblock_flags =
6132 memblock_virt_alloc_node_nopanic(usemapsize,
6136 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6137 unsigned long zone_start_pfn, unsigned long zonesize) {}
6138 #endif /* CONFIG_SPARSEMEM */
6140 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6142 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6143 void __init set_pageblock_order(void)
6147 /* Check that pageblock_nr_pages has not already been setup */
6148 if (pageblock_order)
6151 if (HPAGE_SHIFT > PAGE_SHIFT)
6152 order = HUGETLB_PAGE_ORDER;
6154 order = MAX_ORDER - 1;
6157 * Assume the largest contiguous order of interest is a huge page.
6158 * This value may be variable depending on boot parameters on IA64 and
6161 pageblock_order = order;
6163 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6166 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6167 * is unused as pageblock_order is set at compile-time. See
6168 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6171 void __init set_pageblock_order(void)
6175 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6177 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6178 unsigned long present_pages)
6180 unsigned long pages = spanned_pages;
6183 * Provide a more accurate estimation if there are holes within
6184 * the zone and SPARSEMEM is in use. If there are holes within the
6185 * zone, each populated memory region may cost us one or two extra
6186 * memmap pages due to alignment because memmap pages for each
6187 * populated regions may not be naturally aligned on page boundary.
6188 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6190 if (spanned_pages > present_pages + (present_pages >> 4) &&
6191 IS_ENABLED(CONFIG_SPARSEMEM))
6192 pages = present_pages;
6194 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6197 #ifdef CONFIG_NUMA_BALANCING
6198 static void pgdat_init_numabalancing(struct pglist_data *pgdat)
6200 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6201 pgdat->numabalancing_migrate_nr_pages = 0;
6202 pgdat->numabalancing_migrate_next_window = jiffies;
6205 static void pgdat_init_numabalancing(struct pglist_data *pgdat) {}
6208 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6209 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6211 spin_lock_init(&pgdat->split_queue_lock);
6212 INIT_LIST_HEAD(&pgdat->split_queue);
6213 pgdat->split_queue_len = 0;
6216 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6219 #ifdef CONFIG_COMPACTION
6220 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6222 init_waitqueue_head(&pgdat->kcompactd_wait);
6225 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6228 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6230 pgdat_resize_init(pgdat);
6232 pgdat_init_numabalancing(pgdat);
6233 pgdat_init_split_queue(pgdat);
6234 pgdat_init_kcompactd(pgdat);
6236 init_waitqueue_head(&pgdat->kswapd_wait);
6237 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6239 pgdat_page_ext_init(pgdat);
6240 spin_lock_init(&pgdat->lru_lock);
6241 lruvec_init(node_lruvec(pgdat));
6244 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6245 unsigned long remaining_pages)
6247 zone->managed_pages = remaining_pages;
6248 zone_set_nid(zone, nid);
6249 zone->name = zone_names[idx];
6250 zone->zone_pgdat = NODE_DATA(nid);
6251 spin_lock_init(&zone->lock);
6252 zone_seqlock_init(zone);
6253 zone_pcp_init(zone);
6257 * Set up the zone data structures
6258 * - init pgdat internals
6259 * - init all zones belonging to this node
6261 * NOTE: this function is only called during memory hotplug
6263 #ifdef CONFIG_MEMORY_HOTPLUG
6264 void __ref free_area_init_core_hotplug(int nid)
6267 pg_data_t *pgdat = NODE_DATA(nid);
6269 pgdat_init_internals(pgdat);
6270 for (z = 0; z < MAX_NR_ZONES; z++)
6271 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6276 * Set up the zone data structures:
6277 * - mark all pages reserved
6278 * - mark all memory queues empty
6279 * - clear the memory bitmaps
6281 * NOTE: pgdat should get zeroed by caller.
6282 * NOTE: this function is only called during early init.
6284 static void __init free_area_init_core(struct pglist_data *pgdat)
6287 int nid = pgdat->node_id;
6289 pgdat_init_internals(pgdat);
6290 pgdat->per_cpu_nodestats = &boot_nodestats;
6292 for (j = 0; j < MAX_NR_ZONES; j++) {
6293 struct zone *zone = pgdat->node_zones + j;
6294 unsigned long size, freesize, memmap_pages;
6295 unsigned long zone_start_pfn = zone->zone_start_pfn;
6297 size = zone->spanned_pages;
6298 freesize = zone->present_pages;
6301 * Adjust freesize so that it accounts for how much memory
6302 * is used by this zone for memmap. This affects the watermark
6303 * and per-cpu initialisations
6305 memmap_pages = calc_memmap_size(size, freesize);
6306 if (!is_highmem_idx(j)) {
6307 if (freesize >= memmap_pages) {
6308 freesize -= memmap_pages;
6311 " %s zone: %lu pages used for memmap\n",
6312 zone_names[j], memmap_pages);
6314 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6315 zone_names[j], memmap_pages, freesize);
6318 /* Account for reserved pages */
6319 if (j == 0 && freesize > dma_reserve) {
6320 freesize -= dma_reserve;
6321 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6322 zone_names[0], dma_reserve);
6325 if (!is_highmem_idx(j))
6326 nr_kernel_pages += freesize;
6327 /* Charge for highmem memmap if there are enough kernel pages */
6328 else if (nr_kernel_pages > memmap_pages * 2)
6329 nr_kernel_pages -= memmap_pages;
6330 nr_all_pages += freesize;
6333 * Set an approximate value for lowmem here, it will be adjusted
6334 * when the bootmem allocator frees pages into the buddy system.
6335 * And all highmem pages will be managed by the buddy system.
6337 zone_init_internals(zone, j, nid, freesize);
6342 set_pageblock_order();
6343 setup_usemap(pgdat, zone, zone_start_pfn, size);
6344 init_currently_empty_zone(zone, zone_start_pfn, size);
6345 memmap_init(size, nid, j, zone_start_pfn);
6349 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6350 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6352 unsigned long __maybe_unused start = 0;
6353 unsigned long __maybe_unused offset = 0;
6355 /* Skip empty nodes */
6356 if (!pgdat->node_spanned_pages)
6359 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6360 offset = pgdat->node_start_pfn - start;
6361 /* ia64 gets its own node_mem_map, before this, without bootmem */
6362 if (!pgdat->node_mem_map) {
6363 unsigned long size, end;
6367 * The zone's endpoints aren't required to be MAX_ORDER
6368 * aligned but the node_mem_map endpoints must be in order
6369 * for the buddy allocator to function correctly.
6371 end = pgdat_end_pfn(pgdat);
6372 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6373 size = (end - start) * sizeof(struct page);
6374 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6375 pgdat->node_mem_map = map + offset;
6377 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6378 __func__, pgdat->node_id, (unsigned long)pgdat,
6379 (unsigned long)pgdat->node_mem_map);
6380 #ifndef CONFIG_NEED_MULTIPLE_NODES
6382 * With no DISCONTIG, the global mem_map is just set as node 0's
6384 if (pgdat == NODE_DATA(0)) {
6385 mem_map = NODE_DATA(0)->node_mem_map;
6386 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6387 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6389 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6394 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6395 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6397 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6398 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6401 * We start only with one section of pages, more pages are added as
6402 * needed until the rest of deferred pages are initialized.
6404 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6405 pgdat->node_spanned_pages);
6406 pgdat->first_deferred_pfn = ULONG_MAX;
6409 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6412 void __init free_area_init_node(int nid, unsigned long *zones_size,
6413 unsigned long node_start_pfn,
6414 unsigned long *zholes_size)
6416 pg_data_t *pgdat = NODE_DATA(nid);
6417 unsigned long start_pfn = 0;
6418 unsigned long end_pfn = 0;
6420 /* pg_data_t should be reset to zero when it's allocated */
6421 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6423 pgdat->node_id = nid;
6424 pgdat->node_start_pfn = node_start_pfn;
6425 pgdat->per_cpu_nodestats = NULL;
6426 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6427 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6428 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6429 (u64)start_pfn << PAGE_SHIFT,
6430 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6432 start_pfn = node_start_pfn;
6434 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6435 zones_size, zholes_size);
6437 alloc_node_mem_map(pgdat);
6438 pgdat_set_deferred_range(pgdat);
6440 free_area_init_core(pgdat);
6443 #if defined(CONFIG_HAVE_MEMBLOCK) && !defined(CONFIG_FLAT_NODE_MEM_MAP)
6445 * Only struct pages that are backed by physical memory are zeroed and
6446 * initialized by going through __init_single_page(). But, there are some
6447 * struct pages which are reserved in memblock allocator and their fields
6448 * may be accessed (for example page_to_pfn() on some configuration accesses
6449 * flags). We must explicitly zero those struct pages.
6451 void __init zero_resv_unavail(void)
6453 phys_addr_t start, end;
6458 * Loop through ranges that are reserved, but do not have reported
6459 * physical memory backing.
6462 for_each_resv_unavail_range(i, &start, &end) {
6463 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6464 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6465 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6466 + pageblock_nr_pages - 1;
6469 mm_zero_struct_page(pfn_to_page(pfn));
6475 * Struct pages that do not have backing memory. This could be because
6476 * firmware is using some of this memory, or for some other reasons.
6477 * Once memblock is changed so such behaviour is not allowed: i.e.
6478 * list of "reserved" memory must be a subset of list of "memory", then
6479 * this code can be removed.
6482 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6484 #endif /* CONFIG_HAVE_MEMBLOCK && !CONFIG_FLAT_NODE_MEM_MAP */
6486 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6488 #if MAX_NUMNODES > 1
6490 * Figure out the number of possible node ids.
6492 void __init setup_nr_node_ids(void)
6494 unsigned int highest;
6496 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6497 nr_node_ids = highest + 1;
6502 * node_map_pfn_alignment - determine the maximum internode alignment
6504 * This function should be called after node map is populated and sorted.
6505 * It calculates the maximum power of two alignment which can distinguish
6508 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6509 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6510 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6511 * shifted, 1GiB is enough and this function will indicate so.
6513 * This is used to test whether pfn -> nid mapping of the chosen memory
6514 * model has fine enough granularity to avoid incorrect mapping for the
6515 * populated node map.
6517 * Returns the determined alignment in pfn's. 0 if there is no alignment
6518 * requirement (single node).
6520 unsigned long __init node_map_pfn_alignment(void)
6522 unsigned long accl_mask = 0, last_end = 0;
6523 unsigned long start, end, mask;
6527 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6528 if (!start || last_nid < 0 || last_nid == nid) {
6535 * Start with a mask granular enough to pin-point to the
6536 * start pfn and tick off bits one-by-one until it becomes
6537 * too coarse to separate the current node from the last.
6539 mask = ~((1 << __ffs(start)) - 1);
6540 while (mask && last_end <= (start & (mask << 1)))
6543 /* accumulate all internode masks */
6547 /* convert mask to number of pages */
6548 return ~accl_mask + 1;
6551 /* Find the lowest pfn for a node */
6552 static unsigned long __init find_min_pfn_for_node(int nid)
6554 unsigned long min_pfn = ULONG_MAX;
6555 unsigned long start_pfn;
6558 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6559 min_pfn = min(min_pfn, start_pfn);
6561 if (min_pfn == ULONG_MAX) {
6562 pr_warn("Could not find start_pfn for node %d\n", nid);
6570 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6572 * It returns the minimum PFN based on information provided via
6573 * memblock_set_node().
6575 unsigned long __init find_min_pfn_with_active_regions(void)
6577 return find_min_pfn_for_node(MAX_NUMNODES);
6581 * early_calculate_totalpages()
6582 * Sum pages in active regions for movable zone.
6583 * Populate N_MEMORY for calculating usable_nodes.
6585 static unsigned long __init early_calculate_totalpages(void)
6587 unsigned long totalpages = 0;
6588 unsigned long start_pfn, end_pfn;
6591 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6592 unsigned long pages = end_pfn - start_pfn;
6594 totalpages += pages;
6596 node_set_state(nid, N_MEMORY);
6602 * Find the PFN the Movable zone begins in each node. Kernel memory
6603 * is spread evenly between nodes as long as the nodes have enough
6604 * memory. When they don't, some nodes will have more kernelcore than
6607 static void __init find_zone_movable_pfns_for_nodes(void)
6610 unsigned long usable_startpfn;
6611 unsigned long kernelcore_node, kernelcore_remaining;
6612 /* save the state before borrow the nodemask */
6613 nodemask_t saved_node_state = node_states[N_MEMORY];
6614 unsigned long totalpages = early_calculate_totalpages();
6615 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6616 struct memblock_region *r;
6618 /* Need to find movable_zone earlier when movable_node is specified. */
6619 find_usable_zone_for_movable();
6622 * If movable_node is specified, ignore kernelcore and movablecore
6625 if (movable_node_is_enabled()) {
6626 for_each_memblock(memory, r) {
6627 if (!memblock_is_hotpluggable(r))
6632 usable_startpfn = PFN_DOWN(r->base);
6633 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6634 min(usable_startpfn, zone_movable_pfn[nid]) :
6642 * If kernelcore=mirror is specified, ignore movablecore option
6644 if (mirrored_kernelcore) {
6645 bool mem_below_4gb_not_mirrored = false;
6647 for_each_memblock(memory, r) {
6648 if (memblock_is_mirror(r))
6653 usable_startpfn = memblock_region_memory_base_pfn(r);
6655 if (usable_startpfn < 0x100000) {
6656 mem_below_4gb_not_mirrored = true;
6660 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6661 min(usable_startpfn, zone_movable_pfn[nid]) :
6665 if (mem_below_4gb_not_mirrored)
6666 pr_warn("This configuration results in unmirrored kernel memory.");
6672 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6673 * amount of necessary memory.
6675 if (required_kernelcore_percent)
6676 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6678 if (required_movablecore_percent)
6679 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6683 * If movablecore= was specified, calculate what size of
6684 * kernelcore that corresponds so that memory usable for
6685 * any allocation type is evenly spread. If both kernelcore
6686 * and movablecore are specified, then the value of kernelcore
6687 * will be used for required_kernelcore if it's greater than
6688 * what movablecore would have allowed.
6690 if (required_movablecore) {
6691 unsigned long corepages;
6694 * Round-up so that ZONE_MOVABLE is at least as large as what
6695 * was requested by the user
6697 required_movablecore =
6698 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6699 required_movablecore = min(totalpages, required_movablecore);
6700 corepages = totalpages - required_movablecore;
6702 required_kernelcore = max(required_kernelcore, corepages);
6706 * If kernelcore was not specified or kernelcore size is larger
6707 * than totalpages, there is no ZONE_MOVABLE.
6709 if (!required_kernelcore || required_kernelcore >= totalpages)
6712 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6713 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6716 /* Spread kernelcore memory as evenly as possible throughout nodes */
6717 kernelcore_node = required_kernelcore / usable_nodes;
6718 for_each_node_state(nid, N_MEMORY) {
6719 unsigned long start_pfn, end_pfn;
6722 * Recalculate kernelcore_node if the division per node
6723 * now exceeds what is necessary to satisfy the requested
6724 * amount of memory for the kernel
6726 if (required_kernelcore < kernelcore_node)
6727 kernelcore_node = required_kernelcore / usable_nodes;
6730 * As the map is walked, we track how much memory is usable
6731 * by the kernel using kernelcore_remaining. When it is
6732 * 0, the rest of the node is usable by ZONE_MOVABLE
6734 kernelcore_remaining = kernelcore_node;
6736 /* Go through each range of PFNs within this node */
6737 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6738 unsigned long size_pages;
6740 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6741 if (start_pfn >= end_pfn)
6744 /* Account for what is only usable for kernelcore */
6745 if (start_pfn < usable_startpfn) {
6746 unsigned long kernel_pages;
6747 kernel_pages = min(end_pfn, usable_startpfn)
6750 kernelcore_remaining -= min(kernel_pages,
6751 kernelcore_remaining);
6752 required_kernelcore -= min(kernel_pages,
6753 required_kernelcore);
6755 /* Continue if range is now fully accounted */
6756 if (end_pfn <= usable_startpfn) {
6759 * Push zone_movable_pfn to the end so
6760 * that if we have to rebalance
6761 * kernelcore across nodes, we will
6762 * not double account here
6764 zone_movable_pfn[nid] = end_pfn;
6767 start_pfn = usable_startpfn;
6771 * The usable PFN range for ZONE_MOVABLE is from
6772 * start_pfn->end_pfn. Calculate size_pages as the
6773 * number of pages used as kernelcore
6775 size_pages = end_pfn - start_pfn;
6776 if (size_pages > kernelcore_remaining)
6777 size_pages = kernelcore_remaining;
6778 zone_movable_pfn[nid] = start_pfn + size_pages;
6781 * Some kernelcore has been met, update counts and
6782 * break if the kernelcore for this node has been
6785 required_kernelcore -= min(required_kernelcore,
6787 kernelcore_remaining -= size_pages;
6788 if (!kernelcore_remaining)
6794 * If there is still required_kernelcore, we do another pass with one
6795 * less node in the count. This will push zone_movable_pfn[nid] further
6796 * along on the nodes that still have memory until kernelcore is
6800 if (usable_nodes && required_kernelcore > usable_nodes)
6804 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6805 for (nid = 0; nid < MAX_NUMNODES; nid++)
6806 zone_movable_pfn[nid] =
6807 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6810 /* restore the node_state */
6811 node_states[N_MEMORY] = saved_node_state;
6814 /* Any regular or high memory on that node ? */
6815 static void check_for_memory(pg_data_t *pgdat, int nid)
6817 enum zone_type zone_type;
6819 if (N_MEMORY == N_NORMAL_MEMORY)
6822 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6823 struct zone *zone = &pgdat->node_zones[zone_type];
6824 if (populated_zone(zone)) {
6825 node_set_state(nid, N_HIGH_MEMORY);
6826 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6827 zone_type <= ZONE_NORMAL)
6828 node_set_state(nid, N_NORMAL_MEMORY);
6835 * free_area_init_nodes - Initialise all pg_data_t and zone data
6836 * @max_zone_pfn: an array of max PFNs for each zone
6838 * This will call free_area_init_node() for each active node in the system.
6839 * Using the page ranges provided by memblock_set_node(), the size of each
6840 * zone in each node and their holes is calculated. If the maximum PFN
6841 * between two adjacent zones match, it is assumed that the zone is empty.
6842 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6843 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6844 * starts where the previous one ended. For example, ZONE_DMA32 starts
6845 * at arch_max_dma_pfn.
6847 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6849 unsigned long start_pfn, end_pfn;
6852 /* Record where the zone boundaries are */
6853 memset(arch_zone_lowest_possible_pfn, 0,
6854 sizeof(arch_zone_lowest_possible_pfn));
6855 memset(arch_zone_highest_possible_pfn, 0,
6856 sizeof(arch_zone_highest_possible_pfn));
6858 start_pfn = find_min_pfn_with_active_regions();
6860 for (i = 0; i < MAX_NR_ZONES; i++) {
6861 if (i == ZONE_MOVABLE)
6864 end_pfn = max(max_zone_pfn[i], start_pfn);
6865 arch_zone_lowest_possible_pfn[i] = start_pfn;
6866 arch_zone_highest_possible_pfn[i] = end_pfn;
6868 start_pfn = end_pfn;
6871 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6872 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6873 find_zone_movable_pfns_for_nodes();
6875 /* Print out the zone ranges */
6876 pr_info("Zone ranges:\n");
6877 for (i = 0; i < MAX_NR_ZONES; i++) {
6878 if (i == ZONE_MOVABLE)
6880 pr_info(" %-8s ", zone_names[i]);
6881 if (arch_zone_lowest_possible_pfn[i] ==
6882 arch_zone_highest_possible_pfn[i])
6885 pr_cont("[mem %#018Lx-%#018Lx]\n",
6886 (u64)arch_zone_lowest_possible_pfn[i]
6888 ((u64)arch_zone_highest_possible_pfn[i]
6889 << PAGE_SHIFT) - 1);
6892 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6893 pr_info("Movable zone start for each node\n");
6894 for (i = 0; i < MAX_NUMNODES; i++) {
6895 if (zone_movable_pfn[i])
6896 pr_info(" Node %d: %#018Lx\n", i,
6897 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6900 /* Print out the early node map */
6901 pr_info("Early memory node ranges\n");
6902 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6903 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6904 (u64)start_pfn << PAGE_SHIFT,
6905 ((u64)end_pfn << PAGE_SHIFT) - 1);
6907 /* Initialise every node */
6908 mminit_verify_pageflags_layout();
6909 setup_nr_node_ids();
6910 zero_resv_unavail();
6911 for_each_online_node(nid) {
6912 pg_data_t *pgdat = NODE_DATA(nid);
6913 free_area_init_node(nid, NULL,
6914 find_min_pfn_for_node(nid), NULL);
6916 /* Any memory on that node */
6917 if (pgdat->node_present_pages)
6918 node_set_state(nid, N_MEMORY);
6919 check_for_memory(pgdat, nid);
6923 static int __init cmdline_parse_core(char *p, unsigned long *core,
6924 unsigned long *percent)
6926 unsigned long long coremem;
6932 /* Value may be a percentage of total memory, otherwise bytes */
6933 coremem = simple_strtoull(p, &endptr, 0);
6934 if (*endptr == '%') {
6935 /* Paranoid check for percent values greater than 100 */
6936 WARN_ON(coremem > 100);
6940 coremem = memparse(p, &p);
6941 /* Paranoid check that UL is enough for the coremem value */
6942 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6944 *core = coremem >> PAGE_SHIFT;
6951 * kernelcore=size sets the amount of memory for use for allocations that
6952 * cannot be reclaimed or migrated.
6954 static int __init cmdline_parse_kernelcore(char *p)
6956 /* parse kernelcore=mirror */
6957 if (parse_option_str(p, "mirror")) {
6958 mirrored_kernelcore = true;
6962 return cmdline_parse_core(p, &required_kernelcore,
6963 &required_kernelcore_percent);
6967 * movablecore=size sets the amount of memory for use for allocations that
6968 * can be reclaimed or migrated.
6970 static int __init cmdline_parse_movablecore(char *p)
6972 return cmdline_parse_core(p, &required_movablecore,
6973 &required_movablecore_percent);
6976 early_param("kernelcore", cmdline_parse_kernelcore);
6977 early_param("movablecore", cmdline_parse_movablecore);
6979 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6981 void adjust_managed_page_count(struct page *page, long count)
6983 spin_lock(&managed_page_count_lock);
6984 page_zone(page)->managed_pages += count;
6985 totalram_pages += count;
6986 #ifdef CONFIG_HIGHMEM
6987 if (PageHighMem(page))
6988 totalhigh_pages += count;
6990 spin_unlock(&managed_page_count_lock);
6992 EXPORT_SYMBOL(adjust_managed_page_count);
6994 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6997 unsigned long pages = 0;
6999 start = (void *)PAGE_ALIGN((unsigned long)start);
7000 end = (void *)((unsigned long)end & PAGE_MASK);
7001 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7002 struct page *page = virt_to_page(pos);
7003 void *direct_map_addr;
7006 * 'direct_map_addr' might be different from 'pos'
7007 * because some architectures' virt_to_page()
7008 * work with aliases. Getting the direct map
7009 * address ensures that we get a _writeable_
7010 * alias for the memset().
7012 direct_map_addr = page_address(page);
7013 if ((unsigned int)poison <= 0xFF)
7014 memset(direct_map_addr, poison, PAGE_SIZE);
7016 free_reserved_page(page);
7020 pr_info("Freeing %s memory: %ldK\n",
7021 s, pages << (PAGE_SHIFT - 10));
7025 EXPORT_SYMBOL(free_reserved_area);
7027 #ifdef CONFIG_HIGHMEM
7028 void free_highmem_page(struct page *page)
7030 __free_reserved_page(page);
7032 page_zone(page)->managed_pages++;
7038 void __init mem_init_print_info(const char *str)
7040 unsigned long physpages, codesize, datasize, rosize, bss_size;
7041 unsigned long init_code_size, init_data_size;
7043 physpages = get_num_physpages();
7044 codesize = _etext - _stext;
7045 datasize = _edata - _sdata;
7046 rosize = __end_rodata - __start_rodata;
7047 bss_size = __bss_stop - __bss_start;
7048 init_data_size = __init_end - __init_begin;
7049 init_code_size = _einittext - _sinittext;
7052 * Detect special cases and adjust section sizes accordingly:
7053 * 1) .init.* may be embedded into .data sections
7054 * 2) .init.text.* may be out of [__init_begin, __init_end],
7055 * please refer to arch/tile/kernel/vmlinux.lds.S.
7056 * 3) .rodata.* may be embedded into .text or .data sections.
7058 #define adj_init_size(start, end, size, pos, adj) \
7060 if (start <= pos && pos < end && size > adj) \
7064 adj_init_size(__init_begin, __init_end, init_data_size,
7065 _sinittext, init_code_size);
7066 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7067 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7068 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7069 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7071 #undef adj_init_size
7073 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7074 #ifdef CONFIG_HIGHMEM
7078 nr_free_pages() << (PAGE_SHIFT - 10),
7079 physpages << (PAGE_SHIFT - 10),
7080 codesize >> 10, datasize >> 10, rosize >> 10,
7081 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7082 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7083 totalcma_pages << (PAGE_SHIFT - 10),
7084 #ifdef CONFIG_HIGHMEM
7085 totalhigh_pages << (PAGE_SHIFT - 10),
7087 str ? ", " : "", str ? str : "");
7091 * set_dma_reserve - set the specified number of pages reserved in the first zone
7092 * @new_dma_reserve: The number of pages to mark reserved
7094 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7095 * In the DMA zone, a significant percentage may be consumed by kernel image
7096 * and other unfreeable allocations which can skew the watermarks badly. This
7097 * function may optionally be used to account for unfreeable pages in the
7098 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7099 * smaller per-cpu batchsize.
7101 void __init set_dma_reserve(unsigned long new_dma_reserve)
7103 dma_reserve = new_dma_reserve;
7106 void __init free_area_init(unsigned long *zones_size)
7108 zero_resv_unavail();
7109 free_area_init_node(0, zones_size,
7110 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7113 static int page_alloc_cpu_dead(unsigned int cpu)
7116 lru_add_drain_cpu(cpu);
7120 * Spill the event counters of the dead processor
7121 * into the current processors event counters.
7122 * This artificially elevates the count of the current
7125 vm_events_fold_cpu(cpu);
7128 * Zero the differential counters of the dead processor
7129 * so that the vm statistics are consistent.
7131 * This is only okay since the processor is dead and cannot
7132 * race with what we are doing.
7134 cpu_vm_stats_fold(cpu);
7138 void __init page_alloc_init(void)
7142 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7143 "mm/page_alloc:dead", NULL,
7144 page_alloc_cpu_dead);
7149 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7150 * or min_free_kbytes changes.
7152 static void calculate_totalreserve_pages(void)
7154 struct pglist_data *pgdat;
7155 unsigned long reserve_pages = 0;
7156 enum zone_type i, j;
7158 for_each_online_pgdat(pgdat) {
7160 pgdat->totalreserve_pages = 0;
7162 for (i = 0; i < MAX_NR_ZONES; i++) {
7163 struct zone *zone = pgdat->node_zones + i;
7166 /* Find valid and maximum lowmem_reserve in the zone */
7167 for (j = i; j < MAX_NR_ZONES; j++) {
7168 if (zone->lowmem_reserve[j] > max)
7169 max = zone->lowmem_reserve[j];
7172 /* we treat the high watermark as reserved pages. */
7173 max += high_wmark_pages(zone);
7175 if (max > zone->managed_pages)
7176 max = zone->managed_pages;
7178 pgdat->totalreserve_pages += max;
7180 reserve_pages += max;
7183 totalreserve_pages = reserve_pages;
7187 * setup_per_zone_lowmem_reserve - called whenever
7188 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7189 * has a correct pages reserved value, so an adequate number of
7190 * pages are left in the zone after a successful __alloc_pages().
7192 static void setup_per_zone_lowmem_reserve(void)
7194 struct pglist_data *pgdat;
7195 enum zone_type j, idx;
7197 for_each_online_pgdat(pgdat) {
7198 for (j = 0; j < MAX_NR_ZONES; j++) {
7199 struct zone *zone = pgdat->node_zones + j;
7200 unsigned long managed_pages = zone->managed_pages;
7202 zone->lowmem_reserve[j] = 0;
7206 struct zone *lower_zone;
7209 lower_zone = pgdat->node_zones + idx;
7211 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7212 sysctl_lowmem_reserve_ratio[idx] = 0;
7213 lower_zone->lowmem_reserve[j] = 0;
7215 lower_zone->lowmem_reserve[j] =
7216 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7218 managed_pages += lower_zone->managed_pages;
7223 /* update totalreserve_pages */
7224 calculate_totalreserve_pages();
7227 static void __setup_per_zone_wmarks(void)
7229 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7230 unsigned long lowmem_pages = 0;
7232 unsigned long flags;
7234 /* Calculate total number of !ZONE_HIGHMEM pages */
7235 for_each_zone(zone) {
7236 if (!is_highmem(zone))
7237 lowmem_pages += zone->managed_pages;
7240 for_each_zone(zone) {
7243 spin_lock_irqsave(&zone->lock, flags);
7244 tmp = (u64)pages_min * zone->managed_pages;
7245 do_div(tmp, lowmem_pages);
7246 if (is_highmem(zone)) {
7248 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7249 * need highmem pages, so cap pages_min to a small
7252 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7253 * deltas control asynch page reclaim, and so should
7254 * not be capped for highmem.
7256 unsigned long min_pages;
7258 min_pages = zone->managed_pages / 1024;
7259 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7260 zone->watermark[WMARK_MIN] = min_pages;
7263 * If it's a lowmem zone, reserve a number of pages
7264 * proportionate to the zone's size.
7266 zone->watermark[WMARK_MIN] = tmp;
7270 * Set the kswapd watermarks distance according to the
7271 * scale factor in proportion to available memory, but
7272 * ensure a minimum size on small systems.
7274 tmp = max_t(u64, tmp >> 2,
7275 mult_frac(zone->managed_pages,
7276 watermark_scale_factor, 10000));
7278 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7279 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7281 spin_unlock_irqrestore(&zone->lock, flags);
7284 /* update totalreserve_pages */
7285 calculate_totalreserve_pages();
7289 * setup_per_zone_wmarks - called when min_free_kbytes changes
7290 * or when memory is hot-{added|removed}
7292 * Ensures that the watermark[min,low,high] values for each zone are set
7293 * correctly with respect to min_free_kbytes.
7295 void setup_per_zone_wmarks(void)
7297 static DEFINE_SPINLOCK(lock);
7300 __setup_per_zone_wmarks();
7305 * Initialise min_free_kbytes.
7307 * For small machines we want it small (128k min). For large machines
7308 * we want it large (64MB max). But it is not linear, because network
7309 * bandwidth does not increase linearly with machine size. We use
7311 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7312 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7328 int __meminit init_per_zone_wmark_min(void)
7330 unsigned long lowmem_kbytes;
7331 int new_min_free_kbytes;
7333 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7334 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7336 if (new_min_free_kbytes > user_min_free_kbytes) {
7337 min_free_kbytes = new_min_free_kbytes;
7338 if (min_free_kbytes < 128)
7339 min_free_kbytes = 128;
7340 if (min_free_kbytes > 65536)
7341 min_free_kbytes = 65536;
7343 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7344 new_min_free_kbytes, user_min_free_kbytes);
7346 setup_per_zone_wmarks();
7347 refresh_zone_stat_thresholds();
7348 setup_per_zone_lowmem_reserve();
7351 setup_min_unmapped_ratio();
7352 setup_min_slab_ratio();
7357 core_initcall(init_per_zone_wmark_min)
7360 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7361 * that we can call two helper functions whenever min_free_kbytes
7364 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7365 void __user *buffer, size_t *length, loff_t *ppos)
7369 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7374 user_min_free_kbytes = min_free_kbytes;
7375 setup_per_zone_wmarks();
7380 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7381 void __user *buffer, size_t *length, loff_t *ppos)
7385 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7390 setup_per_zone_wmarks();
7396 static void setup_min_unmapped_ratio(void)
7401 for_each_online_pgdat(pgdat)
7402 pgdat->min_unmapped_pages = 0;
7405 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7406 sysctl_min_unmapped_ratio) / 100;
7410 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7411 void __user *buffer, size_t *length, loff_t *ppos)
7415 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7419 setup_min_unmapped_ratio();
7424 static void setup_min_slab_ratio(void)
7429 for_each_online_pgdat(pgdat)
7430 pgdat->min_slab_pages = 0;
7433 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7434 sysctl_min_slab_ratio) / 100;
7437 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7438 void __user *buffer, size_t *length, loff_t *ppos)
7442 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7446 setup_min_slab_ratio();
7453 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7454 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7455 * whenever sysctl_lowmem_reserve_ratio changes.
7457 * The reserve ratio obviously has absolutely no relation with the
7458 * minimum watermarks. The lowmem reserve ratio can only make sense
7459 * if in function of the boot time zone sizes.
7461 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7462 void __user *buffer, size_t *length, loff_t *ppos)
7464 proc_dointvec_minmax(table, write, buffer, length, ppos);
7465 setup_per_zone_lowmem_reserve();
7470 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7471 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7472 * pagelist can have before it gets flushed back to buddy allocator.
7474 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7475 void __user *buffer, size_t *length, loff_t *ppos)
7478 int old_percpu_pagelist_fraction;
7481 mutex_lock(&pcp_batch_high_lock);
7482 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7484 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7485 if (!write || ret < 0)
7488 /* Sanity checking to avoid pcp imbalance */
7489 if (percpu_pagelist_fraction &&
7490 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7491 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7497 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7500 for_each_populated_zone(zone) {
7503 for_each_possible_cpu(cpu)
7504 pageset_set_high_and_batch(zone,
7505 per_cpu_ptr(zone->pageset, cpu));
7508 mutex_unlock(&pcp_batch_high_lock);
7513 int hashdist = HASHDIST_DEFAULT;
7515 static int __init set_hashdist(char *str)
7519 hashdist = simple_strtoul(str, &str, 0);
7522 __setup("hashdist=", set_hashdist);
7525 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7527 * Returns the number of pages that arch has reserved but
7528 * is not known to alloc_large_system_hash().
7530 static unsigned long __init arch_reserved_kernel_pages(void)
7537 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7538 * machines. As memory size is increased the scale is also increased but at
7539 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7540 * quadruples the scale is increased by one, which means the size of hash table
7541 * only doubles, instead of quadrupling as well.
7542 * Because 32-bit systems cannot have large physical memory, where this scaling
7543 * makes sense, it is disabled on such platforms.
7545 #if __BITS_PER_LONG > 32
7546 #define ADAPT_SCALE_BASE (64ul << 30)
7547 #define ADAPT_SCALE_SHIFT 2
7548 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7552 * allocate a large system hash table from bootmem
7553 * - it is assumed that the hash table must contain an exact power-of-2
7554 * quantity of entries
7555 * - limit is the number of hash buckets, not the total allocation size
7557 void *__init alloc_large_system_hash(const char *tablename,
7558 unsigned long bucketsize,
7559 unsigned long numentries,
7562 unsigned int *_hash_shift,
7563 unsigned int *_hash_mask,
7564 unsigned long low_limit,
7565 unsigned long high_limit)
7567 unsigned long long max = high_limit;
7568 unsigned long log2qty, size;
7572 /* allow the kernel cmdline to have a say */
7574 /* round applicable memory size up to nearest megabyte */
7575 numentries = nr_kernel_pages;
7576 numentries -= arch_reserved_kernel_pages();
7578 /* It isn't necessary when PAGE_SIZE >= 1MB */
7579 if (PAGE_SHIFT < 20)
7580 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7582 #if __BITS_PER_LONG > 32
7584 unsigned long adapt;
7586 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7587 adapt <<= ADAPT_SCALE_SHIFT)
7592 /* limit to 1 bucket per 2^scale bytes of low memory */
7593 if (scale > PAGE_SHIFT)
7594 numentries >>= (scale - PAGE_SHIFT);
7596 numentries <<= (PAGE_SHIFT - scale);
7598 /* Make sure we've got at least a 0-order allocation.. */
7599 if (unlikely(flags & HASH_SMALL)) {
7600 /* Makes no sense without HASH_EARLY */
7601 WARN_ON(!(flags & HASH_EARLY));
7602 if (!(numentries >> *_hash_shift)) {
7603 numentries = 1UL << *_hash_shift;
7604 BUG_ON(!numentries);
7606 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7607 numentries = PAGE_SIZE / bucketsize;
7609 numentries = roundup_pow_of_two(numentries);
7611 /* limit allocation size to 1/16 total memory by default */
7613 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7614 do_div(max, bucketsize);
7616 max = min(max, 0x80000000ULL);
7618 if (numentries < low_limit)
7619 numentries = low_limit;
7620 if (numentries > max)
7623 log2qty = ilog2(numentries);
7625 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7627 size = bucketsize << log2qty;
7628 if (flags & HASH_EARLY) {
7629 if (flags & HASH_ZERO)
7630 table = memblock_virt_alloc_nopanic(size, 0);
7632 table = memblock_virt_alloc_raw(size, 0);
7633 } else if (hashdist) {
7634 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7637 * If bucketsize is not a power-of-two, we may free
7638 * some pages at the end of hash table which
7639 * alloc_pages_exact() automatically does
7641 if (get_order(size) < MAX_ORDER) {
7642 table = alloc_pages_exact(size, gfp_flags);
7643 kmemleak_alloc(table, size, 1, gfp_flags);
7646 } while (!table && size > PAGE_SIZE && --log2qty);
7649 panic("Failed to allocate %s hash table\n", tablename);
7651 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7652 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7655 *_hash_shift = log2qty;
7657 *_hash_mask = (1 << log2qty) - 1;
7663 * This function checks whether pageblock includes unmovable pages or not.
7664 * If @count is not zero, it is okay to include less @count unmovable pages
7666 * PageLRU check without isolation or lru_lock could race so that
7667 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7668 * check without lock_page also may miss some movable non-lru pages at
7669 * race condition. So you can't expect this function should be exact.
7671 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7673 bool skip_hwpoisoned_pages)
7675 unsigned long pfn, iter, found;
7678 * TODO we could make this much more efficient by not checking every
7679 * page in the range if we know all of them are in MOVABLE_ZONE and
7680 * that the movable zone guarantees that pages are migratable but
7681 * the later is not the case right now unfortunatelly. E.g. movablecore
7682 * can still lead to having bootmem allocations in zone_movable.
7686 * CMA allocations (alloc_contig_range) really need to mark isolate
7687 * CMA pageblocks even when they are not movable in fact so consider
7688 * them movable here.
7690 if (is_migrate_cma(migratetype) &&
7691 is_migrate_cma(get_pageblock_migratetype(page)))
7694 pfn = page_to_pfn(page);
7695 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7696 unsigned long check = pfn + iter;
7698 if (!pfn_valid_within(check))
7701 page = pfn_to_page(check);
7703 if (PageReserved(page))
7707 * Hugepages are not in LRU lists, but they're movable.
7708 * We need not scan over tail pages bacause we don't
7709 * handle each tail page individually in migration.
7711 if (PageHuge(page)) {
7712 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7717 * We can't use page_count without pin a page
7718 * because another CPU can free compound page.
7719 * This check already skips compound tails of THP
7720 * because their page->_refcount is zero at all time.
7722 if (!page_ref_count(page)) {
7723 if (PageBuddy(page))
7724 iter += (1 << page_order(page)) - 1;
7729 * The HWPoisoned page may be not in buddy system, and
7730 * page_count() is not 0.
7732 if (skip_hwpoisoned_pages && PageHWPoison(page))
7735 if (__PageMovable(page))
7741 * If there are RECLAIMABLE pages, we need to check
7742 * it. But now, memory offline itself doesn't call
7743 * shrink_node_slabs() and it still to be fixed.
7746 * If the page is not RAM, page_count()should be 0.
7747 * we don't need more check. This is an _used_ not-movable page.
7749 * The problematic thing here is PG_reserved pages. PG_reserved
7750 * is set to both of a memory hole page and a _used_ kernel
7758 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7762 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7764 static unsigned long pfn_max_align_down(unsigned long pfn)
7766 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7767 pageblock_nr_pages) - 1);
7770 static unsigned long pfn_max_align_up(unsigned long pfn)
7772 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7773 pageblock_nr_pages));
7776 /* [start, end) must belong to a single zone. */
7777 static int __alloc_contig_migrate_range(struct compact_control *cc,
7778 unsigned long start, unsigned long end)
7780 /* This function is based on compact_zone() from compaction.c. */
7781 unsigned long nr_reclaimed;
7782 unsigned long pfn = start;
7783 unsigned int tries = 0;
7788 while (pfn < end || !list_empty(&cc->migratepages)) {
7789 if (fatal_signal_pending(current)) {
7794 if (list_empty(&cc->migratepages)) {
7795 cc->nr_migratepages = 0;
7796 pfn = isolate_migratepages_range(cc, pfn, end);
7802 } else if (++tries == 5) {
7803 ret = ret < 0 ? ret : -EBUSY;
7807 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7809 cc->nr_migratepages -= nr_reclaimed;
7811 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7812 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7815 putback_movable_pages(&cc->migratepages);
7822 * alloc_contig_range() -- tries to allocate given range of pages
7823 * @start: start PFN to allocate
7824 * @end: one-past-the-last PFN to allocate
7825 * @migratetype: migratetype of the underlaying pageblocks (either
7826 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7827 * in range must have the same migratetype and it must
7828 * be either of the two.
7829 * @gfp_mask: GFP mask to use during compaction
7831 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7832 * aligned. The PFN range must belong to a single zone.
7834 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7835 * pageblocks in the range. Once isolated, the pageblocks should not
7836 * be modified by others.
7838 * Returns zero on success or negative error code. On success all
7839 * pages which PFN is in [start, end) are allocated for the caller and
7840 * need to be freed with free_contig_range().
7842 int alloc_contig_range(unsigned long start, unsigned long end,
7843 unsigned migratetype, gfp_t gfp_mask)
7845 unsigned long outer_start, outer_end;
7849 struct compact_control cc = {
7850 .nr_migratepages = 0,
7852 .zone = page_zone(pfn_to_page(start)),
7853 .mode = MIGRATE_SYNC,
7854 .ignore_skip_hint = true,
7855 .no_set_skip_hint = true,
7856 .gfp_mask = current_gfp_context(gfp_mask),
7858 INIT_LIST_HEAD(&cc.migratepages);
7861 * What we do here is we mark all pageblocks in range as
7862 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7863 * have different sizes, and due to the way page allocator
7864 * work, we align the range to biggest of the two pages so
7865 * that page allocator won't try to merge buddies from
7866 * different pageblocks and change MIGRATE_ISOLATE to some
7867 * other migration type.
7869 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7870 * migrate the pages from an unaligned range (ie. pages that
7871 * we are interested in). This will put all the pages in
7872 * range back to page allocator as MIGRATE_ISOLATE.
7874 * When this is done, we take the pages in range from page
7875 * allocator removing them from the buddy system. This way
7876 * page allocator will never consider using them.
7878 * This lets us mark the pageblocks back as
7879 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7880 * aligned range but not in the unaligned, original range are
7881 * put back to page allocator so that buddy can use them.
7884 ret = start_isolate_page_range(pfn_max_align_down(start),
7885 pfn_max_align_up(end), migratetype,
7891 * In case of -EBUSY, we'd like to know which page causes problem.
7892 * So, just fall through. test_pages_isolated() has a tracepoint
7893 * which will report the busy page.
7895 * It is possible that busy pages could become available before
7896 * the call to test_pages_isolated, and the range will actually be
7897 * allocated. So, if we fall through be sure to clear ret so that
7898 * -EBUSY is not accidentally used or returned to caller.
7900 ret = __alloc_contig_migrate_range(&cc, start, end);
7901 if (ret && ret != -EBUSY)
7906 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7907 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7908 * more, all pages in [start, end) are free in page allocator.
7909 * What we are going to do is to allocate all pages from
7910 * [start, end) (that is remove them from page allocator).
7912 * The only problem is that pages at the beginning and at the
7913 * end of interesting range may be not aligned with pages that
7914 * page allocator holds, ie. they can be part of higher order
7915 * pages. Because of this, we reserve the bigger range and
7916 * once this is done free the pages we are not interested in.
7918 * We don't have to hold zone->lock here because the pages are
7919 * isolated thus they won't get removed from buddy.
7922 lru_add_drain_all();
7923 drain_all_pages(cc.zone);
7926 outer_start = start;
7927 while (!PageBuddy(pfn_to_page(outer_start))) {
7928 if (++order >= MAX_ORDER) {
7929 outer_start = start;
7932 outer_start &= ~0UL << order;
7935 if (outer_start != start) {
7936 order = page_order(pfn_to_page(outer_start));
7939 * outer_start page could be small order buddy page and
7940 * it doesn't include start page. Adjust outer_start
7941 * in this case to report failed page properly
7942 * on tracepoint in test_pages_isolated()
7944 if (outer_start + (1UL << order) <= start)
7945 outer_start = start;
7948 /* Make sure the range is really isolated. */
7949 if (test_pages_isolated(outer_start, end, false)) {
7950 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7951 __func__, outer_start, end);
7956 /* Grab isolated pages from freelists. */
7957 outer_end = isolate_freepages_range(&cc, outer_start, end);
7963 /* Free head and tail (if any) */
7964 if (start != outer_start)
7965 free_contig_range(outer_start, start - outer_start);
7966 if (end != outer_end)
7967 free_contig_range(end, outer_end - end);
7970 undo_isolate_page_range(pfn_max_align_down(start),
7971 pfn_max_align_up(end), migratetype);
7975 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7977 unsigned int count = 0;
7979 for (; nr_pages--; pfn++) {
7980 struct page *page = pfn_to_page(pfn);
7982 count += page_count(page) != 1;
7985 WARN(count != 0, "%d pages are still in use!\n", count);
7989 #ifdef CONFIG_MEMORY_HOTPLUG
7991 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7992 * page high values need to be recalulated.
7994 void __meminit zone_pcp_update(struct zone *zone)
7997 mutex_lock(&pcp_batch_high_lock);
7998 for_each_possible_cpu(cpu)
7999 pageset_set_high_and_batch(zone,
8000 per_cpu_ptr(zone->pageset, cpu));
8001 mutex_unlock(&pcp_batch_high_lock);
8005 void zone_pcp_reset(struct zone *zone)
8007 unsigned long flags;
8009 struct per_cpu_pageset *pset;
8011 /* avoid races with drain_pages() */
8012 local_irq_save(flags);
8013 if (zone->pageset != &boot_pageset) {
8014 for_each_online_cpu(cpu) {
8015 pset = per_cpu_ptr(zone->pageset, cpu);
8016 drain_zonestat(zone, pset);
8018 free_percpu(zone->pageset);
8019 zone->pageset = &boot_pageset;
8021 local_irq_restore(flags);
8024 #ifdef CONFIG_MEMORY_HOTREMOVE
8026 * All pages in the range must be in a single zone and isolated
8027 * before calling this.
8030 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8034 unsigned int order, i;
8036 unsigned long flags;
8037 /* find the first valid pfn */
8038 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8043 offline_mem_sections(pfn, end_pfn);
8044 zone = page_zone(pfn_to_page(pfn));
8045 spin_lock_irqsave(&zone->lock, flags);
8047 while (pfn < end_pfn) {
8048 if (!pfn_valid(pfn)) {
8052 page = pfn_to_page(pfn);
8054 * The HWPoisoned page may be not in buddy system, and
8055 * page_count() is not 0.
8057 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8059 SetPageReserved(page);
8063 BUG_ON(page_count(page));
8064 BUG_ON(!PageBuddy(page));
8065 order = page_order(page);
8066 #ifdef CONFIG_DEBUG_VM
8067 pr_info("remove from free list %lx %d %lx\n",
8068 pfn, 1 << order, end_pfn);
8070 list_del(&page->lru);
8071 rmv_page_order(page);
8072 zone->free_area[order].nr_free--;
8073 for (i = 0; i < (1 << order); i++)
8074 SetPageReserved((page+i));
8075 pfn += (1 << order);
8077 spin_unlock_irqrestore(&zone->lock, flags);
8081 bool is_free_buddy_page(struct page *page)
8083 struct zone *zone = page_zone(page);
8084 unsigned long pfn = page_to_pfn(page);
8085 unsigned long flags;
8088 spin_lock_irqsave(&zone->lock, flags);
8089 for (order = 0; order < MAX_ORDER; order++) {
8090 struct page *page_head = page - (pfn & ((1 << order) - 1));
8092 if (PageBuddy(page_head) && page_order(page_head) >= order)
8095 spin_unlock_irqrestore(&zone->lock, flags);
8097 return order < MAX_ORDER;
8100 #ifdef CONFIG_MEMORY_FAILURE
8102 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8103 * test is performed under the zone lock to prevent a race against page
8106 bool set_hwpoison_free_buddy_page(struct page *page)
8108 struct zone *zone = page_zone(page);
8109 unsigned long pfn = page_to_pfn(page);
8110 unsigned long flags;
8112 bool hwpoisoned = false;
8114 spin_lock_irqsave(&zone->lock, flags);
8115 for (order = 0; order < MAX_ORDER; order++) {
8116 struct page *page_head = page - (pfn & ((1 << order) - 1));
8118 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8119 if (!TestSetPageHWPoison(page))
8124 spin_unlock_irqrestore(&zone->lock, flags);