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/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node);
83 EXPORT_PER_CPU_SYMBOL(numa_node);
86 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
95 int _node_numa_mem_[MAX_NUMNODES];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex);
100 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy;
104 EXPORT_SYMBOL(latent_entropy);
108 * Array of node states.
110 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
111 [N_POSSIBLE] = NODE_MASK_ALL,
112 [N_ONLINE] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118 [N_MEMORY] = { { [0] = 1UL } },
119 [N_CPU] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock);
127 unsigned long totalram_pages __read_mostly;
128 unsigned long totalreserve_pages __read_mostly;
129 unsigned long totalcma_pages __read_mostly;
131 int percpu_pagelist_fraction;
132 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page *page)
147 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149 page->index = migratetype;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with pm_mutex held (gfp_allowed_mask also should
158 * only be modified with pm_mutex held, unless the suspend/hibernate code is
159 * guaranteed not to run in parallel with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&pm_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&pm_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
213 #ifdef CONFIG_HIGHMEM
219 EXPORT_SYMBOL(totalram_pages);
221 static char * const zone_names[MAX_NR_ZONES] = {
222 #ifdef CONFIG_ZONE_DMA
225 #ifdef CONFIG_ZONE_DMA32
229 #ifdef CONFIG_HIGHMEM
233 #ifdef CONFIG_ZONE_DEVICE
238 char * const migratetype_names[MIGRATE_TYPES] = {
246 #ifdef CONFIG_MEMORY_ISOLATION
251 compound_page_dtor * const compound_page_dtors[] = {
254 #ifdef CONFIG_HUGETLB_PAGE
257 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
262 int min_free_kbytes = 1024;
263 int user_min_free_kbytes = -1;
264 int watermark_scale_factor = 10;
266 static unsigned long __meminitdata nr_kernel_pages;
267 static unsigned long __meminitdata nr_all_pages;
268 static unsigned long __meminitdata dma_reserve;
270 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
271 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
272 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
273 static unsigned long __initdata required_kernelcore;
274 static unsigned long __initdata required_movablecore;
275 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
276 static bool mirrored_kernelcore;
278 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
280 EXPORT_SYMBOL(movable_zone);
281 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
284 int nr_node_ids __read_mostly = MAX_NUMNODES;
285 int nr_online_nodes __read_mostly = 1;
286 EXPORT_SYMBOL(nr_node_ids);
287 EXPORT_SYMBOL(nr_online_nodes);
290 int page_group_by_mobility_disabled __read_mostly;
292 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
293 static inline void reset_deferred_meminit(pg_data_t *pgdat)
295 unsigned long max_initialise;
296 unsigned long reserved_lowmem;
299 * Initialise at least 2G of a node but also take into account that
300 * two large system hashes that can take up 1GB for 0.25TB/node.
302 max_initialise = max(2UL << (30 - PAGE_SHIFT),
303 (pgdat->node_spanned_pages >> 8));
306 * Compensate the all the memblock reservations (e.g. crash kernel)
307 * from the initial estimation to make sure we will initialize enough
310 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
311 pgdat->node_start_pfn + max_initialise);
312 max_initialise += reserved_lowmem;
314 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
315 pgdat->first_deferred_pfn = ULONG_MAX;
318 /* Returns true if the struct page for the pfn is uninitialised */
319 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
321 int nid = early_pfn_to_nid(pfn);
323 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
330 * Returns false when the remaining initialisation should be deferred until
331 * later in the boot cycle when it can be parallelised.
333 static inline bool update_defer_init(pg_data_t *pgdat,
334 unsigned long pfn, unsigned long zone_end,
335 unsigned long *nr_initialised)
337 /* Always populate low zones for address-contrained allocations */
338 if (zone_end < pgdat_end_pfn(pgdat))
341 if ((*nr_initialised > pgdat->static_init_size) &&
342 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
343 pgdat->first_deferred_pfn = pfn;
350 static inline void reset_deferred_meminit(pg_data_t *pgdat)
354 static inline bool early_page_uninitialised(unsigned long pfn)
359 static inline bool update_defer_init(pg_data_t *pgdat,
360 unsigned long pfn, unsigned long zone_end,
361 unsigned long *nr_initialised)
367 /* Return a pointer to the bitmap storing bits affecting a block of pages */
368 static inline unsigned long *get_pageblock_bitmap(struct page *page,
371 #ifdef CONFIG_SPARSEMEM
372 return __pfn_to_section(pfn)->pageblock_flags;
374 return page_zone(page)->pageblock_flags;
375 #endif /* CONFIG_SPARSEMEM */
378 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
380 #ifdef CONFIG_SPARSEMEM
381 pfn &= (PAGES_PER_SECTION-1);
382 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
384 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
385 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
386 #endif /* CONFIG_SPARSEMEM */
390 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
391 * @page: The page within the block of interest
392 * @pfn: The target page frame number
393 * @end_bitidx: The last bit of interest to retrieve
394 * @mask: mask of bits that the caller is interested in
396 * Return: pageblock_bits flags
398 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
400 unsigned long end_bitidx,
403 unsigned long *bitmap;
404 unsigned long bitidx, word_bitidx;
407 bitmap = get_pageblock_bitmap(page, pfn);
408 bitidx = pfn_to_bitidx(page, pfn);
409 word_bitidx = bitidx / BITS_PER_LONG;
410 bitidx &= (BITS_PER_LONG-1);
412 word = bitmap[word_bitidx];
413 bitidx += end_bitidx;
414 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
417 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
418 unsigned long end_bitidx,
421 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
424 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
426 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
430 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
431 * @page: The page within the block of interest
432 * @flags: The flags to set
433 * @pfn: The target page frame number
434 * @end_bitidx: The last bit of interest
435 * @mask: mask of bits that the caller is interested in
437 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
439 unsigned long end_bitidx,
442 unsigned long *bitmap;
443 unsigned long bitidx, word_bitidx;
444 unsigned long old_word, word;
446 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
448 bitmap = get_pageblock_bitmap(page, pfn);
449 bitidx = pfn_to_bitidx(page, pfn);
450 word_bitidx = bitidx / BITS_PER_LONG;
451 bitidx &= (BITS_PER_LONG-1);
453 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
455 bitidx += end_bitidx;
456 mask <<= (BITS_PER_LONG - bitidx - 1);
457 flags <<= (BITS_PER_LONG - bitidx - 1);
459 word = READ_ONCE(bitmap[word_bitidx]);
461 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
462 if (word == old_word)
468 void set_pageblock_migratetype(struct page *page, int migratetype)
470 if (unlikely(page_group_by_mobility_disabled &&
471 migratetype < MIGRATE_PCPTYPES))
472 migratetype = MIGRATE_UNMOVABLE;
474 set_pageblock_flags_group(page, (unsigned long)migratetype,
475 PB_migrate, PB_migrate_end);
478 #ifdef CONFIG_DEBUG_VM
479 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
483 unsigned long pfn = page_to_pfn(page);
484 unsigned long sp, start_pfn;
487 seq = zone_span_seqbegin(zone);
488 start_pfn = zone->zone_start_pfn;
489 sp = zone->spanned_pages;
490 if (!zone_spans_pfn(zone, pfn))
492 } while (zone_span_seqretry(zone, seq));
495 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
496 pfn, zone_to_nid(zone), zone->name,
497 start_pfn, start_pfn + sp);
502 static int page_is_consistent(struct zone *zone, struct page *page)
504 if (!pfn_valid_within(page_to_pfn(page)))
506 if (zone != page_zone(page))
512 * Temporary debugging check for pages not lying within a given zone.
514 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
516 if (page_outside_zone_boundaries(zone, page))
518 if (!page_is_consistent(zone, page))
524 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
530 static void bad_page(struct page *page, const char *reason,
531 unsigned long bad_flags)
533 static unsigned long resume;
534 static unsigned long nr_shown;
535 static unsigned long nr_unshown;
538 * Allow a burst of 60 reports, then keep quiet for that minute;
539 * or allow a steady drip of one report per second.
541 if (nr_shown == 60) {
542 if (time_before(jiffies, resume)) {
548 "BUG: Bad page state: %lu messages suppressed\n",
555 resume = jiffies + 60 * HZ;
557 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
558 current->comm, page_to_pfn(page));
559 __dump_page(page, reason);
560 bad_flags &= page->flags;
562 pr_alert("bad because of flags: %#lx(%pGp)\n",
563 bad_flags, &bad_flags);
564 dump_page_owner(page);
569 /* Leave bad fields for debug, except PageBuddy could make trouble */
570 page_mapcount_reset(page); /* remove PageBuddy */
571 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
575 * Higher-order pages are called "compound pages". They are structured thusly:
577 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
579 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
580 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
582 * The first tail page's ->compound_dtor holds the offset in array of compound
583 * page destructors. See compound_page_dtors.
585 * The first tail page's ->compound_order holds the order of allocation.
586 * This usage means that zero-order pages may not be compound.
589 void free_compound_page(struct page *page)
591 __free_pages_ok(page, compound_order(page));
594 void prep_compound_page(struct page *page, unsigned int order)
597 int nr_pages = 1 << order;
599 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
600 set_compound_order(page, order);
602 for (i = 1; i < nr_pages; i++) {
603 struct page *p = page + i;
604 set_page_count(p, 0);
605 p->mapping = TAIL_MAPPING;
606 set_compound_head(p, page);
608 atomic_set(compound_mapcount_ptr(page), -1);
611 #ifdef CONFIG_DEBUG_PAGEALLOC
612 unsigned int _debug_guardpage_minorder;
613 bool _debug_pagealloc_enabled __read_mostly
614 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
615 EXPORT_SYMBOL(_debug_pagealloc_enabled);
616 bool _debug_guardpage_enabled __read_mostly;
618 static int __init early_debug_pagealloc(char *buf)
622 return kstrtobool(buf, &_debug_pagealloc_enabled);
624 early_param("debug_pagealloc", early_debug_pagealloc);
626 static bool need_debug_guardpage(void)
628 /* If we don't use debug_pagealloc, we don't need guard page */
629 if (!debug_pagealloc_enabled())
632 if (!debug_guardpage_minorder())
638 static void init_debug_guardpage(void)
640 if (!debug_pagealloc_enabled())
643 if (!debug_guardpage_minorder())
646 _debug_guardpage_enabled = true;
649 struct page_ext_operations debug_guardpage_ops = {
650 .need = need_debug_guardpage,
651 .init = init_debug_guardpage,
654 static int __init debug_guardpage_minorder_setup(char *buf)
658 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
659 pr_err("Bad debug_guardpage_minorder value\n");
662 _debug_guardpage_minorder = res;
663 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
666 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
668 static inline bool set_page_guard(struct zone *zone, struct page *page,
669 unsigned int order, int migratetype)
671 struct page_ext *page_ext;
673 if (!debug_guardpage_enabled())
676 if (order >= debug_guardpage_minorder())
679 page_ext = lookup_page_ext(page);
680 if (unlikely(!page_ext))
683 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
685 INIT_LIST_HEAD(&page->lru);
686 set_page_private(page, order);
687 /* Guard pages are not available for any usage */
688 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
693 static inline void clear_page_guard(struct zone *zone, struct page *page,
694 unsigned int order, int migratetype)
696 struct page_ext *page_ext;
698 if (!debug_guardpage_enabled())
701 page_ext = lookup_page_ext(page);
702 if (unlikely(!page_ext))
705 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
707 set_page_private(page, 0);
708 if (!is_migrate_isolate(migratetype))
709 __mod_zone_freepage_state(zone, (1 << order), migratetype);
712 struct page_ext_operations debug_guardpage_ops;
713 static inline bool set_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype) { return false; }
715 static inline void clear_page_guard(struct zone *zone, struct page *page,
716 unsigned int order, int migratetype) {}
719 static inline void set_page_order(struct page *page, unsigned int order)
721 set_page_private(page, order);
722 __SetPageBuddy(page);
725 static inline void rmv_page_order(struct page *page)
727 __ClearPageBuddy(page);
728 set_page_private(page, 0);
732 * This function checks whether a page is free && is the buddy
733 * we can do coalesce a page and its buddy if
734 * (a) the buddy is not in a hole (check before calling!) &&
735 * (b) the buddy is in the buddy system &&
736 * (c) a page and its buddy have the same order &&
737 * (d) a page and its buddy are in the same zone.
739 * For recording whether a page is in the buddy system, we set ->_mapcount
740 * PAGE_BUDDY_MAPCOUNT_VALUE.
741 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
742 * serialized by zone->lock.
744 * For recording page's order, we use page_private(page).
746 static inline int page_is_buddy(struct page *page, struct page *buddy,
749 if (page_is_guard(buddy) && page_order(buddy) == order) {
750 if (page_zone_id(page) != page_zone_id(buddy))
753 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
758 if (PageBuddy(buddy) && page_order(buddy) == order) {
760 * zone check is done late to avoid uselessly
761 * calculating zone/node ids for pages that could
764 if (page_zone_id(page) != page_zone_id(buddy))
767 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
775 * Freeing function for a buddy system allocator.
777 * The concept of a buddy system is to maintain direct-mapped table
778 * (containing bit values) for memory blocks of various "orders".
779 * The bottom level table contains the map for the smallest allocatable
780 * units of memory (here, pages), and each level above it describes
781 * pairs of units from the levels below, hence, "buddies".
782 * At a high level, all that happens here is marking the table entry
783 * at the bottom level available, and propagating the changes upward
784 * as necessary, plus some accounting needed to play nicely with other
785 * parts of the VM system.
786 * At each level, we keep a list of pages, which are heads of continuous
787 * free pages of length of (1 << order) and marked with _mapcount
788 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
790 * So when we are allocating or freeing one, we can derive the state of the
791 * other. That is, if we allocate a small block, and both were
792 * free, the remainder of the region must be split into blocks.
793 * If a block is freed, and its buddy is also free, then this
794 * triggers coalescing into a block of larger size.
799 static inline void __free_one_page(struct page *page,
801 struct zone *zone, unsigned int order,
804 unsigned long combined_pfn;
805 unsigned long uninitialized_var(buddy_pfn);
807 unsigned int max_order;
809 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
811 VM_BUG_ON(!zone_is_initialized(zone));
812 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
814 VM_BUG_ON(migratetype == -1);
815 if (likely(!is_migrate_isolate(migratetype)))
816 __mod_zone_freepage_state(zone, 1 << order, migratetype);
818 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
819 VM_BUG_ON_PAGE(bad_range(zone, page), page);
822 while (order < max_order - 1) {
823 buddy_pfn = __find_buddy_pfn(pfn, order);
824 buddy = page + (buddy_pfn - pfn);
826 if (!pfn_valid_within(buddy_pfn))
828 if (!page_is_buddy(page, buddy, order))
831 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
832 * merge with it and move up one order.
834 if (page_is_guard(buddy)) {
835 clear_page_guard(zone, buddy, order, migratetype);
837 list_del(&buddy->lru);
838 zone->free_area[order].nr_free--;
839 rmv_page_order(buddy);
841 combined_pfn = buddy_pfn & pfn;
842 page = page + (combined_pfn - pfn);
846 if (max_order < MAX_ORDER) {
847 /* If we are here, it means order is >= pageblock_order.
848 * We want to prevent merge between freepages on isolate
849 * pageblock and normal pageblock. Without this, pageblock
850 * isolation could cause incorrect freepage or CMA accounting.
852 * We don't want to hit this code for the more frequent
855 if (unlikely(has_isolate_pageblock(zone))) {
858 buddy_pfn = __find_buddy_pfn(pfn, order);
859 buddy = page + (buddy_pfn - pfn);
860 buddy_mt = get_pageblock_migratetype(buddy);
862 if (migratetype != buddy_mt
863 && (is_migrate_isolate(migratetype) ||
864 is_migrate_isolate(buddy_mt)))
868 goto continue_merging;
872 set_page_order(page, order);
875 * If this is not the largest possible page, check if the buddy
876 * of the next-highest order is free. If it is, it's possible
877 * that pages are being freed that will coalesce soon. In case,
878 * that is happening, add the free page to the tail of the list
879 * so it's less likely to be used soon and more likely to be merged
880 * as a higher order page
882 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
883 struct page *higher_page, *higher_buddy;
884 combined_pfn = buddy_pfn & pfn;
885 higher_page = page + (combined_pfn - pfn);
886 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
887 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
888 if (pfn_valid_within(buddy_pfn) &&
889 page_is_buddy(higher_page, higher_buddy, order + 1)) {
890 list_add_tail(&page->lru,
891 &zone->free_area[order].free_list[migratetype]);
896 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
898 zone->free_area[order].nr_free++;
902 * A bad page could be due to a number of fields. Instead of multiple branches,
903 * try and check multiple fields with one check. The caller must do a detailed
904 * check if necessary.
906 static inline bool page_expected_state(struct page *page,
907 unsigned long check_flags)
909 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 if (unlikely((unsigned long)page->mapping |
913 page_ref_count(page) |
915 (unsigned long)page->mem_cgroup |
917 (page->flags & check_flags)))
923 static void free_pages_check_bad(struct page *page)
925 const char *bad_reason;
926 unsigned long bad_flags;
931 if (unlikely(atomic_read(&page->_mapcount) != -1))
932 bad_reason = "nonzero mapcount";
933 if (unlikely(page->mapping != NULL))
934 bad_reason = "non-NULL mapping";
935 if (unlikely(page_ref_count(page) != 0))
936 bad_reason = "nonzero _refcount";
937 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
938 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
939 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
942 if (unlikely(page->mem_cgroup))
943 bad_reason = "page still charged to cgroup";
945 bad_page(page, bad_reason, bad_flags);
948 static inline int free_pages_check(struct page *page)
950 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
953 /* Something has gone sideways, find it */
954 free_pages_check_bad(page);
958 static int free_tail_pages_check(struct page *head_page, struct page *page)
963 * We rely page->lru.next never has bit 0 set, unless the page
964 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
966 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
968 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
972 switch (page - head_page) {
974 /* the first tail page: ->mapping is compound_mapcount() */
975 if (unlikely(compound_mapcount(page))) {
976 bad_page(page, "nonzero compound_mapcount", 0);
982 * the second tail page: ->mapping is
983 * page_deferred_list().next -- ignore value.
987 if (page->mapping != TAIL_MAPPING) {
988 bad_page(page, "corrupted mapping in tail page", 0);
993 if (unlikely(!PageTail(page))) {
994 bad_page(page, "PageTail not set", 0);
997 if (unlikely(compound_head(page) != head_page)) {
998 bad_page(page, "compound_head not consistent", 0);
1003 page->mapping = NULL;
1004 clear_compound_head(page);
1008 static __always_inline bool free_pages_prepare(struct page *page,
1009 unsigned int order, bool check_free)
1013 VM_BUG_ON_PAGE(PageTail(page), page);
1015 trace_mm_page_free(page, order);
1016 kmemcheck_free_shadow(page, order);
1019 * Check tail pages before head page information is cleared to
1020 * avoid checking PageCompound for order-0 pages.
1022 if (unlikely(order)) {
1023 bool compound = PageCompound(page);
1026 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1029 ClearPageDoubleMap(page);
1030 for (i = 1; i < (1 << order); i++) {
1032 bad += free_tail_pages_check(page, page + i);
1033 if (unlikely(free_pages_check(page + i))) {
1037 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1040 if (PageMappingFlags(page))
1041 page->mapping = NULL;
1042 if (memcg_kmem_enabled() && PageKmemcg(page))
1043 memcg_kmem_uncharge(page, order);
1045 bad += free_pages_check(page);
1049 page_cpupid_reset_last(page);
1050 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1051 reset_page_owner(page, order);
1053 if (!PageHighMem(page)) {
1054 debug_check_no_locks_freed(page_address(page),
1055 PAGE_SIZE << order);
1056 debug_check_no_obj_freed(page_address(page),
1057 PAGE_SIZE << order);
1059 arch_free_page(page, order);
1060 kernel_poison_pages(page, 1 << order, 0);
1061 kernel_map_pages(page, 1 << order, 0);
1062 kasan_free_pages(page, order);
1067 #ifdef CONFIG_DEBUG_VM
1068 static inline bool free_pcp_prepare(struct page *page)
1070 return free_pages_prepare(page, 0, true);
1073 static inline bool bulkfree_pcp_prepare(struct page *page)
1078 static bool free_pcp_prepare(struct page *page)
1080 return free_pages_prepare(page, 0, false);
1083 static bool bulkfree_pcp_prepare(struct page *page)
1085 return free_pages_check(page);
1087 #endif /* CONFIG_DEBUG_VM */
1090 * Frees a number of pages from the PCP lists
1091 * Assumes all pages on list are in same zone, and of same order.
1092 * count is the number of pages to free.
1094 * If the zone was previously in an "all pages pinned" state then look to
1095 * see if this freeing clears that state.
1097 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1098 * pinned" detection logic.
1100 static void free_pcppages_bulk(struct zone *zone, int count,
1101 struct per_cpu_pages *pcp)
1103 int migratetype = 0;
1105 bool isolated_pageblocks;
1107 spin_lock(&zone->lock);
1108 isolated_pageblocks = has_isolate_pageblock(zone);
1112 struct list_head *list;
1115 * Remove pages from lists in a round-robin fashion. A
1116 * batch_free count is maintained that is incremented when an
1117 * empty list is encountered. This is so more pages are freed
1118 * off fuller lists instead of spinning excessively around empty
1123 if (++migratetype == MIGRATE_PCPTYPES)
1125 list = &pcp->lists[migratetype];
1126 } while (list_empty(list));
1128 /* This is the only non-empty list. Free them all. */
1129 if (batch_free == MIGRATE_PCPTYPES)
1133 int mt; /* migratetype of the to-be-freed page */
1135 page = list_last_entry(list, struct page, lru);
1136 /* must delete as __free_one_page list manipulates */
1137 list_del(&page->lru);
1139 mt = get_pcppage_migratetype(page);
1140 /* MIGRATE_ISOLATE page should not go to pcplists */
1141 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1142 /* Pageblock could have been isolated meanwhile */
1143 if (unlikely(isolated_pageblocks))
1144 mt = get_pageblock_migratetype(page);
1146 if (bulkfree_pcp_prepare(page))
1149 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1150 trace_mm_page_pcpu_drain(page, 0, mt);
1151 } while (--count && --batch_free && !list_empty(list));
1153 spin_unlock(&zone->lock);
1156 static void free_one_page(struct zone *zone,
1157 struct page *page, unsigned long pfn,
1161 spin_lock(&zone->lock);
1162 if (unlikely(has_isolate_pageblock(zone) ||
1163 is_migrate_isolate(migratetype))) {
1164 migratetype = get_pfnblock_migratetype(page, pfn);
1166 __free_one_page(page, pfn, zone, order, migratetype);
1167 spin_unlock(&zone->lock);
1170 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1171 unsigned long zone, int nid)
1173 set_page_links(page, zone, nid, pfn);
1174 init_page_count(page);
1175 page_mapcount_reset(page);
1176 page_cpupid_reset_last(page);
1178 INIT_LIST_HEAD(&page->lru);
1179 #ifdef WANT_PAGE_VIRTUAL
1180 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1181 if (!is_highmem_idx(zone))
1182 set_page_address(page, __va(pfn << PAGE_SHIFT));
1186 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1189 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1192 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1193 static void init_reserved_page(unsigned long pfn)
1198 if (!early_page_uninitialised(pfn))
1201 nid = early_pfn_to_nid(pfn);
1202 pgdat = NODE_DATA(nid);
1204 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1205 struct zone *zone = &pgdat->node_zones[zid];
1207 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1210 __init_single_pfn(pfn, zid, nid);
1213 static inline void init_reserved_page(unsigned long pfn)
1216 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1219 * Initialised pages do not have PageReserved set. This function is
1220 * called for each range allocated by the bootmem allocator and
1221 * marks the pages PageReserved. The remaining valid pages are later
1222 * sent to the buddy page allocator.
1224 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1226 unsigned long start_pfn = PFN_DOWN(start);
1227 unsigned long end_pfn = PFN_UP(end);
1229 for (; start_pfn < end_pfn; start_pfn++) {
1230 if (pfn_valid(start_pfn)) {
1231 struct page *page = pfn_to_page(start_pfn);
1233 init_reserved_page(start_pfn);
1235 /* Avoid false-positive PageTail() */
1236 INIT_LIST_HEAD(&page->lru);
1238 SetPageReserved(page);
1243 static void __free_pages_ok(struct page *page, unsigned int order)
1245 unsigned long flags;
1247 unsigned long pfn = page_to_pfn(page);
1249 if (!free_pages_prepare(page, order, true))
1252 migratetype = get_pfnblock_migratetype(page, pfn);
1253 local_irq_save(flags);
1254 __count_vm_events(PGFREE, 1 << order);
1255 free_one_page(page_zone(page), page, pfn, order, migratetype);
1256 local_irq_restore(flags);
1259 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1261 unsigned int nr_pages = 1 << order;
1262 struct page *p = page;
1266 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1268 __ClearPageReserved(p);
1269 set_page_count(p, 0);
1271 __ClearPageReserved(p);
1272 set_page_count(p, 0);
1274 page_zone(page)->managed_pages += nr_pages;
1275 set_page_refcounted(page);
1276 __free_pages(page, order);
1279 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1280 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1282 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1284 int __meminit early_pfn_to_nid(unsigned long pfn)
1286 static DEFINE_SPINLOCK(early_pfn_lock);
1289 spin_lock(&early_pfn_lock);
1290 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1292 nid = first_online_node;
1293 spin_unlock(&early_pfn_lock);
1299 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1300 static inline bool __meminit __maybe_unused
1301 meminit_pfn_in_nid(unsigned long pfn, int node,
1302 struct mminit_pfnnid_cache *state)
1306 nid = __early_pfn_to_nid(pfn, state);
1307 if (nid >= 0 && nid != node)
1312 /* Only safe to use early in boot when initialisation is single-threaded */
1313 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1315 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1320 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1324 static inline bool __meminit __maybe_unused
1325 meminit_pfn_in_nid(unsigned long pfn, int node,
1326 struct mminit_pfnnid_cache *state)
1333 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1336 if (early_page_uninitialised(pfn))
1338 return __free_pages_boot_core(page, order);
1342 * Check that the whole (or subset of) a pageblock given by the interval of
1343 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1344 * with the migration of free compaction scanner. The scanners then need to
1345 * use only pfn_valid_within() check for arches that allow holes within
1348 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1350 * It's possible on some configurations to have a setup like node0 node1 node0
1351 * i.e. it's possible that all pages within a zones range of pages do not
1352 * belong to a single zone. We assume that a border between node0 and node1
1353 * can occur within a single pageblock, but not a node0 node1 node0
1354 * interleaving within a single pageblock. It is therefore sufficient to check
1355 * the first and last page of a pageblock and avoid checking each individual
1356 * page in a pageblock.
1358 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1359 unsigned long end_pfn, struct zone *zone)
1361 struct page *start_page;
1362 struct page *end_page;
1364 /* end_pfn is one past the range we are checking */
1367 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1370 start_page = pfn_to_online_page(start_pfn);
1374 if (page_zone(start_page) != zone)
1377 end_page = pfn_to_page(end_pfn);
1379 /* This gives a shorter code than deriving page_zone(end_page) */
1380 if (page_zone_id(start_page) != page_zone_id(end_page))
1386 void set_zone_contiguous(struct zone *zone)
1388 unsigned long block_start_pfn = zone->zone_start_pfn;
1389 unsigned long block_end_pfn;
1391 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1392 for (; block_start_pfn < zone_end_pfn(zone);
1393 block_start_pfn = block_end_pfn,
1394 block_end_pfn += pageblock_nr_pages) {
1396 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1398 if (!__pageblock_pfn_to_page(block_start_pfn,
1399 block_end_pfn, zone))
1403 /* We confirm that there is no hole */
1404 zone->contiguous = true;
1407 void clear_zone_contiguous(struct zone *zone)
1409 zone->contiguous = false;
1412 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1413 static void __init deferred_free_range(struct page *page,
1414 unsigned long pfn, int nr_pages)
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);
1446 /* Initialise remaining memory on a node */
1447 static int __init deferred_init_memmap(void *data)
1449 pg_data_t *pgdat = data;
1450 int nid = pgdat->node_id;
1451 struct mminit_pfnnid_cache nid_init_state = { };
1452 unsigned long start = jiffies;
1453 unsigned long nr_pages = 0;
1454 unsigned long walk_start, walk_end;
1457 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1458 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1460 if (first_init_pfn == ULONG_MAX) {
1461 pgdat_init_report_one_done();
1465 /* Bind memory initialisation thread to a local node if possible */
1466 if (!cpumask_empty(cpumask))
1467 set_cpus_allowed_ptr(current, cpumask);
1469 /* Sanity check boundaries */
1470 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1471 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1472 pgdat->first_deferred_pfn = ULONG_MAX;
1474 /* Only the highest zone is deferred so find it */
1475 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1476 zone = pgdat->node_zones + zid;
1477 if (first_init_pfn < zone_end_pfn(zone))
1481 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1482 unsigned long pfn, end_pfn;
1483 struct page *page = NULL;
1484 struct page *free_base_page = NULL;
1485 unsigned long free_base_pfn = 0;
1488 end_pfn = min(walk_end, zone_end_pfn(zone));
1489 pfn = first_init_pfn;
1490 if (pfn < walk_start)
1492 if (pfn < zone->zone_start_pfn)
1493 pfn = zone->zone_start_pfn;
1495 for (; pfn < end_pfn; pfn++) {
1496 if (!pfn_valid_within(pfn))
1500 * Ensure pfn_valid is checked every
1501 * pageblock_nr_pages for memory holes
1503 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1504 if (!pfn_valid(pfn)) {
1510 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1515 /* Minimise pfn page lookups and scheduler checks */
1516 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1519 nr_pages += nr_to_free;
1520 deferred_free_range(free_base_page,
1521 free_base_pfn, nr_to_free);
1522 free_base_page = NULL;
1523 free_base_pfn = nr_to_free = 0;
1525 page = pfn_to_page(pfn);
1530 VM_BUG_ON(page_zone(page) != zone);
1534 __init_single_page(page, pfn, zid, nid);
1535 if (!free_base_page) {
1536 free_base_page = page;
1537 free_base_pfn = pfn;
1542 /* Where possible, batch up pages for a single free */
1545 /* Free the current block of pages to allocator */
1546 nr_pages += nr_to_free;
1547 deferred_free_range(free_base_page, free_base_pfn,
1549 free_base_page = NULL;
1550 free_base_pfn = nr_to_free = 0;
1552 /* Free the last block of pages to allocator */
1553 nr_pages += nr_to_free;
1554 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1556 first_init_pfn = max(end_pfn, first_init_pfn);
1559 /* Sanity check that the next zone really is unpopulated */
1560 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1562 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1563 jiffies_to_msecs(jiffies - start));
1565 pgdat_init_report_one_done();
1568 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1570 void __init page_alloc_init_late(void)
1574 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577 /* There will be num_node_state(N_MEMORY) threads */
1578 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1579 for_each_node_state(nid, N_MEMORY) {
1580 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1583 /* Block until all are initialised */
1584 wait_for_completion(&pgdat_init_all_done_comp);
1586 /* Reinit limits that are based on free pages after the kernel is up */
1587 files_maxfiles_init();
1589 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1590 /* Discard memblock private memory */
1594 for_each_populated_zone(zone)
1595 set_zone_contiguous(zone);
1599 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1600 void __init init_cma_reserved_pageblock(struct page *page)
1602 unsigned i = pageblock_nr_pages;
1603 struct page *p = page;
1606 __ClearPageReserved(p);
1607 set_page_count(p, 0);
1610 set_pageblock_migratetype(page, MIGRATE_CMA);
1612 if (pageblock_order >= MAX_ORDER) {
1613 i = pageblock_nr_pages;
1616 set_page_refcounted(p);
1617 __free_pages(p, MAX_ORDER - 1);
1618 p += MAX_ORDER_NR_PAGES;
1619 } while (i -= MAX_ORDER_NR_PAGES);
1621 set_page_refcounted(page);
1622 __free_pages(page, pageblock_order);
1625 adjust_managed_page_count(page, pageblock_nr_pages);
1630 * The order of subdivision here is critical for the IO subsystem.
1631 * Please do not alter this order without good reasons and regression
1632 * testing. Specifically, as large blocks of memory are subdivided,
1633 * the order in which smaller blocks are delivered depends on the order
1634 * they're subdivided in this function. This is the primary factor
1635 * influencing the order in which pages are delivered to the IO
1636 * subsystem according to empirical testing, and this is also justified
1637 * by considering the behavior of a buddy system containing a single
1638 * large block of memory acted on by a series of small allocations.
1639 * This behavior is a critical factor in sglist merging's success.
1643 static inline void expand(struct zone *zone, struct page *page,
1644 int low, int high, struct free_area *area,
1647 unsigned long size = 1 << high;
1649 while (high > low) {
1653 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1656 * Mark as guard pages (or page), that will allow to
1657 * merge back to allocator when buddy will be freed.
1658 * Corresponding page table entries will not be touched,
1659 * pages will stay not present in virtual address space
1661 if (set_page_guard(zone, &page[size], high, migratetype))
1664 list_add(&page[size].lru, &area->free_list[migratetype]);
1666 set_page_order(&page[size], high);
1670 static void check_new_page_bad(struct page *page)
1672 const char *bad_reason = NULL;
1673 unsigned long bad_flags = 0;
1675 if (unlikely(atomic_read(&page->_mapcount) != -1))
1676 bad_reason = "nonzero mapcount";
1677 if (unlikely(page->mapping != NULL))
1678 bad_reason = "non-NULL mapping";
1679 if (unlikely(page_ref_count(page) != 0))
1680 bad_reason = "nonzero _count";
1681 if (unlikely(page->flags & __PG_HWPOISON)) {
1682 bad_reason = "HWPoisoned (hardware-corrupted)";
1683 bad_flags = __PG_HWPOISON;
1684 /* Don't complain about hwpoisoned pages */
1685 page_mapcount_reset(page); /* remove PageBuddy */
1688 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1689 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1690 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1693 if (unlikely(page->mem_cgroup))
1694 bad_reason = "page still charged to cgroup";
1696 bad_page(page, bad_reason, bad_flags);
1700 * This page is about to be returned from the page allocator
1702 static inline int check_new_page(struct page *page)
1704 if (likely(page_expected_state(page,
1705 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1708 check_new_page_bad(page);
1712 static inline bool free_pages_prezeroed(void)
1714 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1715 page_poisoning_enabled();
1718 #ifdef CONFIG_DEBUG_VM
1719 static bool check_pcp_refill(struct page *page)
1724 static bool check_new_pcp(struct page *page)
1726 return check_new_page(page);
1729 static bool check_pcp_refill(struct page *page)
1731 return check_new_page(page);
1733 static bool check_new_pcp(struct page *page)
1737 #endif /* CONFIG_DEBUG_VM */
1739 static bool check_new_pages(struct page *page, unsigned int order)
1742 for (i = 0; i < (1 << order); i++) {
1743 struct page *p = page + i;
1745 if (unlikely(check_new_page(p)))
1752 inline void post_alloc_hook(struct page *page, unsigned int order,
1755 set_page_private(page, 0);
1756 set_page_refcounted(page);
1758 arch_alloc_page(page, order);
1759 kernel_map_pages(page, 1 << order, 1);
1760 kernel_poison_pages(page, 1 << order, 1);
1761 kasan_alloc_pages(page, order);
1762 set_page_owner(page, order, gfp_flags);
1765 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1766 unsigned int alloc_flags)
1770 post_alloc_hook(page, order, gfp_flags);
1772 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1773 for (i = 0; i < (1 << order); i++)
1774 clear_highpage(page + i);
1776 if (order && (gfp_flags & __GFP_COMP))
1777 prep_compound_page(page, order);
1780 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1781 * allocate the page. The expectation is that the caller is taking
1782 * steps that will free more memory. The caller should avoid the page
1783 * being used for !PFMEMALLOC purposes.
1785 if (alloc_flags & ALLOC_NO_WATERMARKS)
1786 set_page_pfmemalloc(page);
1788 clear_page_pfmemalloc(page);
1792 * Go through the free lists for the given migratetype and remove
1793 * the smallest available page from the freelists
1796 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1799 unsigned int current_order;
1800 struct free_area *area;
1803 /* Find a page of the appropriate size in the preferred list */
1804 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1805 area = &(zone->free_area[current_order]);
1806 page = list_first_entry_or_null(&area->free_list[migratetype],
1810 list_del(&page->lru);
1811 rmv_page_order(page);
1813 expand(zone, page, order, current_order, area, migratetype);
1814 set_pcppage_migratetype(page, migratetype);
1823 * This array describes the order lists are fallen back to when
1824 * the free lists for the desirable migrate type are depleted
1826 static int fallbacks[MIGRATE_TYPES][4] = {
1827 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1828 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1829 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1831 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1833 #ifdef CONFIG_MEMORY_ISOLATION
1834 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1839 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1842 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1845 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1846 unsigned int order) { return NULL; }
1850 * Move the free pages in a range to the free lists of the requested type.
1851 * Note that start_page and end_pages are not aligned on a pageblock
1852 * boundary. If alignment is required, use move_freepages_block()
1854 static int move_freepages(struct zone *zone,
1855 struct page *start_page, struct page *end_page,
1856 int migratetype, int *num_movable)
1860 int pages_moved = 0;
1862 #ifndef CONFIG_HOLES_IN_ZONE
1864 * page_zone is not safe to call in this context when
1865 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1866 * anyway as we check zone boundaries in move_freepages_block().
1867 * Remove at a later date when no bug reports exist related to
1868 * grouping pages by mobility
1870 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1876 for (page = start_page; page <= end_page;) {
1877 if (!pfn_valid_within(page_to_pfn(page))) {
1882 /* Make sure we are not inadvertently changing nodes */
1883 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1885 if (!PageBuddy(page)) {
1887 * We assume that pages that could be isolated for
1888 * migration are movable. But we don't actually try
1889 * isolating, as that would be expensive.
1892 (PageLRU(page) || __PageMovable(page)))
1899 order = page_order(page);
1900 list_move(&page->lru,
1901 &zone->free_area[order].free_list[migratetype]);
1903 pages_moved += 1 << order;
1909 int move_freepages_block(struct zone *zone, struct page *page,
1910 int migratetype, int *num_movable)
1912 unsigned long start_pfn, end_pfn;
1913 struct page *start_page, *end_page;
1915 start_pfn = page_to_pfn(page);
1916 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1917 start_page = pfn_to_page(start_pfn);
1918 end_page = start_page + pageblock_nr_pages - 1;
1919 end_pfn = start_pfn + pageblock_nr_pages - 1;
1921 /* Do not cross zone boundaries */
1922 if (!zone_spans_pfn(zone, start_pfn))
1924 if (!zone_spans_pfn(zone, end_pfn))
1927 return move_freepages(zone, start_page, end_page, migratetype,
1931 static void change_pageblock_range(struct page *pageblock_page,
1932 int start_order, int migratetype)
1934 int nr_pageblocks = 1 << (start_order - pageblock_order);
1936 while (nr_pageblocks--) {
1937 set_pageblock_migratetype(pageblock_page, migratetype);
1938 pageblock_page += pageblock_nr_pages;
1943 * When we are falling back to another migratetype during allocation, try to
1944 * steal extra free pages from the same pageblocks to satisfy further
1945 * allocations, instead of polluting multiple pageblocks.
1947 * If we are stealing a relatively large buddy page, it is likely there will
1948 * be more free pages in the pageblock, so try to steal them all. For
1949 * reclaimable and unmovable allocations, we steal regardless of page size,
1950 * as fragmentation caused by those allocations polluting movable pageblocks
1951 * is worse than movable allocations stealing from unmovable and reclaimable
1954 static bool can_steal_fallback(unsigned int order, int start_mt)
1957 * Leaving this order check is intended, although there is
1958 * relaxed order check in next check. The reason is that
1959 * we can actually steal whole pageblock if this condition met,
1960 * but, below check doesn't guarantee it and that is just heuristic
1961 * so could be changed anytime.
1963 if (order >= pageblock_order)
1966 if (order >= pageblock_order / 2 ||
1967 start_mt == MIGRATE_RECLAIMABLE ||
1968 start_mt == MIGRATE_UNMOVABLE ||
1969 page_group_by_mobility_disabled)
1976 * This function implements actual steal behaviour. If order is large enough,
1977 * we can steal whole pageblock. If not, we first move freepages in this
1978 * pageblock to our migratetype and determine how many already-allocated pages
1979 * are there in the pageblock with a compatible migratetype. If at least half
1980 * of pages are free or compatible, we can change migratetype of the pageblock
1981 * itself, so pages freed in the future will be put on the correct free list.
1983 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1984 int start_type, bool whole_block)
1986 unsigned int current_order = page_order(page);
1987 struct free_area *area;
1988 int free_pages, movable_pages, alike_pages;
1991 old_block_type = get_pageblock_migratetype(page);
1994 * This can happen due to races and we want to prevent broken
1995 * highatomic accounting.
1997 if (is_migrate_highatomic(old_block_type))
2000 /* Take ownership for orders >= pageblock_order */
2001 if (current_order >= pageblock_order) {
2002 change_pageblock_range(page, current_order, start_type);
2006 /* We are not allowed to try stealing from the whole block */
2010 free_pages = move_freepages_block(zone, page, start_type,
2013 * Determine how many pages are compatible with our allocation.
2014 * For movable allocation, it's the number of movable pages which
2015 * we just obtained. For other types it's a bit more tricky.
2017 if (start_type == MIGRATE_MOVABLE) {
2018 alike_pages = movable_pages;
2021 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2022 * to MOVABLE pageblock, consider all non-movable pages as
2023 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2024 * vice versa, be conservative since we can't distinguish the
2025 * exact migratetype of non-movable pages.
2027 if (old_block_type == MIGRATE_MOVABLE)
2028 alike_pages = pageblock_nr_pages
2029 - (free_pages + movable_pages);
2034 /* moving whole block can fail due to zone boundary conditions */
2039 * If a sufficient number of pages in the block are either free or of
2040 * comparable migratability as our allocation, claim the whole block.
2042 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2043 page_group_by_mobility_disabled)
2044 set_pageblock_migratetype(page, start_type);
2049 area = &zone->free_area[current_order];
2050 list_move(&page->lru, &area->free_list[start_type]);
2054 * Check whether there is a suitable fallback freepage with requested order.
2055 * If only_stealable is true, this function returns fallback_mt only if
2056 * we can steal other freepages all together. This would help to reduce
2057 * fragmentation due to mixed migratetype pages in one pageblock.
2059 int find_suitable_fallback(struct free_area *area, unsigned int order,
2060 int migratetype, bool only_stealable, bool *can_steal)
2065 if (area->nr_free == 0)
2070 fallback_mt = fallbacks[migratetype][i];
2071 if (fallback_mt == MIGRATE_TYPES)
2074 if (list_empty(&area->free_list[fallback_mt]))
2077 if (can_steal_fallback(order, migratetype))
2080 if (!only_stealable)
2091 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2092 * there are no empty page blocks that contain a page with a suitable order
2094 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2095 unsigned int alloc_order)
2098 unsigned long max_managed, flags;
2101 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2102 * Check is race-prone but harmless.
2104 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2105 if (zone->nr_reserved_highatomic >= max_managed)
2108 spin_lock_irqsave(&zone->lock, flags);
2110 /* Recheck the nr_reserved_highatomic limit under the lock */
2111 if (zone->nr_reserved_highatomic >= max_managed)
2115 mt = get_pageblock_migratetype(page);
2116 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2117 && !is_migrate_cma(mt)) {
2118 zone->nr_reserved_highatomic += pageblock_nr_pages;
2119 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2120 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2124 spin_unlock_irqrestore(&zone->lock, flags);
2128 * Used when an allocation is about to fail under memory pressure. This
2129 * potentially hurts the reliability of high-order allocations when under
2130 * intense memory pressure but failed atomic allocations should be easier
2131 * to recover from than an OOM.
2133 * If @force is true, try to unreserve a pageblock even though highatomic
2134 * pageblock is exhausted.
2136 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2139 struct zonelist *zonelist = ac->zonelist;
2140 unsigned long flags;
2147 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2150 * Preserve at least one pageblock unless memory pressure
2153 if (!force && zone->nr_reserved_highatomic <=
2157 spin_lock_irqsave(&zone->lock, flags);
2158 for (order = 0; order < MAX_ORDER; order++) {
2159 struct free_area *area = &(zone->free_area[order]);
2161 page = list_first_entry_or_null(
2162 &area->free_list[MIGRATE_HIGHATOMIC],
2168 * In page freeing path, migratetype change is racy so
2169 * we can counter several free pages in a pageblock
2170 * in this loop althoug we changed the pageblock type
2171 * from highatomic to ac->migratetype. So we should
2172 * adjust the count once.
2174 if (is_migrate_highatomic_page(page)) {
2176 * It should never happen but changes to
2177 * locking could inadvertently allow a per-cpu
2178 * drain to add pages to MIGRATE_HIGHATOMIC
2179 * while unreserving so be safe and watch for
2182 zone->nr_reserved_highatomic -= min(
2184 zone->nr_reserved_highatomic);
2188 * Convert to ac->migratetype and avoid the normal
2189 * pageblock stealing heuristics. Minimally, the caller
2190 * is doing the work and needs the pages. More
2191 * importantly, if the block was always converted to
2192 * MIGRATE_UNMOVABLE or another type then the number
2193 * of pageblocks that cannot be completely freed
2196 set_pageblock_migratetype(page, ac->migratetype);
2197 ret = move_freepages_block(zone, page, ac->migratetype,
2200 spin_unlock_irqrestore(&zone->lock, flags);
2204 spin_unlock_irqrestore(&zone->lock, flags);
2211 * Try finding a free buddy page on the fallback list and put it on the free
2212 * list of requested migratetype, possibly along with other pages from the same
2213 * block, depending on fragmentation avoidance heuristics. Returns true if
2214 * fallback was found so that __rmqueue_smallest() can grab it.
2216 * The use of signed ints for order and current_order is a deliberate
2217 * deviation from the rest of this file, to make the for loop
2218 * condition simpler.
2221 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2223 struct free_area *area;
2230 * Find the largest available free page in the other list. This roughly
2231 * approximates finding the pageblock with the most free pages, which
2232 * would be too costly to do exactly.
2234 for (current_order = MAX_ORDER - 1; current_order >= order;
2236 area = &(zone->free_area[current_order]);
2237 fallback_mt = find_suitable_fallback(area, current_order,
2238 start_migratetype, false, &can_steal);
2239 if (fallback_mt == -1)
2243 * We cannot steal all free pages from the pageblock and the
2244 * requested migratetype is movable. In that case it's better to
2245 * steal and split the smallest available page instead of the
2246 * largest available page, because even if the next movable
2247 * allocation falls back into a different pageblock than this
2248 * one, it won't cause permanent fragmentation.
2250 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2251 && current_order > order)
2260 for (current_order = order; current_order < MAX_ORDER;
2262 area = &(zone->free_area[current_order]);
2263 fallback_mt = find_suitable_fallback(area, current_order,
2264 start_migratetype, false, &can_steal);
2265 if (fallback_mt != -1)
2270 * This should not happen - we already found a suitable fallback
2271 * when looking for the largest page.
2273 VM_BUG_ON(current_order == MAX_ORDER);
2276 page = list_first_entry(&area->free_list[fallback_mt],
2279 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2281 trace_mm_page_alloc_extfrag(page, order, current_order,
2282 start_migratetype, fallback_mt);
2289 * Do the hard work of removing an element from the buddy allocator.
2290 * Call me with the zone->lock already held.
2292 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2298 page = __rmqueue_smallest(zone, order, migratetype);
2299 if (unlikely(!page)) {
2300 if (migratetype == MIGRATE_MOVABLE)
2301 page = __rmqueue_cma_fallback(zone, order);
2303 if (!page && __rmqueue_fallback(zone, order, migratetype))
2307 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2312 * Obtain a specified number of elements from the buddy allocator, all under
2313 * a single hold of the lock, for efficiency. Add them to the supplied list.
2314 * Returns the number of new pages which were placed at *list.
2316 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2317 unsigned long count, struct list_head *list,
2318 int migratetype, bool cold)
2322 spin_lock(&zone->lock);
2323 for (i = 0; i < count; ++i) {
2324 struct page *page = __rmqueue(zone, order, migratetype);
2325 if (unlikely(page == NULL))
2328 if (unlikely(check_pcp_refill(page)))
2332 * Split buddy pages returned by expand() are received here
2333 * in physical page order. The page is added to the callers and
2334 * list and the list head then moves forward. From the callers
2335 * perspective, the linked list is ordered by page number in
2336 * some conditions. This is useful for IO devices that can
2337 * merge IO requests if the physical pages are ordered
2341 list_add(&page->lru, list);
2343 list_add_tail(&page->lru, list);
2346 if (is_migrate_cma(get_pcppage_migratetype(page)))
2347 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2352 * i pages were removed from the buddy list even if some leak due
2353 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2354 * on i. Do not confuse with 'alloced' which is the number of
2355 * pages added to the pcp list.
2357 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2358 spin_unlock(&zone->lock);
2364 * Called from the vmstat counter updater to drain pagesets of this
2365 * currently executing processor on remote nodes after they have
2368 * Note that this function must be called with the thread pinned to
2369 * a single processor.
2371 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2373 unsigned long flags;
2374 int to_drain, batch;
2376 local_irq_save(flags);
2377 batch = READ_ONCE(pcp->batch);
2378 to_drain = min(pcp->count, batch);
2380 free_pcppages_bulk(zone, to_drain, pcp);
2381 pcp->count -= to_drain;
2383 local_irq_restore(flags);
2388 * Drain pcplists of the indicated processor and zone.
2390 * The processor must either be the current processor and the
2391 * thread pinned to the current processor or a processor that
2394 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2396 unsigned long flags;
2397 struct per_cpu_pageset *pset;
2398 struct per_cpu_pages *pcp;
2400 local_irq_save(flags);
2401 pset = per_cpu_ptr(zone->pageset, cpu);
2405 free_pcppages_bulk(zone, pcp->count, pcp);
2408 local_irq_restore(flags);
2412 * Drain pcplists of all zones on the indicated processor.
2414 * The processor must either be the current processor and the
2415 * thread pinned to the current processor or a processor that
2418 static void drain_pages(unsigned int cpu)
2422 for_each_populated_zone(zone) {
2423 drain_pages_zone(cpu, zone);
2428 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2430 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2431 * the single zone's pages.
2433 void drain_local_pages(struct zone *zone)
2435 int cpu = smp_processor_id();
2438 drain_pages_zone(cpu, zone);
2443 static void drain_local_pages_wq(struct work_struct *work)
2446 * drain_all_pages doesn't use proper cpu hotplug protection so
2447 * we can race with cpu offline when the WQ can move this from
2448 * a cpu pinned worker to an unbound one. We can operate on a different
2449 * cpu which is allright but we also have to make sure to not move to
2453 drain_local_pages(NULL);
2458 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2460 * When zone parameter is non-NULL, spill just the single zone's pages.
2462 * Note that this can be extremely slow as the draining happens in a workqueue.
2464 void drain_all_pages(struct zone *zone)
2469 * Allocate in the BSS so we wont require allocation in
2470 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2472 static cpumask_t cpus_with_pcps;
2475 * Make sure nobody triggers this path before mm_percpu_wq is fully
2478 if (WARN_ON_ONCE(!mm_percpu_wq))
2481 /* Workqueues cannot recurse */
2482 if (current->flags & PF_WQ_WORKER)
2486 * Do not drain if one is already in progress unless it's specific to
2487 * a zone. Such callers are primarily CMA and memory hotplug and need
2488 * the drain to be complete when the call returns.
2490 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2493 mutex_lock(&pcpu_drain_mutex);
2497 * We don't care about racing with CPU hotplug event
2498 * as offline notification will cause the notified
2499 * cpu to drain that CPU pcps and on_each_cpu_mask
2500 * disables preemption as part of its processing
2502 for_each_online_cpu(cpu) {
2503 struct per_cpu_pageset *pcp;
2505 bool has_pcps = false;
2508 pcp = per_cpu_ptr(zone->pageset, cpu);
2512 for_each_populated_zone(z) {
2513 pcp = per_cpu_ptr(z->pageset, cpu);
2514 if (pcp->pcp.count) {
2522 cpumask_set_cpu(cpu, &cpus_with_pcps);
2524 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2527 for_each_cpu(cpu, &cpus_with_pcps) {
2528 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2529 INIT_WORK(work, drain_local_pages_wq);
2530 queue_work_on(cpu, mm_percpu_wq, work);
2532 for_each_cpu(cpu, &cpus_with_pcps)
2533 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2535 mutex_unlock(&pcpu_drain_mutex);
2538 #ifdef CONFIG_HIBERNATION
2541 * Touch the watchdog for every WD_PAGE_COUNT pages.
2543 #define WD_PAGE_COUNT (128*1024)
2545 void mark_free_pages(struct zone *zone)
2547 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2548 unsigned long flags;
2549 unsigned int order, t;
2552 if (zone_is_empty(zone))
2555 spin_lock_irqsave(&zone->lock, flags);
2557 max_zone_pfn = zone_end_pfn(zone);
2558 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2559 if (pfn_valid(pfn)) {
2560 page = pfn_to_page(pfn);
2562 if (!--page_count) {
2563 touch_nmi_watchdog();
2564 page_count = WD_PAGE_COUNT;
2567 if (page_zone(page) != zone)
2570 if (!swsusp_page_is_forbidden(page))
2571 swsusp_unset_page_free(page);
2574 for_each_migratetype_order(order, t) {
2575 list_for_each_entry(page,
2576 &zone->free_area[order].free_list[t], lru) {
2579 pfn = page_to_pfn(page);
2580 for (i = 0; i < (1UL << order); i++) {
2581 if (!--page_count) {
2582 touch_nmi_watchdog();
2583 page_count = WD_PAGE_COUNT;
2585 swsusp_set_page_free(pfn_to_page(pfn + i));
2589 spin_unlock_irqrestore(&zone->lock, flags);
2591 #endif /* CONFIG_PM */
2594 * Free a 0-order page
2595 * cold == true ? free a cold page : free a hot page
2597 void free_hot_cold_page(struct page *page, bool cold)
2599 struct zone *zone = page_zone(page);
2600 struct per_cpu_pages *pcp;
2601 unsigned long flags;
2602 unsigned long pfn = page_to_pfn(page);
2605 if (!free_pcp_prepare(page))
2608 migratetype = get_pfnblock_migratetype(page, pfn);
2609 set_pcppage_migratetype(page, migratetype);
2610 local_irq_save(flags);
2611 __count_vm_event(PGFREE);
2614 * We only track unmovable, reclaimable and movable on pcp lists.
2615 * Free ISOLATE pages back to the allocator because they are being
2616 * offlined but treat HIGHATOMIC as movable pages so we can get those
2617 * areas back if necessary. Otherwise, we may have to free
2618 * excessively into the page allocator
2620 if (migratetype >= MIGRATE_PCPTYPES) {
2621 if (unlikely(is_migrate_isolate(migratetype))) {
2622 free_one_page(zone, page, pfn, 0, migratetype);
2625 migratetype = MIGRATE_MOVABLE;
2628 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2630 list_add(&page->lru, &pcp->lists[migratetype]);
2632 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2634 if (pcp->count >= pcp->high) {
2635 unsigned long batch = READ_ONCE(pcp->batch);
2636 free_pcppages_bulk(zone, batch, pcp);
2637 pcp->count -= batch;
2641 local_irq_restore(flags);
2645 * Free a list of 0-order pages
2647 void free_hot_cold_page_list(struct list_head *list, bool cold)
2649 struct page *page, *next;
2651 list_for_each_entry_safe(page, next, list, lru) {
2652 trace_mm_page_free_batched(page, cold);
2653 free_hot_cold_page(page, cold);
2658 * split_page takes a non-compound higher-order page, and splits it into
2659 * n (1<<order) sub-pages: page[0..n]
2660 * Each sub-page must be freed individually.
2662 * Note: this is probably too low level an operation for use in drivers.
2663 * Please consult with lkml before using this in your driver.
2665 void split_page(struct page *page, unsigned int order)
2669 VM_BUG_ON_PAGE(PageCompound(page), page);
2670 VM_BUG_ON_PAGE(!page_count(page), page);
2672 #ifdef CONFIG_KMEMCHECK
2674 * Split shadow pages too, because free(page[0]) would
2675 * otherwise free the whole shadow.
2677 if (kmemcheck_page_is_tracked(page))
2678 split_page(virt_to_page(page[0].shadow), order);
2681 for (i = 1; i < (1 << order); i++)
2682 set_page_refcounted(page + i);
2683 split_page_owner(page, order);
2685 EXPORT_SYMBOL_GPL(split_page);
2687 int __isolate_free_page(struct page *page, unsigned int order)
2689 unsigned long watermark;
2693 BUG_ON(!PageBuddy(page));
2695 zone = page_zone(page);
2696 mt = get_pageblock_migratetype(page);
2698 if (!is_migrate_isolate(mt)) {
2700 * Obey watermarks as if the page was being allocated. We can
2701 * emulate a high-order watermark check with a raised order-0
2702 * watermark, because we already know our high-order page
2705 watermark = min_wmark_pages(zone) + (1UL << order);
2706 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2709 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2712 /* Remove page from free list */
2713 list_del(&page->lru);
2714 zone->free_area[order].nr_free--;
2715 rmv_page_order(page);
2718 * Set the pageblock if the isolated page is at least half of a
2721 if (order >= pageblock_order - 1) {
2722 struct page *endpage = page + (1 << order) - 1;
2723 for (; page < endpage; page += pageblock_nr_pages) {
2724 int mt = get_pageblock_migratetype(page);
2725 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2726 && !is_migrate_highatomic(mt))
2727 set_pageblock_migratetype(page,
2733 return 1UL << order;
2737 * Update NUMA hit/miss statistics
2739 * Must be called with interrupts disabled.
2741 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2744 enum zone_stat_item local_stat = NUMA_LOCAL;
2746 if (z->node != numa_node_id())
2747 local_stat = NUMA_OTHER;
2749 if (z->node == preferred_zone->node)
2750 __inc_zone_state(z, NUMA_HIT);
2752 __inc_zone_state(z, NUMA_MISS);
2753 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2755 __inc_zone_state(z, local_stat);
2759 /* Remove page from the per-cpu list, caller must protect the list */
2760 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2761 bool cold, struct per_cpu_pages *pcp,
2762 struct list_head *list)
2767 if (list_empty(list)) {
2768 pcp->count += rmqueue_bulk(zone, 0,
2771 if (unlikely(list_empty(list)))
2776 page = list_last_entry(list, struct page, lru);
2778 page = list_first_entry(list, struct page, lru);
2780 list_del(&page->lru);
2782 } while (check_new_pcp(page));
2787 /* Lock and remove page from the per-cpu list */
2788 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2789 struct zone *zone, unsigned int order,
2790 gfp_t gfp_flags, int migratetype)
2792 struct per_cpu_pages *pcp;
2793 struct list_head *list;
2794 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2796 unsigned long flags;
2798 local_irq_save(flags);
2799 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2800 list = &pcp->lists[migratetype];
2801 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2803 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2804 zone_statistics(preferred_zone, zone);
2806 local_irq_restore(flags);
2811 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2814 struct page *rmqueue(struct zone *preferred_zone,
2815 struct zone *zone, unsigned int order,
2816 gfp_t gfp_flags, unsigned int alloc_flags,
2819 unsigned long flags;
2822 if (likely(order == 0)) {
2823 page = rmqueue_pcplist(preferred_zone, zone, order,
2824 gfp_flags, migratetype);
2829 * We most definitely don't want callers attempting to
2830 * allocate greater than order-1 page units with __GFP_NOFAIL.
2832 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2833 spin_lock_irqsave(&zone->lock, flags);
2837 if (alloc_flags & ALLOC_HARDER) {
2838 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2840 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2843 page = __rmqueue(zone, order, migratetype);
2844 } while (page && check_new_pages(page, order));
2845 spin_unlock(&zone->lock);
2848 __mod_zone_freepage_state(zone, -(1 << order),
2849 get_pcppage_migratetype(page));
2851 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2852 zone_statistics(preferred_zone, zone);
2853 local_irq_restore(flags);
2856 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2860 local_irq_restore(flags);
2864 #ifdef CONFIG_FAIL_PAGE_ALLOC
2867 struct fault_attr attr;
2869 bool ignore_gfp_highmem;
2870 bool ignore_gfp_reclaim;
2872 } fail_page_alloc = {
2873 .attr = FAULT_ATTR_INITIALIZER,
2874 .ignore_gfp_reclaim = true,
2875 .ignore_gfp_highmem = true,
2879 static int __init setup_fail_page_alloc(char *str)
2881 return setup_fault_attr(&fail_page_alloc.attr, str);
2883 __setup("fail_page_alloc=", setup_fail_page_alloc);
2885 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2887 if (order < fail_page_alloc.min_order)
2889 if (gfp_mask & __GFP_NOFAIL)
2891 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2893 if (fail_page_alloc.ignore_gfp_reclaim &&
2894 (gfp_mask & __GFP_DIRECT_RECLAIM))
2897 return should_fail(&fail_page_alloc.attr, 1 << order);
2900 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2902 static int __init fail_page_alloc_debugfs(void)
2904 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2907 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2908 &fail_page_alloc.attr);
2910 return PTR_ERR(dir);
2912 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2913 &fail_page_alloc.ignore_gfp_reclaim))
2915 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2916 &fail_page_alloc.ignore_gfp_highmem))
2918 if (!debugfs_create_u32("min-order", mode, dir,
2919 &fail_page_alloc.min_order))
2924 debugfs_remove_recursive(dir);
2929 late_initcall(fail_page_alloc_debugfs);
2931 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2933 #else /* CONFIG_FAIL_PAGE_ALLOC */
2935 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2940 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2943 * Return true if free base pages are above 'mark'. For high-order checks it
2944 * will return true of the order-0 watermark is reached and there is at least
2945 * one free page of a suitable size. Checking now avoids taking the zone lock
2946 * to check in the allocation paths if no pages are free.
2948 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2949 int classzone_idx, unsigned int alloc_flags,
2954 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2956 /* free_pages may go negative - that's OK */
2957 free_pages -= (1 << order) - 1;
2959 if (alloc_flags & ALLOC_HIGH)
2963 * If the caller does not have rights to ALLOC_HARDER then subtract
2964 * the high-atomic reserves. This will over-estimate the size of the
2965 * atomic reserve but it avoids a search.
2967 if (likely(!alloc_harder))
2968 free_pages -= z->nr_reserved_highatomic;
2973 /* If allocation can't use CMA areas don't use free CMA pages */
2974 if (!(alloc_flags & ALLOC_CMA))
2975 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2979 * Check watermarks for an order-0 allocation request. If these
2980 * are not met, then a high-order request also cannot go ahead
2981 * even if a suitable page happened to be free.
2983 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2986 /* If this is an order-0 request then the watermark is fine */
2990 /* For a high-order request, check at least one suitable page is free */
2991 for (o = order; o < MAX_ORDER; o++) {
2992 struct free_area *area = &z->free_area[o];
3001 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3002 if (!list_empty(&area->free_list[mt]))
3007 if ((alloc_flags & ALLOC_CMA) &&
3008 !list_empty(&area->free_list[MIGRATE_CMA])) {
3016 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3017 int classzone_idx, unsigned int alloc_flags)
3019 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3020 zone_page_state(z, NR_FREE_PAGES));
3023 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3024 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3026 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3030 /* If allocation can't use CMA areas don't use free CMA pages */
3031 if (!(alloc_flags & ALLOC_CMA))
3032 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3036 * Fast check for order-0 only. If this fails then the reserves
3037 * need to be calculated. There is a corner case where the check
3038 * passes but only the high-order atomic reserve are free. If
3039 * the caller is !atomic then it'll uselessly search the free
3040 * list. That corner case is then slower but it is harmless.
3042 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3045 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3049 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3050 unsigned long mark, int classzone_idx)
3052 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3054 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3055 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3057 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3062 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3064 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3067 #else /* CONFIG_NUMA */
3068 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3072 #endif /* CONFIG_NUMA */
3075 * get_page_from_freelist goes through the zonelist trying to allocate
3078 static struct page *
3079 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3080 const struct alloc_context *ac)
3082 struct zoneref *z = ac->preferred_zoneref;
3084 struct pglist_data *last_pgdat_dirty_limit = NULL;
3087 * Scan zonelist, looking for a zone with enough free.
3088 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3090 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3095 if (cpusets_enabled() &&
3096 (alloc_flags & ALLOC_CPUSET) &&
3097 !__cpuset_zone_allowed(zone, gfp_mask))
3100 * When allocating a page cache page for writing, we
3101 * want to get it from a node that is within its dirty
3102 * limit, such that no single node holds more than its
3103 * proportional share of globally allowed dirty pages.
3104 * The dirty limits take into account the node's
3105 * lowmem reserves and high watermark so that kswapd
3106 * should be able to balance it without having to
3107 * write pages from its LRU list.
3109 * XXX: For now, allow allocations to potentially
3110 * exceed the per-node dirty limit in the slowpath
3111 * (spread_dirty_pages unset) before going into reclaim,
3112 * which is important when on a NUMA setup the allowed
3113 * nodes are together not big enough to reach the
3114 * global limit. The proper fix for these situations
3115 * will require awareness of nodes in the
3116 * dirty-throttling and the flusher threads.
3118 if (ac->spread_dirty_pages) {
3119 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3122 if (!node_dirty_ok(zone->zone_pgdat)) {
3123 last_pgdat_dirty_limit = zone->zone_pgdat;
3128 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3129 if (!zone_watermark_fast(zone, order, mark,
3130 ac_classzone_idx(ac), alloc_flags)) {
3133 /* Checked here to keep the fast path fast */
3134 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3135 if (alloc_flags & ALLOC_NO_WATERMARKS)
3138 if (node_reclaim_mode == 0 ||
3139 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3142 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3144 case NODE_RECLAIM_NOSCAN:
3147 case NODE_RECLAIM_FULL:
3148 /* scanned but unreclaimable */
3151 /* did we reclaim enough */
3152 if (zone_watermark_ok(zone, order, mark,
3153 ac_classzone_idx(ac), alloc_flags))
3161 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3162 gfp_mask, alloc_flags, ac->migratetype);
3164 prep_new_page(page, order, gfp_mask, alloc_flags);
3167 * If this is a high-order atomic allocation then check
3168 * if the pageblock should be reserved for the future
3170 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3171 reserve_highatomic_pageblock(page, zone, order);
3181 * Large machines with many possible nodes should not always dump per-node
3182 * meminfo in irq context.
3184 static inline bool should_suppress_show_mem(void)
3189 ret = in_interrupt();
3194 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3196 unsigned int filter = SHOW_MEM_FILTER_NODES;
3197 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3199 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3203 * This documents exceptions given to allocations in certain
3204 * contexts that are allowed to allocate outside current's set
3207 if (!(gfp_mask & __GFP_NOMEMALLOC))
3208 if (test_thread_flag(TIF_MEMDIE) ||
3209 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3210 filter &= ~SHOW_MEM_FILTER_NODES;
3211 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3212 filter &= ~SHOW_MEM_FILTER_NODES;
3214 show_mem(filter, nodemask);
3217 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3219 struct va_format vaf;
3221 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3222 DEFAULT_RATELIMIT_BURST);
3224 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3227 pr_warn("%s: ", current->comm);
3229 va_start(args, fmt);
3232 pr_cont("%pV", &vaf);
3235 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3237 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3239 pr_cont("(null)\n");
3241 cpuset_print_current_mems_allowed();
3244 warn_alloc_show_mem(gfp_mask, nodemask);
3247 static inline struct page *
3248 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3249 unsigned int alloc_flags,
3250 const struct alloc_context *ac)
3254 page = get_page_from_freelist(gfp_mask, order,
3255 alloc_flags|ALLOC_CPUSET, ac);
3257 * fallback to ignore cpuset restriction if our nodes
3261 page = get_page_from_freelist(gfp_mask, order,
3267 static inline struct page *
3268 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3269 const struct alloc_context *ac, unsigned long *did_some_progress)
3271 struct oom_control oc = {
3272 .zonelist = ac->zonelist,
3273 .nodemask = ac->nodemask,
3275 .gfp_mask = gfp_mask,
3280 *did_some_progress = 0;
3283 * Acquire the oom lock. If that fails, somebody else is
3284 * making progress for us.
3286 if (!mutex_trylock(&oom_lock)) {
3287 *did_some_progress = 1;
3288 schedule_timeout_uninterruptible(1);
3293 * Go through the zonelist yet one more time, keep very high watermark
3294 * here, this is only to catch a parallel oom killing, we must fail if
3295 * we're still under heavy pressure. But make sure that this reclaim
3296 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3297 * allocation which will never fail due to oom_lock already held.
3299 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3300 ~__GFP_DIRECT_RECLAIM, order,
3301 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3305 /* Coredumps can quickly deplete all memory reserves */
3306 if (current->flags & PF_DUMPCORE)
3308 /* The OOM killer will not help higher order allocs */
3309 if (order > PAGE_ALLOC_COSTLY_ORDER)
3312 * We have already exhausted all our reclaim opportunities without any
3313 * success so it is time to admit defeat. We will skip the OOM killer
3314 * because it is very likely that the caller has a more reasonable
3315 * fallback than shooting a random task.
3317 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3319 /* The OOM killer does not needlessly kill tasks for lowmem */
3320 if (ac->high_zoneidx < ZONE_NORMAL)
3322 if (pm_suspended_storage())
3325 * XXX: GFP_NOFS allocations should rather fail than rely on
3326 * other request to make a forward progress.
3327 * We are in an unfortunate situation where out_of_memory cannot
3328 * do much for this context but let's try it to at least get
3329 * access to memory reserved if the current task is killed (see
3330 * out_of_memory). Once filesystems are ready to handle allocation
3331 * failures more gracefully we should just bail out here.
3334 /* The OOM killer may not free memory on a specific node */
3335 if (gfp_mask & __GFP_THISNODE)
3338 /* Exhausted what can be done so it's blamo time */
3339 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3340 *did_some_progress = 1;
3343 * Help non-failing allocations by giving them access to memory
3346 if (gfp_mask & __GFP_NOFAIL)
3347 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3348 ALLOC_NO_WATERMARKS, ac);
3351 mutex_unlock(&oom_lock);
3356 * Maximum number of compaction retries wit a progress before OOM
3357 * killer is consider as the only way to move forward.
3359 #define MAX_COMPACT_RETRIES 16
3361 #ifdef CONFIG_COMPACTION
3362 /* Try memory compaction for high-order allocations before reclaim */
3363 static struct page *
3364 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3365 unsigned int alloc_flags, const struct alloc_context *ac,
3366 enum compact_priority prio, enum compact_result *compact_result)
3369 unsigned int noreclaim_flag;
3374 noreclaim_flag = memalloc_noreclaim_save();
3375 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3377 memalloc_noreclaim_restore(noreclaim_flag);
3379 if (*compact_result <= COMPACT_INACTIVE)
3383 * At least in one zone compaction wasn't deferred or skipped, so let's
3384 * count a compaction stall
3386 count_vm_event(COMPACTSTALL);
3388 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3391 struct zone *zone = page_zone(page);
3393 zone->compact_blockskip_flush = false;
3394 compaction_defer_reset(zone, order, true);
3395 count_vm_event(COMPACTSUCCESS);
3400 * It's bad if compaction run occurs and fails. The most likely reason
3401 * is that pages exist, but not enough to satisfy watermarks.
3403 count_vm_event(COMPACTFAIL);
3411 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3412 enum compact_result compact_result,
3413 enum compact_priority *compact_priority,
3414 int *compaction_retries)
3416 int max_retries = MAX_COMPACT_RETRIES;
3419 int retries = *compaction_retries;
3420 enum compact_priority priority = *compact_priority;
3425 if (compaction_made_progress(compact_result))
3426 (*compaction_retries)++;
3429 * compaction considers all the zone as desperately out of memory
3430 * so it doesn't really make much sense to retry except when the
3431 * failure could be caused by insufficient priority
3433 if (compaction_failed(compact_result))
3434 goto check_priority;
3437 * make sure the compaction wasn't deferred or didn't bail out early
3438 * due to locks contention before we declare that we should give up.
3439 * But do not retry if the given zonelist is not suitable for
3442 if (compaction_withdrawn(compact_result)) {
3443 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3448 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3449 * costly ones because they are de facto nofail and invoke OOM
3450 * killer to move on while costly can fail and users are ready
3451 * to cope with that. 1/4 retries is rather arbitrary but we
3452 * would need much more detailed feedback from compaction to
3453 * make a better decision.
3455 if (order > PAGE_ALLOC_COSTLY_ORDER)
3457 if (*compaction_retries <= max_retries) {
3463 * Make sure there are attempts at the highest priority if we exhausted
3464 * all retries or failed at the lower priorities.
3467 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3468 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3470 if (*compact_priority > min_priority) {
3471 (*compact_priority)--;
3472 *compaction_retries = 0;
3476 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3480 static inline struct page *
3481 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3482 unsigned int alloc_flags, const struct alloc_context *ac,
3483 enum compact_priority prio, enum compact_result *compact_result)
3485 *compact_result = COMPACT_SKIPPED;
3490 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3491 enum compact_result compact_result,
3492 enum compact_priority *compact_priority,
3493 int *compaction_retries)
3498 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3502 * There are setups with compaction disabled which would prefer to loop
3503 * inside the allocator rather than hit the oom killer prematurely.
3504 * Let's give them a good hope and keep retrying while the order-0
3505 * watermarks are OK.
3507 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3509 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3510 ac_classzone_idx(ac), alloc_flags))
3515 #endif /* CONFIG_COMPACTION */
3517 #ifdef CONFIG_LOCKDEP
3518 struct lockdep_map __fs_reclaim_map =
3519 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3521 static bool __need_fs_reclaim(gfp_t gfp_mask)
3523 gfp_mask = current_gfp_context(gfp_mask);
3525 /* no reclaim without waiting on it */
3526 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3529 /* this guy won't enter reclaim */
3530 if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
3533 /* We're only interested __GFP_FS allocations for now */
3534 if (!(gfp_mask & __GFP_FS))
3537 if (gfp_mask & __GFP_NOLOCKDEP)
3543 void fs_reclaim_acquire(gfp_t gfp_mask)
3545 if (__need_fs_reclaim(gfp_mask))
3546 lock_map_acquire(&__fs_reclaim_map);
3548 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3550 void fs_reclaim_release(gfp_t gfp_mask)
3552 if (__need_fs_reclaim(gfp_mask))
3553 lock_map_release(&__fs_reclaim_map);
3555 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3558 /* Perform direct synchronous page reclaim */
3560 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3561 const struct alloc_context *ac)
3563 struct reclaim_state reclaim_state;
3565 unsigned int noreclaim_flag;
3569 /* We now go into synchronous reclaim */
3570 cpuset_memory_pressure_bump();
3571 noreclaim_flag = memalloc_noreclaim_save();
3572 fs_reclaim_acquire(gfp_mask);
3573 reclaim_state.reclaimed_slab = 0;
3574 current->reclaim_state = &reclaim_state;
3576 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3579 current->reclaim_state = NULL;
3580 fs_reclaim_release(gfp_mask);
3581 memalloc_noreclaim_restore(noreclaim_flag);
3588 /* The really slow allocator path where we enter direct reclaim */
3589 static inline struct page *
3590 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3591 unsigned int alloc_flags, const struct alloc_context *ac,
3592 unsigned long *did_some_progress)
3594 struct page *page = NULL;
3595 bool drained = false;
3597 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3598 if (unlikely(!(*did_some_progress)))
3602 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3605 * If an allocation failed after direct reclaim, it could be because
3606 * pages are pinned on the per-cpu lists or in high alloc reserves.
3607 * Shrink them them and try again
3609 if (!page && !drained) {
3610 unreserve_highatomic_pageblock(ac, false);
3611 drain_all_pages(NULL);
3619 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3623 pg_data_t *last_pgdat = NULL;
3625 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3626 ac->high_zoneidx, ac->nodemask) {
3627 if (last_pgdat != zone->zone_pgdat)
3628 wakeup_kswapd(zone, order, ac->high_zoneidx);
3629 last_pgdat = zone->zone_pgdat;
3633 static inline unsigned int
3634 gfp_to_alloc_flags(gfp_t gfp_mask)
3636 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3638 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3639 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3642 * The caller may dip into page reserves a bit more if the caller
3643 * cannot run direct reclaim, or if the caller has realtime scheduling
3644 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3645 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3647 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3649 if (gfp_mask & __GFP_ATOMIC) {
3651 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3652 * if it can't schedule.
3654 if (!(gfp_mask & __GFP_NOMEMALLOC))
3655 alloc_flags |= ALLOC_HARDER;
3657 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3658 * comment for __cpuset_node_allowed().
3660 alloc_flags &= ~ALLOC_CPUSET;
3661 } else if (unlikely(rt_task(current)) && !in_interrupt())
3662 alloc_flags |= ALLOC_HARDER;
3665 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3666 alloc_flags |= ALLOC_CMA;
3671 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3673 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3676 if (gfp_mask & __GFP_MEMALLOC)
3678 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3680 if (!in_interrupt() &&
3681 ((current->flags & PF_MEMALLOC) ||
3682 unlikely(test_thread_flag(TIF_MEMDIE))))
3689 * Checks whether it makes sense to retry the reclaim to make a forward progress
3690 * for the given allocation request.
3692 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3693 * without success, or when we couldn't even meet the watermark if we
3694 * reclaimed all remaining pages on the LRU lists.
3696 * Returns true if a retry is viable or false to enter the oom path.
3699 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3700 struct alloc_context *ac, int alloc_flags,
3701 bool did_some_progress, int *no_progress_loops)
3707 * Costly allocations might have made a progress but this doesn't mean
3708 * their order will become available due to high fragmentation so
3709 * always increment the no progress counter for them
3711 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3712 *no_progress_loops = 0;
3714 (*no_progress_loops)++;
3717 * Make sure we converge to OOM if we cannot make any progress
3718 * several times in the row.
3720 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3721 /* Before OOM, exhaust highatomic_reserve */
3722 return unreserve_highatomic_pageblock(ac, true);
3726 * Keep reclaiming pages while there is a chance this will lead
3727 * somewhere. If none of the target zones can satisfy our allocation
3728 * request even if all reclaimable pages are considered then we are
3729 * screwed and have to go OOM.
3731 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3733 unsigned long available;
3734 unsigned long reclaimable;
3735 unsigned long min_wmark = min_wmark_pages(zone);
3738 available = reclaimable = zone_reclaimable_pages(zone);
3739 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3742 * Would the allocation succeed if we reclaimed all
3743 * reclaimable pages?
3745 wmark = __zone_watermark_ok(zone, order, min_wmark,
3746 ac_classzone_idx(ac), alloc_flags, available);
3747 trace_reclaim_retry_zone(z, order, reclaimable,
3748 available, min_wmark, *no_progress_loops, wmark);
3751 * If we didn't make any progress and have a lot of
3752 * dirty + writeback pages then we should wait for
3753 * an IO to complete to slow down the reclaim and
3754 * prevent from pre mature OOM
3756 if (!did_some_progress) {
3757 unsigned long write_pending;
3759 write_pending = zone_page_state_snapshot(zone,
3760 NR_ZONE_WRITE_PENDING);
3762 if (2 * write_pending > reclaimable) {
3763 congestion_wait(BLK_RW_ASYNC, HZ/10);
3769 * Memory allocation/reclaim might be called from a WQ
3770 * context and the current implementation of the WQ
3771 * concurrency control doesn't recognize that
3772 * a particular WQ is congested if the worker thread is
3773 * looping without ever sleeping. Therefore we have to
3774 * do a short sleep here rather than calling
3777 if (current->flags & PF_WQ_WORKER)
3778 schedule_timeout_uninterruptible(1);
3790 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3793 * It's possible that cpuset's mems_allowed and the nodemask from
3794 * mempolicy don't intersect. This should be normally dealt with by
3795 * policy_nodemask(), but it's possible to race with cpuset update in
3796 * such a way the check therein was true, and then it became false
3797 * before we got our cpuset_mems_cookie here.
3798 * This assumes that for all allocations, ac->nodemask can come only
3799 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3800 * when it does not intersect with the cpuset restrictions) or the
3801 * caller can deal with a violated nodemask.
3803 if (cpusets_enabled() && ac->nodemask &&
3804 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3805 ac->nodemask = NULL;
3810 * When updating a task's mems_allowed or mempolicy nodemask, it is
3811 * possible to race with parallel threads in such a way that our
3812 * allocation can fail while the mask is being updated. If we are about
3813 * to fail, check if the cpuset changed during allocation and if so,
3816 if (read_mems_allowed_retry(cpuset_mems_cookie))
3822 static inline struct page *
3823 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3824 struct alloc_context *ac)
3826 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3827 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3828 struct page *page = NULL;
3829 unsigned int alloc_flags;
3830 unsigned long did_some_progress;
3831 enum compact_priority compact_priority;
3832 enum compact_result compact_result;
3833 int compaction_retries;
3834 int no_progress_loops;
3835 unsigned long alloc_start = jiffies;
3836 unsigned int stall_timeout = 10 * HZ;
3837 unsigned int cpuset_mems_cookie;
3840 * In the slowpath, we sanity check order to avoid ever trying to
3841 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3842 * be using allocators in order of preference for an area that is
3845 if (order >= MAX_ORDER) {
3846 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3851 * We also sanity check to catch abuse of atomic reserves being used by
3852 * callers that are not in atomic context.
3854 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3855 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3856 gfp_mask &= ~__GFP_ATOMIC;
3859 compaction_retries = 0;
3860 no_progress_loops = 0;
3861 compact_priority = DEF_COMPACT_PRIORITY;
3862 cpuset_mems_cookie = read_mems_allowed_begin();
3865 * The fast path uses conservative alloc_flags to succeed only until
3866 * kswapd needs to be woken up, and to avoid the cost of setting up
3867 * alloc_flags precisely. So we do that now.
3869 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3872 * We need to recalculate the starting point for the zonelist iterator
3873 * because we might have used different nodemask in the fast path, or
3874 * there was a cpuset modification and we are retrying - otherwise we
3875 * could end up iterating over non-eligible zones endlessly.
3877 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3878 ac->high_zoneidx, ac->nodemask);
3879 if (!ac->preferred_zoneref->zone)
3882 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3883 wake_all_kswapds(order, ac);
3886 * The adjusted alloc_flags might result in immediate success, so try
3889 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3894 * For costly allocations, try direct compaction first, as it's likely
3895 * that we have enough base pages and don't need to reclaim. For non-
3896 * movable high-order allocations, do that as well, as compaction will
3897 * try prevent permanent fragmentation by migrating from blocks of the
3899 * Don't try this for allocations that are allowed to ignore
3900 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3902 if (can_direct_reclaim &&
3904 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3905 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3906 page = __alloc_pages_direct_compact(gfp_mask, order,
3908 INIT_COMPACT_PRIORITY,
3914 * Checks for costly allocations with __GFP_NORETRY, which
3915 * includes THP page fault allocations
3917 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3919 * If compaction is deferred for high-order allocations,
3920 * it is because sync compaction recently failed. If
3921 * this is the case and the caller requested a THP
3922 * allocation, we do not want to heavily disrupt the
3923 * system, so we fail the allocation instead of entering
3926 if (compact_result == COMPACT_DEFERRED)
3930 * Looks like reclaim/compaction is worth trying, but
3931 * sync compaction could be very expensive, so keep
3932 * using async compaction.
3934 compact_priority = INIT_COMPACT_PRIORITY;
3939 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3940 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3941 wake_all_kswapds(order, ac);
3943 if (gfp_pfmemalloc_allowed(gfp_mask))
3944 alloc_flags = ALLOC_NO_WATERMARKS;
3947 * Reset the zonelist iterators if memory policies can be ignored.
3948 * These allocations are high priority and system rather than user
3951 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3952 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3953 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3954 ac->high_zoneidx, ac->nodemask);
3957 /* Attempt with potentially adjusted zonelist and alloc_flags */
3958 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3962 /* Caller is not willing to reclaim, we can't balance anything */
3963 if (!can_direct_reclaim)
3966 /* Make sure we know about allocations which stall for too long */
3967 if (time_after(jiffies, alloc_start + stall_timeout)) {
3968 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
3969 "page allocation stalls for %ums, order:%u",
3970 jiffies_to_msecs(jiffies-alloc_start), order);
3971 stall_timeout += 10 * HZ;
3974 /* Avoid recursion of direct reclaim */
3975 if (current->flags & PF_MEMALLOC)
3978 /* Try direct reclaim and then allocating */
3979 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3980 &did_some_progress);
3984 /* Try direct compaction and then allocating */
3985 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3986 compact_priority, &compact_result);
3990 /* Do not loop if specifically requested */
3991 if (gfp_mask & __GFP_NORETRY)
3995 * Do not retry costly high order allocations unless they are
3996 * __GFP_RETRY_MAYFAIL
3998 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4001 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4002 did_some_progress > 0, &no_progress_loops))
4006 * It doesn't make any sense to retry for the compaction if the order-0
4007 * reclaim is not able to make any progress because the current
4008 * implementation of the compaction depends on the sufficient amount
4009 * of free memory (see __compaction_suitable)
4011 if (did_some_progress > 0 &&
4012 should_compact_retry(ac, order, alloc_flags,
4013 compact_result, &compact_priority,
4014 &compaction_retries))
4018 /* Deal with possible cpuset update races before we start OOM killing */
4019 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4022 /* Reclaim has failed us, start killing things */
4023 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4027 /* Avoid allocations with no watermarks from looping endlessly */
4028 if (test_thread_flag(TIF_MEMDIE) &&
4029 (alloc_flags == ALLOC_NO_WATERMARKS ||
4030 (gfp_mask & __GFP_NOMEMALLOC)))
4033 /* Retry as long as the OOM killer is making progress */
4034 if (did_some_progress) {
4035 no_progress_loops = 0;
4040 /* Deal with possible cpuset update races before we fail */
4041 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4045 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4048 if (gfp_mask & __GFP_NOFAIL) {
4050 * All existing users of the __GFP_NOFAIL are blockable, so warn
4051 * of any new users that actually require GFP_NOWAIT
4053 if (WARN_ON_ONCE(!can_direct_reclaim))
4057 * PF_MEMALLOC request from this context is rather bizarre
4058 * because we cannot reclaim anything and only can loop waiting
4059 * for somebody to do a work for us
4061 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4064 * non failing costly orders are a hard requirement which we
4065 * are not prepared for much so let's warn about these users
4066 * so that we can identify them and convert them to something
4069 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4072 * Help non-failing allocations by giving them access to memory
4073 * reserves but do not use ALLOC_NO_WATERMARKS because this
4074 * could deplete whole memory reserves which would just make
4075 * the situation worse
4077 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4085 warn_alloc(gfp_mask, ac->nodemask,
4086 "page allocation failure: order:%u", order);
4091 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4092 int preferred_nid, nodemask_t *nodemask,
4093 struct alloc_context *ac, gfp_t *alloc_mask,
4094 unsigned int *alloc_flags)
4096 ac->high_zoneidx = gfp_zone(gfp_mask);
4097 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4098 ac->nodemask = nodemask;
4099 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4101 if (cpusets_enabled()) {
4102 *alloc_mask |= __GFP_HARDWALL;
4104 ac->nodemask = &cpuset_current_mems_allowed;
4106 *alloc_flags |= ALLOC_CPUSET;
4109 fs_reclaim_acquire(gfp_mask);
4110 fs_reclaim_release(gfp_mask);
4112 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4114 if (should_fail_alloc_page(gfp_mask, order))
4117 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4118 *alloc_flags |= ALLOC_CMA;
4123 /* Determine whether to spread dirty pages and what the first usable zone */
4124 static inline void finalise_ac(gfp_t gfp_mask,
4125 unsigned int order, struct alloc_context *ac)
4127 /* Dirty zone balancing only done in the fast path */
4128 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4131 * The preferred zone is used for statistics but crucially it is
4132 * also used as the starting point for the zonelist iterator. It
4133 * may get reset for allocations that ignore memory policies.
4135 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4136 ac->high_zoneidx, ac->nodemask);
4140 * This is the 'heart' of the zoned buddy allocator.
4143 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4144 nodemask_t *nodemask)
4147 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4148 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
4149 struct alloc_context ac = { };
4151 gfp_mask &= gfp_allowed_mask;
4152 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4155 finalise_ac(gfp_mask, order, &ac);
4157 /* First allocation attempt */
4158 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4163 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4164 * resp. GFP_NOIO which has to be inherited for all allocation requests
4165 * from a particular context which has been marked by
4166 * memalloc_no{fs,io}_{save,restore}.
4168 alloc_mask = current_gfp_context(gfp_mask);
4169 ac.spread_dirty_pages = false;
4172 * Restore the original nodemask if it was potentially replaced with
4173 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4175 if (unlikely(ac.nodemask != nodemask))
4176 ac.nodemask = nodemask;
4178 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4181 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4182 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4183 __free_pages(page, order);
4187 if (kmemcheck_enabled && page)
4188 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4190 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4194 EXPORT_SYMBOL(__alloc_pages_nodemask);
4197 * Common helper functions.
4199 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4204 * __get_free_pages() returns a 32-bit address, which cannot represent
4207 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4209 page = alloc_pages(gfp_mask, order);
4212 return (unsigned long) page_address(page);
4214 EXPORT_SYMBOL(__get_free_pages);
4216 unsigned long get_zeroed_page(gfp_t gfp_mask)
4218 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4220 EXPORT_SYMBOL(get_zeroed_page);
4222 void __free_pages(struct page *page, unsigned int order)
4224 if (put_page_testzero(page)) {
4226 free_hot_cold_page(page, false);
4228 __free_pages_ok(page, order);
4232 EXPORT_SYMBOL(__free_pages);
4234 void free_pages(unsigned long addr, unsigned int order)
4237 VM_BUG_ON(!virt_addr_valid((void *)addr));
4238 __free_pages(virt_to_page((void *)addr), order);
4242 EXPORT_SYMBOL(free_pages);
4246 * An arbitrary-length arbitrary-offset area of memory which resides
4247 * within a 0 or higher order page. Multiple fragments within that page
4248 * are individually refcounted, in the page's reference counter.
4250 * The page_frag functions below provide a simple allocation framework for
4251 * page fragments. This is used by the network stack and network device
4252 * drivers to provide a backing region of memory for use as either an
4253 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4255 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4258 struct page *page = NULL;
4259 gfp_t gfp = gfp_mask;
4261 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4262 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4264 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4265 PAGE_FRAG_CACHE_MAX_ORDER);
4266 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4268 if (unlikely(!page))
4269 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4271 nc->va = page ? page_address(page) : NULL;
4276 void __page_frag_cache_drain(struct page *page, unsigned int count)
4278 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4280 if (page_ref_sub_and_test(page, count)) {
4281 unsigned int order = compound_order(page);
4284 free_hot_cold_page(page, false);
4286 __free_pages_ok(page, order);
4289 EXPORT_SYMBOL(__page_frag_cache_drain);
4291 void *page_frag_alloc(struct page_frag_cache *nc,
4292 unsigned int fragsz, gfp_t gfp_mask)
4294 unsigned int size = PAGE_SIZE;
4298 if (unlikely(!nc->va)) {
4300 page = __page_frag_cache_refill(nc, gfp_mask);
4304 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4305 /* if size can vary use size else just use PAGE_SIZE */
4308 /* Even if we own the page, we do not use atomic_set().
4309 * This would break get_page_unless_zero() users.
4311 page_ref_add(page, size - 1);
4313 /* reset page count bias and offset to start of new frag */
4314 nc->pfmemalloc = page_is_pfmemalloc(page);
4315 nc->pagecnt_bias = size;
4319 offset = nc->offset - fragsz;
4320 if (unlikely(offset < 0)) {
4321 page = virt_to_page(nc->va);
4323 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4326 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4327 /* if size can vary use size else just use PAGE_SIZE */
4330 /* OK, page count is 0, we can safely set it */
4331 set_page_count(page, size);
4333 /* reset page count bias and offset to start of new frag */
4334 nc->pagecnt_bias = size;
4335 offset = size - fragsz;
4339 nc->offset = offset;
4341 return nc->va + offset;
4343 EXPORT_SYMBOL(page_frag_alloc);
4346 * Frees a page fragment allocated out of either a compound or order 0 page.
4348 void page_frag_free(void *addr)
4350 struct page *page = virt_to_head_page(addr);
4352 if (unlikely(put_page_testzero(page)))
4353 __free_pages_ok(page, compound_order(page));
4355 EXPORT_SYMBOL(page_frag_free);
4357 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4361 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4362 unsigned long used = addr + PAGE_ALIGN(size);
4364 split_page(virt_to_page((void *)addr), order);
4365 while (used < alloc_end) {
4370 return (void *)addr;
4374 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4375 * @size: the number of bytes to allocate
4376 * @gfp_mask: GFP flags for the allocation
4378 * This function is similar to alloc_pages(), except that it allocates the
4379 * minimum number of pages to satisfy the request. alloc_pages() can only
4380 * allocate memory in power-of-two pages.
4382 * This function is also limited by MAX_ORDER.
4384 * Memory allocated by this function must be released by free_pages_exact().
4386 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4388 unsigned int order = get_order(size);
4391 addr = __get_free_pages(gfp_mask, order);
4392 return make_alloc_exact(addr, order, size);
4394 EXPORT_SYMBOL(alloc_pages_exact);
4397 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4399 * @nid: the preferred node ID where memory should be allocated
4400 * @size: the number of bytes to allocate
4401 * @gfp_mask: GFP flags for the allocation
4403 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4406 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4408 unsigned int order = get_order(size);
4409 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4412 return make_alloc_exact((unsigned long)page_address(p), order, size);
4416 * free_pages_exact - release memory allocated via alloc_pages_exact()
4417 * @virt: the value returned by alloc_pages_exact.
4418 * @size: size of allocation, same value as passed to alloc_pages_exact().
4420 * Release the memory allocated by a previous call to alloc_pages_exact.
4422 void free_pages_exact(void *virt, size_t size)
4424 unsigned long addr = (unsigned long)virt;
4425 unsigned long end = addr + PAGE_ALIGN(size);
4427 while (addr < end) {
4432 EXPORT_SYMBOL(free_pages_exact);
4435 * nr_free_zone_pages - count number of pages beyond high watermark
4436 * @offset: The zone index of the highest zone
4438 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4439 * high watermark within all zones at or below a given zone index. For each
4440 * zone, the number of pages is calculated as:
4442 * nr_free_zone_pages = managed_pages - high_pages
4444 static unsigned long nr_free_zone_pages(int offset)
4449 /* Just pick one node, since fallback list is circular */
4450 unsigned long sum = 0;
4452 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4454 for_each_zone_zonelist(zone, z, zonelist, offset) {
4455 unsigned long size = zone->managed_pages;
4456 unsigned long high = high_wmark_pages(zone);
4465 * nr_free_buffer_pages - count number of pages beyond high watermark
4467 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4468 * watermark within ZONE_DMA and ZONE_NORMAL.
4470 unsigned long nr_free_buffer_pages(void)
4472 return nr_free_zone_pages(gfp_zone(GFP_USER));
4474 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4477 * nr_free_pagecache_pages - count number of pages beyond high watermark
4479 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4480 * high watermark within all zones.
4482 unsigned long nr_free_pagecache_pages(void)
4484 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4487 static inline void show_node(struct zone *zone)
4489 if (IS_ENABLED(CONFIG_NUMA))
4490 printk("Node %d ", zone_to_nid(zone));
4493 long si_mem_available(void)
4496 unsigned long pagecache;
4497 unsigned long wmark_low = 0;
4498 unsigned long pages[NR_LRU_LISTS];
4502 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4503 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4506 wmark_low += zone->watermark[WMARK_LOW];
4509 * Estimate the amount of memory available for userspace allocations,
4510 * without causing swapping.
4512 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4515 * Not all the page cache can be freed, otherwise the system will
4516 * start swapping. Assume at least half of the page cache, or the
4517 * low watermark worth of cache, needs to stay.
4519 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4520 pagecache -= min(pagecache / 2, wmark_low);
4521 available += pagecache;
4524 * Part of the reclaimable slab consists of items that are in use,
4525 * and cannot be freed. Cap this estimate at the low watermark.
4527 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4528 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4535 EXPORT_SYMBOL_GPL(si_mem_available);
4537 void si_meminfo(struct sysinfo *val)
4539 val->totalram = totalram_pages;
4540 val->sharedram = global_node_page_state(NR_SHMEM);
4541 val->freeram = global_page_state(NR_FREE_PAGES);
4542 val->bufferram = nr_blockdev_pages();
4543 val->totalhigh = totalhigh_pages;
4544 val->freehigh = nr_free_highpages();
4545 val->mem_unit = PAGE_SIZE;
4548 EXPORT_SYMBOL(si_meminfo);
4551 void si_meminfo_node(struct sysinfo *val, int nid)
4553 int zone_type; /* needs to be signed */
4554 unsigned long managed_pages = 0;
4555 unsigned long managed_highpages = 0;
4556 unsigned long free_highpages = 0;
4557 pg_data_t *pgdat = NODE_DATA(nid);
4559 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4560 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4561 val->totalram = managed_pages;
4562 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4563 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4564 #ifdef CONFIG_HIGHMEM
4565 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4566 struct zone *zone = &pgdat->node_zones[zone_type];
4568 if (is_highmem(zone)) {
4569 managed_highpages += zone->managed_pages;
4570 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4573 val->totalhigh = managed_highpages;
4574 val->freehigh = free_highpages;
4576 val->totalhigh = managed_highpages;
4577 val->freehigh = free_highpages;
4579 val->mem_unit = PAGE_SIZE;
4584 * Determine whether the node should be displayed or not, depending on whether
4585 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4587 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4589 if (!(flags & SHOW_MEM_FILTER_NODES))
4593 * no node mask - aka implicit memory numa policy. Do not bother with
4594 * the synchronization - read_mems_allowed_begin - because we do not
4595 * have to be precise here.
4598 nodemask = &cpuset_current_mems_allowed;
4600 return !node_isset(nid, *nodemask);
4603 #define K(x) ((x) << (PAGE_SHIFT-10))
4605 static void show_migration_types(unsigned char type)
4607 static const char types[MIGRATE_TYPES] = {
4608 [MIGRATE_UNMOVABLE] = 'U',
4609 [MIGRATE_MOVABLE] = 'M',
4610 [MIGRATE_RECLAIMABLE] = 'E',
4611 [MIGRATE_HIGHATOMIC] = 'H',
4613 [MIGRATE_CMA] = 'C',
4615 #ifdef CONFIG_MEMORY_ISOLATION
4616 [MIGRATE_ISOLATE] = 'I',
4619 char tmp[MIGRATE_TYPES + 1];
4623 for (i = 0; i < MIGRATE_TYPES; i++) {
4624 if (type & (1 << i))
4629 printk(KERN_CONT "(%s) ", tmp);
4633 * Show free area list (used inside shift_scroll-lock stuff)
4634 * We also calculate the percentage fragmentation. We do this by counting the
4635 * memory on each free list with the exception of the first item on the list.
4638 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4641 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4643 unsigned long free_pcp = 0;
4648 for_each_populated_zone(zone) {
4649 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4652 for_each_online_cpu(cpu)
4653 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4656 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4657 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4658 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4659 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4660 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4661 " free:%lu free_pcp:%lu free_cma:%lu\n",
4662 global_node_page_state(NR_ACTIVE_ANON),
4663 global_node_page_state(NR_INACTIVE_ANON),
4664 global_node_page_state(NR_ISOLATED_ANON),
4665 global_node_page_state(NR_ACTIVE_FILE),
4666 global_node_page_state(NR_INACTIVE_FILE),
4667 global_node_page_state(NR_ISOLATED_FILE),
4668 global_node_page_state(NR_UNEVICTABLE),
4669 global_node_page_state(NR_FILE_DIRTY),
4670 global_node_page_state(NR_WRITEBACK),
4671 global_node_page_state(NR_UNSTABLE_NFS),
4672 global_node_page_state(NR_SLAB_RECLAIMABLE),
4673 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4674 global_node_page_state(NR_FILE_MAPPED),
4675 global_node_page_state(NR_SHMEM),
4676 global_page_state(NR_PAGETABLE),
4677 global_page_state(NR_BOUNCE),
4678 global_page_state(NR_FREE_PAGES),
4680 global_page_state(NR_FREE_CMA_PAGES));
4682 for_each_online_pgdat(pgdat) {
4683 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4687 " active_anon:%lukB"
4688 " inactive_anon:%lukB"
4689 " active_file:%lukB"
4690 " inactive_file:%lukB"
4691 " unevictable:%lukB"
4692 " isolated(anon):%lukB"
4693 " isolated(file):%lukB"
4698 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4700 " shmem_pmdmapped: %lukB"
4703 " writeback_tmp:%lukB"
4705 " all_unreclaimable? %s"
4708 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4709 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4710 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4711 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4712 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4713 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4714 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4715 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4716 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4717 K(node_page_state(pgdat, NR_WRITEBACK)),
4718 K(node_page_state(pgdat, NR_SHMEM)),
4719 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4720 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4721 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4723 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4725 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4726 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4727 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4731 for_each_populated_zone(zone) {
4734 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4738 for_each_online_cpu(cpu)
4739 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4748 " active_anon:%lukB"
4749 " inactive_anon:%lukB"
4750 " active_file:%lukB"
4751 " inactive_file:%lukB"
4752 " unevictable:%lukB"
4753 " writepending:%lukB"
4757 " kernel_stack:%lukB"
4765 K(zone_page_state(zone, NR_FREE_PAGES)),
4766 K(min_wmark_pages(zone)),
4767 K(low_wmark_pages(zone)),
4768 K(high_wmark_pages(zone)),
4769 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4770 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4771 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4772 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4773 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4774 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4775 K(zone->present_pages),
4776 K(zone->managed_pages),
4777 K(zone_page_state(zone, NR_MLOCK)),
4778 zone_page_state(zone, NR_KERNEL_STACK_KB),
4779 K(zone_page_state(zone, NR_PAGETABLE)),
4780 K(zone_page_state(zone, NR_BOUNCE)),
4782 K(this_cpu_read(zone->pageset->pcp.count)),
4783 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4784 printk("lowmem_reserve[]:");
4785 for (i = 0; i < MAX_NR_ZONES; i++)
4786 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4787 printk(KERN_CONT "\n");
4790 for_each_populated_zone(zone) {
4792 unsigned long nr[MAX_ORDER], flags, total = 0;
4793 unsigned char types[MAX_ORDER];
4795 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4798 printk(KERN_CONT "%s: ", zone->name);
4800 spin_lock_irqsave(&zone->lock, flags);
4801 for (order = 0; order < MAX_ORDER; order++) {
4802 struct free_area *area = &zone->free_area[order];
4805 nr[order] = area->nr_free;
4806 total += nr[order] << order;
4809 for (type = 0; type < MIGRATE_TYPES; type++) {
4810 if (!list_empty(&area->free_list[type]))
4811 types[order] |= 1 << type;
4814 spin_unlock_irqrestore(&zone->lock, flags);
4815 for (order = 0; order < MAX_ORDER; order++) {
4816 printk(KERN_CONT "%lu*%lukB ",
4817 nr[order], K(1UL) << order);
4819 show_migration_types(types[order]);
4821 printk(KERN_CONT "= %lukB\n", K(total));
4824 hugetlb_show_meminfo();
4826 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4828 show_swap_cache_info();
4831 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4833 zoneref->zone = zone;
4834 zoneref->zone_idx = zone_idx(zone);
4838 * Builds allocation fallback zone lists.
4840 * Add all populated zones of a node to the zonelist.
4842 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4846 enum zone_type zone_type = MAX_NR_ZONES;
4850 zone = pgdat->node_zones + zone_type;
4851 if (managed_zone(zone)) {
4852 zoneref_set_zone(zone,
4853 &zonelist->_zonerefs[nr_zones++]);
4854 check_highest_zone(zone_type);
4856 } while (zone_type);
4864 * 0 = automatic detection of better ordering.
4865 * 1 = order by ([node] distance, -zonetype)
4866 * 2 = order by (-zonetype, [node] distance)
4868 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4869 * the same zonelist. So only NUMA can configure this param.
4871 #define ZONELIST_ORDER_DEFAULT 0
4872 #define ZONELIST_ORDER_NODE 1
4873 #define ZONELIST_ORDER_ZONE 2
4875 /* zonelist order in the kernel.
4876 * set_zonelist_order() will set this to NODE or ZONE.
4878 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4879 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4883 /* The value user specified ....changed by config */
4884 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4885 /* string for sysctl */
4886 #define NUMA_ZONELIST_ORDER_LEN 16
4887 char numa_zonelist_order[16] = "default";
4890 * interface for configure zonelist ordering.
4891 * command line option "numa_zonelist_order"
4892 * = "[dD]efault - default, automatic configuration.
4893 * = "[nN]ode - order by node locality, then by zone within node
4894 * = "[zZ]one - order by zone, then by locality within zone
4897 static int __parse_numa_zonelist_order(char *s)
4899 if (*s == 'd' || *s == 'D') {
4900 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4901 } else if (*s == 'n' || *s == 'N') {
4902 user_zonelist_order = ZONELIST_ORDER_NODE;
4903 } else if (*s == 'z' || *s == 'Z') {
4904 user_zonelist_order = ZONELIST_ORDER_ZONE;
4906 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4912 static __init int setup_numa_zonelist_order(char *s)
4919 ret = __parse_numa_zonelist_order(s);
4921 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4925 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4928 * sysctl handler for numa_zonelist_order
4930 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4931 void __user *buffer, size_t *length,
4934 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4936 static DEFINE_MUTEX(zl_order_mutex);
4938 mutex_lock(&zl_order_mutex);
4940 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4944 strcpy(saved_string, (char *)table->data);
4946 ret = proc_dostring(table, write, buffer, length, ppos);
4950 int oldval = user_zonelist_order;
4952 ret = __parse_numa_zonelist_order((char *)table->data);
4955 * bogus value. restore saved string
4957 strncpy((char *)table->data, saved_string,
4958 NUMA_ZONELIST_ORDER_LEN);
4959 user_zonelist_order = oldval;
4960 } else if (oldval != user_zonelist_order) {
4961 mem_hotplug_begin();
4962 mutex_lock(&zonelists_mutex);
4963 build_all_zonelists(NULL, NULL);
4964 mutex_unlock(&zonelists_mutex);
4969 mutex_unlock(&zl_order_mutex);
4974 #define MAX_NODE_LOAD (nr_online_nodes)
4975 static int node_load[MAX_NUMNODES];
4978 * find_next_best_node - find the next node that should appear in a given node's fallback list
4979 * @node: node whose fallback list we're appending
4980 * @used_node_mask: nodemask_t of already used nodes
4982 * We use a number of factors to determine which is the next node that should
4983 * appear on a given node's fallback list. The node should not have appeared
4984 * already in @node's fallback list, and it should be the next closest node
4985 * according to the distance array (which contains arbitrary distance values
4986 * from each node to each node in the system), and should also prefer nodes
4987 * with no CPUs, since presumably they'll have very little allocation pressure
4988 * on them otherwise.
4989 * It returns -1 if no node is found.
4991 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4994 int min_val = INT_MAX;
4995 int best_node = NUMA_NO_NODE;
4996 const struct cpumask *tmp = cpumask_of_node(0);
4998 /* Use the local node if we haven't already */
4999 if (!node_isset(node, *used_node_mask)) {
5000 node_set(node, *used_node_mask);
5004 for_each_node_state(n, N_MEMORY) {
5006 /* Don't want a node to appear more than once */
5007 if (node_isset(n, *used_node_mask))
5010 /* Use the distance array to find the distance */
5011 val = node_distance(node, n);
5013 /* Penalize nodes under us ("prefer the next node") */
5016 /* Give preference to headless and unused nodes */
5017 tmp = cpumask_of_node(n);
5018 if (!cpumask_empty(tmp))
5019 val += PENALTY_FOR_NODE_WITH_CPUS;
5021 /* Slight preference for less loaded node */
5022 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5023 val += node_load[n];
5025 if (val < min_val) {
5032 node_set(best_node, *used_node_mask);
5039 * Build zonelists ordered by node and zones within node.
5040 * This results in maximum locality--normal zone overflows into local
5041 * DMA zone, if any--but risks exhausting DMA zone.
5043 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
5046 struct zonelist *zonelist;
5048 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5049 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
5051 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5052 zonelist->_zonerefs[j].zone = NULL;
5053 zonelist->_zonerefs[j].zone_idx = 0;
5057 * Build gfp_thisnode zonelists
5059 static void build_thisnode_zonelists(pg_data_t *pgdat)
5062 struct zonelist *zonelist;
5064 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
5065 j = build_zonelists_node(pgdat, zonelist, 0);
5066 zonelist->_zonerefs[j].zone = NULL;
5067 zonelist->_zonerefs[j].zone_idx = 0;
5071 * Build zonelists ordered by zone and nodes within zones.
5072 * This results in conserving DMA zone[s] until all Normal memory is
5073 * exhausted, but results in overflowing to remote node while memory
5074 * may still exist in local DMA zone.
5076 static int node_order[MAX_NUMNODES];
5078 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
5081 int zone_type; /* needs to be signed */
5083 struct zonelist *zonelist;
5085 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5087 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
5088 for (j = 0; j < nr_nodes; j++) {
5089 node = node_order[j];
5090 z = &NODE_DATA(node)->node_zones[zone_type];
5091 if (managed_zone(z)) {
5093 &zonelist->_zonerefs[pos++]);
5094 check_highest_zone(zone_type);
5098 zonelist->_zonerefs[pos].zone = NULL;
5099 zonelist->_zonerefs[pos].zone_idx = 0;
5102 #if defined(CONFIG_64BIT)
5104 * Devices that require DMA32/DMA are relatively rare and do not justify a
5105 * penalty to every machine in case the specialised case applies. Default
5106 * to Node-ordering on 64-bit NUMA machines
5108 static int default_zonelist_order(void)
5110 return ZONELIST_ORDER_NODE;
5114 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
5115 * by the kernel. If processes running on node 0 deplete the low memory zone
5116 * then reclaim will occur more frequency increasing stalls and potentially
5117 * be easier to OOM if a large percentage of the zone is under writeback or
5118 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
5119 * Hence, default to zone ordering on 32-bit.
5121 static int default_zonelist_order(void)
5123 return ZONELIST_ORDER_ZONE;
5125 #endif /* CONFIG_64BIT */
5127 static void set_zonelist_order(void)
5129 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
5130 current_zonelist_order = default_zonelist_order();
5132 current_zonelist_order = user_zonelist_order;
5135 static void build_zonelists(pg_data_t *pgdat)
5138 nodemask_t used_mask;
5139 int local_node, prev_node;
5140 struct zonelist *zonelist;
5141 unsigned int order = current_zonelist_order;
5143 /* initialize zonelists */
5144 for (i = 0; i < MAX_ZONELISTS; i++) {
5145 zonelist = pgdat->node_zonelists + i;
5146 zonelist->_zonerefs[0].zone = NULL;
5147 zonelist->_zonerefs[0].zone_idx = 0;
5150 /* NUMA-aware ordering of nodes */
5151 local_node = pgdat->node_id;
5152 load = nr_online_nodes;
5153 prev_node = local_node;
5154 nodes_clear(used_mask);
5156 memset(node_order, 0, sizeof(node_order));
5159 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5161 * We don't want to pressure a particular node.
5162 * So adding penalty to the first node in same
5163 * distance group to make it round-robin.
5165 if (node_distance(local_node, node) !=
5166 node_distance(local_node, prev_node))
5167 node_load[node] = load;
5171 if (order == ZONELIST_ORDER_NODE)
5172 build_zonelists_in_node_order(pgdat, node);
5174 node_order[i++] = node; /* remember order */
5177 if (order == ZONELIST_ORDER_ZONE) {
5178 /* calculate node order -- i.e., DMA last! */
5179 build_zonelists_in_zone_order(pgdat, i);
5182 build_thisnode_zonelists(pgdat);
5185 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5187 * Return node id of node used for "local" allocations.
5188 * I.e., first node id of first zone in arg node's generic zonelist.
5189 * Used for initializing percpu 'numa_mem', which is used primarily
5190 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5192 int local_memory_node(int node)
5196 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5197 gfp_zone(GFP_KERNEL),
5199 return z->zone->node;
5203 static void setup_min_unmapped_ratio(void);
5204 static void setup_min_slab_ratio(void);
5205 #else /* CONFIG_NUMA */
5207 static void set_zonelist_order(void)
5209 current_zonelist_order = ZONELIST_ORDER_ZONE;
5212 static void build_zonelists(pg_data_t *pgdat)
5214 int node, local_node;
5216 struct zonelist *zonelist;
5218 local_node = pgdat->node_id;
5220 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5221 j = build_zonelists_node(pgdat, zonelist, 0);
5224 * Now we build the zonelist so that it contains the zones
5225 * of all the other nodes.
5226 * We don't want to pressure a particular node, so when
5227 * building the zones for node N, we make sure that the
5228 * zones coming right after the local ones are those from
5229 * node N+1 (modulo N)
5231 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5232 if (!node_online(node))
5234 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5236 for (node = 0; node < local_node; node++) {
5237 if (!node_online(node))
5239 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5242 zonelist->_zonerefs[j].zone = NULL;
5243 zonelist->_zonerefs[j].zone_idx = 0;
5246 #endif /* CONFIG_NUMA */
5249 * Boot pageset table. One per cpu which is going to be used for all
5250 * zones and all nodes. The parameters will be set in such a way
5251 * that an item put on a list will immediately be handed over to
5252 * the buddy list. This is safe since pageset manipulation is done
5253 * with interrupts disabled.
5255 * The boot_pagesets must be kept even after bootup is complete for
5256 * unused processors and/or zones. They do play a role for bootstrapping
5257 * hotplugged processors.
5259 * zoneinfo_show() and maybe other functions do
5260 * not check if the processor is online before following the pageset pointer.
5261 * Other parts of the kernel may not check if the zone is available.
5263 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5264 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5265 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5266 static void setup_zone_pageset(struct zone *zone);
5269 * Global mutex to protect against size modification of zonelists
5270 * as well as to serialize pageset setup for the new populated zone.
5272 DEFINE_MUTEX(zonelists_mutex);
5274 /* return values int ....just for stop_machine() */
5275 static int __build_all_zonelists(void *data)
5279 pg_data_t *self = data;
5282 memset(node_load, 0, sizeof(node_load));
5285 if (self && !node_online(self->node_id)) {
5286 build_zonelists(self);
5289 for_each_online_node(nid) {
5290 pg_data_t *pgdat = NODE_DATA(nid);
5292 build_zonelists(pgdat);
5296 * Initialize the boot_pagesets that are going to be used
5297 * for bootstrapping processors. The real pagesets for
5298 * each zone will be allocated later when the per cpu
5299 * allocator is available.
5301 * boot_pagesets are used also for bootstrapping offline
5302 * cpus if the system is already booted because the pagesets
5303 * are needed to initialize allocators on a specific cpu too.
5304 * F.e. the percpu allocator needs the page allocator which
5305 * needs the percpu allocator in order to allocate its pagesets
5306 * (a chicken-egg dilemma).
5308 for_each_possible_cpu(cpu) {
5309 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5311 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5313 * We now know the "local memory node" for each node--
5314 * i.e., the node of the first zone in the generic zonelist.
5315 * Set up numa_mem percpu variable for on-line cpus. During
5316 * boot, only the boot cpu should be on-line; we'll init the
5317 * secondary cpus' numa_mem as they come on-line. During
5318 * node/memory hotplug, we'll fixup all on-line cpus.
5320 if (cpu_online(cpu))
5321 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5328 static noinline void __init
5329 build_all_zonelists_init(void)
5331 __build_all_zonelists(NULL);
5332 mminit_verify_zonelist();
5333 cpuset_init_current_mems_allowed();
5337 * Called with zonelists_mutex held always
5338 * unless system_state == SYSTEM_BOOTING.
5340 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5341 * [we're only called with non-NULL zone through __meminit paths] and
5342 * (2) call of __init annotated helper build_all_zonelists_init
5343 * [protected by SYSTEM_BOOTING].
5345 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5347 set_zonelist_order();
5349 if (system_state == SYSTEM_BOOTING) {
5350 build_all_zonelists_init();
5352 #ifdef CONFIG_MEMORY_HOTPLUG
5354 setup_zone_pageset(zone);
5356 /* we have to stop all cpus to guarantee there is no user
5358 stop_machine_cpuslocked(__build_all_zonelists, pgdat, NULL);
5359 /* cpuset refresh routine should be here */
5361 vm_total_pages = nr_free_pagecache_pages();
5363 * Disable grouping by mobility if the number of pages in the
5364 * system is too low to allow the mechanism to work. It would be
5365 * more accurate, but expensive to check per-zone. This check is
5366 * made on memory-hotadd so a system can start with mobility
5367 * disabled and enable it later
5369 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5370 page_group_by_mobility_disabled = 1;
5372 page_group_by_mobility_disabled = 0;
5374 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5376 zonelist_order_name[current_zonelist_order],
5377 page_group_by_mobility_disabled ? "off" : "on",
5380 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5385 * Initially all pages are reserved - free ones are freed
5386 * up by free_all_bootmem() once the early boot process is
5387 * done. Non-atomic initialization, single-pass.
5389 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5390 unsigned long start_pfn, enum memmap_context context)
5392 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5393 unsigned long end_pfn = start_pfn + size;
5394 pg_data_t *pgdat = NODE_DATA(nid);
5396 unsigned long nr_initialised = 0;
5397 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5398 struct memblock_region *r = NULL, *tmp;
5401 if (highest_memmap_pfn < end_pfn - 1)
5402 highest_memmap_pfn = end_pfn - 1;
5405 * Honor reservation requested by the driver for this ZONE_DEVICE
5408 if (altmap && start_pfn == altmap->base_pfn)
5409 start_pfn += altmap->reserve;
5411 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5413 * There can be holes in boot-time mem_map[]s handed to this
5414 * function. They do not exist on hotplugged memory.
5416 if (context != MEMMAP_EARLY)
5419 if (!early_pfn_valid(pfn)) {
5420 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5422 * Skip to the pfn preceding the next valid one (or
5423 * end_pfn), such that we hit a valid pfn (or end_pfn)
5424 * on our next iteration of the loop.
5426 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5430 if (!early_pfn_in_nid(pfn, nid))
5432 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5435 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5437 * Check given memblock attribute by firmware which can affect
5438 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5439 * mirrored, it's an overlapped memmap init. skip it.
5441 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5442 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5443 for_each_memblock(memory, tmp)
5444 if (pfn < memblock_region_memory_end_pfn(tmp))
5448 if (pfn >= memblock_region_memory_base_pfn(r) &&
5449 memblock_is_mirror(r)) {
5450 /* already initialized as NORMAL */
5451 pfn = memblock_region_memory_end_pfn(r);
5459 * Mark the block movable so that blocks are reserved for
5460 * movable at startup. This will force kernel allocations
5461 * to reserve their blocks rather than leaking throughout
5462 * the address space during boot when many long-lived
5463 * kernel allocations are made.
5465 * bitmap is created for zone's valid pfn range. but memmap
5466 * can be created for invalid pages (for alignment)
5467 * check here not to call set_pageblock_migratetype() against
5470 if (!(pfn & (pageblock_nr_pages - 1))) {
5471 struct page *page = pfn_to_page(pfn);
5473 __init_single_page(page, pfn, zone, nid);
5474 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5476 __init_single_pfn(pfn, zone, nid);
5481 static void __meminit zone_init_free_lists(struct zone *zone)
5483 unsigned int order, t;
5484 for_each_migratetype_order(order, t) {
5485 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5486 zone->free_area[order].nr_free = 0;
5490 #ifndef __HAVE_ARCH_MEMMAP_INIT
5491 #define memmap_init(size, nid, zone, start_pfn) \
5492 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5495 static int zone_batchsize(struct zone *zone)
5501 * The per-cpu-pages pools are set to around 1000th of the
5502 * size of the zone. But no more than 1/2 of a meg.
5504 * OK, so we don't know how big the cache is. So guess.
5506 batch = zone->managed_pages / 1024;
5507 if (batch * PAGE_SIZE > 512 * 1024)
5508 batch = (512 * 1024) / PAGE_SIZE;
5509 batch /= 4; /* We effectively *= 4 below */
5514 * Clamp the batch to a 2^n - 1 value. Having a power
5515 * of 2 value was found to be more likely to have
5516 * suboptimal cache aliasing properties in some cases.
5518 * For example if 2 tasks are alternately allocating
5519 * batches of pages, one task can end up with a lot
5520 * of pages of one half of the possible page colors
5521 * and the other with pages of the other colors.
5523 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5528 /* The deferral and batching of frees should be suppressed under NOMMU
5531 * The problem is that NOMMU needs to be able to allocate large chunks
5532 * of contiguous memory as there's no hardware page translation to
5533 * assemble apparent contiguous memory from discontiguous pages.
5535 * Queueing large contiguous runs of pages for batching, however,
5536 * causes the pages to actually be freed in smaller chunks. As there
5537 * can be a significant delay between the individual batches being
5538 * recycled, this leads to the once large chunks of space being
5539 * fragmented and becoming unavailable for high-order allocations.
5546 * pcp->high and pcp->batch values are related and dependent on one another:
5547 * ->batch must never be higher then ->high.
5548 * The following function updates them in a safe manner without read side
5551 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5552 * those fields changing asynchronously (acording the the above rule).
5554 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5555 * outside of boot time (or some other assurance that no concurrent updaters
5558 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5559 unsigned long batch)
5561 /* start with a fail safe value for batch */
5565 /* Update high, then batch, in order */
5572 /* a companion to pageset_set_high() */
5573 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5575 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5578 static void pageset_init(struct per_cpu_pageset *p)
5580 struct per_cpu_pages *pcp;
5583 memset(p, 0, sizeof(*p));
5587 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5588 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5591 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5594 pageset_set_batch(p, batch);
5598 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5599 * to the value high for the pageset p.
5601 static void pageset_set_high(struct per_cpu_pageset *p,
5604 unsigned long batch = max(1UL, high / 4);
5605 if ((high / 4) > (PAGE_SHIFT * 8))
5606 batch = PAGE_SHIFT * 8;
5608 pageset_update(&p->pcp, high, batch);
5611 static void pageset_set_high_and_batch(struct zone *zone,
5612 struct per_cpu_pageset *pcp)
5614 if (percpu_pagelist_fraction)
5615 pageset_set_high(pcp,
5616 (zone->managed_pages /
5617 percpu_pagelist_fraction));
5619 pageset_set_batch(pcp, zone_batchsize(zone));
5622 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5624 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5627 pageset_set_high_and_batch(zone, pcp);
5630 static void __meminit setup_zone_pageset(struct zone *zone)
5633 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5634 for_each_possible_cpu(cpu)
5635 zone_pageset_init(zone, cpu);
5639 * Allocate per cpu pagesets and initialize them.
5640 * Before this call only boot pagesets were available.
5642 void __init setup_per_cpu_pageset(void)
5644 struct pglist_data *pgdat;
5647 for_each_populated_zone(zone)
5648 setup_zone_pageset(zone);
5650 for_each_online_pgdat(pgdat)
5651 pgdat->per_cpu_nodestats =
5652 alloc_percpu(struct per_cpu_nodestat);
5655 static __meminit void zone_pcp_init(struct zone *zone)
5658 * per cpu subsystem is not up at this point. The following code
5659 * relies on the ability of the linker to provide the
5660 * offset of a (static) per cpu variable into the per cpu area.
5662 zone->pageset = &boot_pageset;
5664 if (populated_zone(zone))
5665 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5666 zone->name, zone->present_pages,
5667 zone_batchsize(zone));
5670 void __meminit init_currently_empty_zone(struct zone *zone,
5671 unsigned long zone_start_pfn,
5674 struct pglist_data *pgdat = zone->zone_pgdat;
5676 pgdat->nr_zones = zone_idx(zone) + 1;
5678 zone->zone_start_pfn = zone_start_pfn;
5680 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5681 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5683 (unsigned long)zone_idx(zone),
5684 zone_start_pfn, (zone_start_pfn + size));
5686 zone_init_free_lists(zone);
5687 zone->initialized = 1;
5690 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5691 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5694 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5696 int __meminit __early_pfn_to_nid(unsigned long pfn,
5697 struct mminit_pfnnid_cache *state)
5699 unsigned long start_pfn, end_pfn;
5702 if (state->last_start <= pfn && pfn < state->last_end)
5703 return state->last_nid;
5705 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5707 state->last_start = start_pfn;
5708 state->last_end = end_pfn;
5709 state->last_nid = nid;
5714 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5717 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5718 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5719 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5721 * If an architecture guarantees that all ranges registered contain no holes
5722 * and may be freed, this this function may be used instead of calling
5723 * memblock_free_early_nid() manually.
5725 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5727 unsigned long start_pfn, end_pfn;
5730 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5731 start_pfn = min(start_pfn, max_low_pfn);
5732 end_pfn = min(end_pfn, max_low_pfn);
5734 if (start_pfn < end_pfn)
5735 memblock_free_early_nid(PFN_PHYS(start_pfn),
5736 (end_pfn - start_pfn) << PAGE_SHIFT,
5742 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5743 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5745 * If an architecture guarantees that all ranges registered contain no holes and may
5746 * be freed, this function may be used instead of calling memory_present() manually.
5748 void __init sparse_memory_present_with_active_regions(int nid)
5750 unsigned long start_pfn, end_pfn;
5753 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5754 memory_present(this_nid, start_pfn, end_pfn);
5758 * get_pfn_range_for_nid - Return the start and end page frames for a node
5759 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5760 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5761 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5763 * It returns the start and end page frame of a node based on information
5764 * provided by memblock_set_node(). If called for a node
5765 * with no available memory, a warning is printed and the start and end
5768 void __meminit get_pfn_range_for_nid(unsigned int nid,
5769 unsigned long *start_pfn, unsigned long *end_pfn)
5771 unsigned long this_start_pfn, this_end_pfn;
5777 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5778 *start_pfn = min(*start_pfn, this_start_pfn);
5779 *end_pfn = max(*end_pfn, this_end_pfn);
5782 if (*start_pfn == -1UL)
5787 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5788 * assumption is made that zones within a node are ordered in monotonic
5789 * increasing memory addresses so that the "highest" populated zone is used
5791 static void __init find_usable_zone_for_movable(void)
5794 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5795 if (zone_index == ZONE_MOVABLE)
5798 if (arch_zone_highest_possible_pfn[zone_index] >
5799 arch_zone_lowest_possible_pfn[zone_index])
5803 VM_BUG_ON(zone_index == -1);
5804 movable_zone = zone_index;
5808 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5809 * because it is sized independent of architecture. Unlike the other zones,
5810 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5811 * in each node depending on the size of each node and how evenly kernelcore
5812 * is distributed. This helper function adjusts the zone ranges
5813 * provided by the architecture for a given node by using the end of the
5814 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5815 * zones within a node are in order of monotonic increases memory addresses
5817 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5818 unsigned long zone_type,
5819 unsigned long node_start_pfn,
5820 unsigned long node_end_pfn,
5821 unsigned long *zone_start_pfn,
5822 unsigned long *zone_end_pfn)
5824 /* Only adjust if ZONE_MOVABLE is on this node */
5825 if (zone_movable_pfn[nid]) {
5826 /* Size ZONE_MOVABLE */
5827 if (zone_type == ZONE_MOVABLE) {
5828 *zone_start_pfn = zone_movable_pfn[nid];
5829 *zone_end_pfn = min(node_end_pfn,
5830 arch_zone_highest_possible_pfn[movable_zone]);
5832 /* Adjust for ZONE_MOVABLE starting within this range */
5833 } else if (!mirrored_kernelcore &&
5834 *zone_start_pfn < zone_movable_pfn[nid] &&
5835 *zone_end_pfn > zone_movable_pfn[nid]) {
5836 *zone_end_pfn = zone_movable_pfn[nid];
5838 /* Check if this whole range is within ZONE_MOVABLE */
5839 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5840 *zone_start_pfn = *zone_end_pfn;
5845 * Return the number of pages a zone spans in a node, including holes
5846 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5848 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5849 unsigned long zone_type,
5850 unsigned long node_start_pfn,
5851 unsigned long node_end_pfn,
5852 unsigned long *zone_start_pfn,
5853 unsigned long *zone_end_pfn,
5854 unsigned long *ignored)
5856 /* When hotadd a new node from cpu_up(), the node should be empty */
5857 if (!node_start_pfn && !node_end_pfn)
5860 /* Get the start and end of the zone */
5861 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5862 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5863 adjust_zone_range_for_zone_movable(nid, zone_type,
5864 node_start_pfn, node_end_pfn,
5865 zone_start_pfn, zone_end_pfn);
5867 /* Check that this node has pages within the zone's required range */
5868 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5871 /* Move the zone boundaries inside the node if necessary */
5872 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5873 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5875 /* Return the spanned pages */
5876 return *zone_end_pfn - *zone_start_pfn;
5880 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5881 * then all holes in the requested range will be accounted for.
5883 unsigned long __meminit __absent_pages_in_range(int nid,
5884 unsigned long range_start_pfn,
5885 unsigned long range_end_pfn)
5887 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5888 unsigned long start_pfn, end_pfn;
5891 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5892 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5893 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5894 nr_absent -= end_pfn - start_pfn;
5900 * absent_pages_in_range - Return number of page frames in holes within a range
5901 * @start_pfn: The start PFN to start searching for holes
5902 * @end_pfn: The end PFN to stop searching for holes
5904 * It returns the number of pages frames in memory holes within a range.
5906 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5907 unsigned long end_pfn)
5909 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5912 /* Return the number of page frames in holes in a zone on a node */
5913 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5914 unsigned long zone_type,
5915 unsigned long node_start_pfn,
5916 unsigned long node_end_pfn,
5917 unsigned long *ignored)
5919 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5920 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5921 unsigned long zone_start_pfn, zone_end_pfn;
5922 unsigned long nr_absent;
5924 /* When hotadd a new node from cpu_up(), the node should be empty */
5925 if (!node_start_pfn && !node_end_pfn)
5928 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5929 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5931 adjust_zone_range_for_zone_movable(nid, zone_type,
5932 node_start_pfn, node_end_pfn,
5933 &zone_start_pfn, &zone_end_pfn);
5934 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5937 * ZONE_MOVABLE handling.
5938 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5941 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5942 unsigned long start_pfn, end_pfn;
5943 struct memblock_region *r;
5945 for_each_memblock(memory, r) {
5946 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5947 zone_start_pfn, zone_end_pfn);
5948 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5949 zone_start_pfn, zone_end_pfn);
5951 if (zone_type == ZONE_MOVABLE &&
5952 memblock_is_mirror(r))
5953 nr_absent += end_pfn - start_pfn;
5955 if (zone_type == ZONE_NORMAL &&
5956 !memblock_is_mirror(r))
5957 nr_absent += end_pfn - start_pfn;
5964 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5965 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5966 unsigned long zone_type,
5967 unsigned long node_start_pfn,
5968 unsigned long node_end_pfn,
5969 unsigned long *zone_start_pfn,
5970 unsigned long *zone_end_pfn,
5971 unsigned long *zones_size)
5975 *zone_start_pfn = node_start_pfn;
5976 for (zone = 0; zone < zone_type; zone++)
5977 *zone_start_pfn += zones_size[zone];
5979 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5981 return zones_size[zone_type];
5984 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5985 unsigned long zone_type,
5986 unsigned long node_start_pfn,
5987 unsigned long node_end_pfn,
5988 unsigned long *zholes_size)
5993 return zholes_size[zone_type];
5996 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5998 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5999 unsigned long node_start_pfn,
6000 unsigned long node_end_pfn,
6001 unsigned long *zones_size,
6002 unsigned long *zholes_size)
6004 unsigned long realtotalpages = 0, totalpages = 0;
6007 for (i = 0; i < MAX_NR_ZONES; i++) {
6008 struct zone *zone = pgdat->node_zones + i;
6009 unsigned long zone_start_pfn, zone_end_pfn;
6010 unsigned long size, real_size;
6012 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6018 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6019 node_start_pfn, node_end_pfn,
6022 zone->zone_start_pfn = zone_start_pfn;
6024 zone->zone_start_pfn = 0;
6025 zone->spanned_pages = size;
6026 zone->present_pages = real_size;
6029 realtotalpages += real_size;
6032 pgdat->node_spanned_pages = totalpages;
6033 pgdat->node_present_pages = realtotalpages;
6034 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6038 #ifndef CONFIG_SPARSEMEM
6040 * Calculate the size of the zone->blockflags rounded to an unsigned long
6041 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6042 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6043 * round what is now in bits to nearest long in bits, then return it in
6046 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6048 unsigned long usemapsize;
6050 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6051 usemapsize = roundup(zonesize, pageblock_nr_pages);
6052 usemapsize = usemapsize >> pageblock_order;
6053 usemapsize *= NR_PAGEBLOCK_BITS;
6054 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6056 return usemapsize / 8;
6059 static void __init setup_usemap(struct pglist_data *pgdat,
6061 unsigned long zone_start_pfn,
6062 unsigned long zonesize)
6064 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6065 zone->pageblock_flags = NULL;
6067 zone->pageblock_flags =
6068 memblock_virt_alloc_node_nopanic(usemapsize,
6072 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6073 unsigned long zone_start_pfn, unsigned long zonesize) {}
6074 #endif /* CONFIG_SPARSEMEM */
6076 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6078 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6079 void __paginginit set_pageblock_order(void)
6083 /* Check that pageblock_nr_pages has not already been setup */
6084 if (pageblock_order)
6087 if (HPAGE_SHIFT > PAGE_SHIFT)
6088 order = HUGETLB_PAGE_ORDER;
6090 order = MAX_ORDER - 1;
6093 * Assume the largest contiguous order of interest is a huge page.
6094 * This value may be variable depending on boot parameters on IA64 and
6097 pageblock_order = order;
6099 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6102 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6103 * is unused as pageblock_order is set at compile-time. See
6104 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6107 void __paginginit set_pageblock_order(void)
6111 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6113 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6114 unsigned long present_pages)
6116 unsigned long pages = spanned_pages;
6119 * Provide a more accurate estimation if there are holes within
6120 * the zone and SPARSEMEM is in use. If there are holes within the
6121 * zone, each populated memory region may cost us one or two extra
6122 * memmap pages due to alignment because memmap pages for each
6123 * populated regions may not be naturally aligned on page boundary.
6124 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6126 if (spanned_pages > present_pages + (present_pages >> 4) &&
6127 IS_ENABLED(CONFIG_SPARSEMEM))
6128 pages = present_pages;
6130 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6134 * Set up the zone data structures:
6135 * - mark all pages reserved
6136 * - mark all memory queues empty
6137 * - clear the memory bitmaps
6139 * NOTE: pgdat should get zeroed by caller.
6141 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6144 int nid = pgdat->node_id;
6146 pgdat_resize_init(pgdat);
6147 #ifdef CONFIG_NUMA_BALANCING
6148 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6149 pgdat->numabalancing_migrate_nr_pages = 0;
6150 pgdat->numabalancing_migrate_next_window = jiffies;
6152 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6153 spin_lock_init(&pgdat->split_queue_lock);
6154 INIT_LIST_HEAD(&pgdat->split_queue);
6155 pgdat->split_queue_len = 0;
6157 init_waitqueue_head(&pgdat->kswapd_wait);
6158 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6159 #ifdef CONFIG_COMPACTION
6160 init_waitqueue_head(&pgdat->kcompactd_wait);
6162 pgdat_page_ext_init(pgdat);
6163 spin_lock_init(&pgdat->lru_lock);
6164 lruvec_init(node_lruvec(pgdat));
6166 pgdat->per_cpu_nodestats = &boot_nodestats;
6168 for (j = 0; j < MAX_NR_ZONES; j++) {
6169 struct zone *zone = pgdat->node_zones + j;
6170 unsigned long size, realsize, freesize, memmap_pages;
6171 unsigned long zone_start_pfn = zone->zone_start_pfn;
6173 size = zone->spanned_pages;
6174 realsize = freesize = zone->present_pages;
6177 * Adjust freesize so that it accounts for how much memory
6178 * is used by this zone for memmap. This affects the watermark
6179 * and per-cpu initialisations
6181 memmap_pages = calc_memmap_size(size, realsize);
6182 if (!is_highmem_idx(j)) {
6183 if (freesize >= memmap_pages) {
6184 freesize -= memmap_pages;
6187 " %s zone: %lu pages used for memmap\n",
6188 zone_names[j], memmap_pages);
6190 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6191 zone_names[j], memmap_pages, freesize);
6194 /* Account for reserved pages */
6195 if (j == 0 && freesize > dma_reserve) {
6196 freesize -= dma_reserve;
6197 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6198 zone_names[0], dma_reserve);
6201 if (!is_highmem_idx(j))
6202 nr_kernel_pages += freesize;
6203 /* Charge for highmem memmap if there are enough kernel pages */
6204 else if (nr_kernel_pages > memmap_pages * 2)
6205 nr_kernel_pages -= memmap_pages;
6206 nr_all_pages += freesize;
6209 * Set an approximate value for lowmem here, it will be adjusted
6210 * when the bootmem allocator frees pages into the buddy system.
6211 * And all highmem pages will be managed by the buddy system.
6213 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6217 zone->name = zone_names[j];
6218 zone->zone_pgdat = pgdat;
6219 spin_lock_init(&zone->lock);
6220 zone_seqlock_init(zone);
6221 zone_pcp_init(zone);
6226 set_pageblock_order();
6227 setup_usemap(pgdat, zone, zone_start_pfn, size);
6228 init_currently_empty_zone(zone, zone_start_pfn, size);
6229 memmap_init(size, nid, j, zone_start_pfn);
6233 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6235 unsigned long __maybe_unused start = 0;
6236 unsigned long __maybe_unused offset = 0;
6238 /* Skip empty nodes */
6239 if (!pgdat->node_spanned_pages)
6242 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6243 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6244 offset = pgdat->node_start_pfn - start;
6245 /* ia64 gets its own node_mem_map, before this, without bootmem */
6246 if (!pgdat->node_mem_map) {
6247 unsigned long size, end;
6251 * The zone's endpoints aren't required to be MAX_ORDER
6252 * aligned but the node_mem_map endpoints must be in order
6253 * for the buddy allocator to function correctly.
6255 end = pgdat_end_pfn(pgdat);
6256 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6257 size = (end - start) * sizeof(struct page);
6258 map = alloc_remap(pgdat->node_id, size);
6260 map = memblock_virt_alloc_node_nopanic(size,
6262 pgdat->node_mem_map = map + offset;
6264 #ifndef CONFIG_NEED_MULTIPLE_NODES
6266 * With no DISCONTIG, the global mem_map is just set as node 0's
6268 if (pgdat == NODE_DATA(0)) {
6269 mem_map = NODE_DATA(0)->node_mem_map;
6270 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6271 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6273 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6276 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6279 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6280 unsigned long node_start_pfn, unsigned long *zholes_size)
6282 pg_data_t *pgdat = NODE_DATA(nid);
6283 unsigned long start_pfn = 0;
6284 unsigned long end_pfn = 0;
6286 /* pg_data_t should be reset to zero when it's allocated */
6287 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6289 pgdat->node_id = nid;
6290 pgdat->node_start_pfn = node_start_pfn;
6291 pgdat->per_cpu_nodestats = NULL;
6292 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6293 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6294 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6295 (u64)start_pfn << PAGE_SHIFT,
6296 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6298 start_pfn = node_start_pfn;
6300 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6301 zones_size, zholes_size);
6303 alloc_node_mem_map(pgdat);
6304 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6305 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6306 nid, (unsigned long)pgdat,
6307 (unsigned long)pgdat->node_mem_map);
6310 reset_deferred_meminit(pgdat);
6311 free_area_init_core(pgdat);
6314 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6316 #if MAX_NUMNODES > 1
6318 * Figure out the number of possible node ids.
6320 void __init setup_nr_node_ids(void)
6322 unsigned int highest;
6324 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6325 nr_node_ids = highest + 1;
6330 * node_map_pfn_alignment - determine the maximum internode alignment
6332 * This function should be called after node map is populated and sorted.
6333 * It calculates the maximum power of two alignment which can distinguish
6336 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6337 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6338 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6339 * shifted, 1GiB is enough and this function will indicate so.
6341 * This is used to test whether pfn -> nid mapping of the chosen memory
6342 * model has fine enough granularity to avoid incorrect mapping for the
6343 * populated node map.
6345 * Returns the determined alignment in pfn's. 0 if there is no alignment
6346 * requirement (single node).
6348 unsigned long __init node_map_pfn_alignment(void)
6350 unsigned long accl_mask = 0, last_end = 0;
6351 unsigned long start, end, mask;
6355 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6356 if (!start || last_nid < 0 || last_nid == nid) {
6363 * Start with a mask granular enough to pin-point to the
6364 * start pfn and tick off bits one-by-one until it becomes
6365 * too coarse to separate the current node from the last.
6367 mask = ~((1 << __ffs(start)) - 1);
6368 while (mask && last_end <= (start & (mask << 1)))
6371 /* accumulate all internode masks */
6375 /* convert mask to number of pages */
6376 return ~accl_mask + 1;
6379 /* Find the lowest pfn for a node */
6380 static unsigned long __init find_min_pfn_for_node(int nid)
6382 unsigned long min_pfn = ULONG_MAX;
6383 unsigned long start_pfn;
6386 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6387 min_pfn = min(min_pfn, start_pfn);
6389 if (min_pfn == ULONG_MAX) {
6390 pr_warn("Could not find start_pfn for node %d\n", nid);
6398 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6400 * It returns the minimum PFN based on information provided via
6401 * memblock_set_node().
6403 unsigned long __init find_min_pfn_with_active_regions(void)
6405 return find_min_pfn_for_node(MAX_NUMNODES);
6409 * early_calculate_totalpages()
6410 * Sum pages in active regions for movable zone.
6411 * Populate N_MEMORY for calculating usable_nodes.
6413 static unsigned long __init early_calculate_totalpages(void)
6415 unsigned long totalpages = 0;
6416 unsigned long start_pfn, end_pfn;
6419 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6420 unsigned long pages = end_pfn - start_pfn;
6422 totalpages += pages;
6424 node_set_state(nid, N_MEMORY);
6430 * Find the PFN the Movable zone begins in each node. Kernel memory
6431 * is spread evenly between nodes as long as the nodes have enough
6432 * memory. When they don't, some nodes will have more kernelcore than
6435 static void __init find_zone_movable_pfns_for_nodes(void)
6438 unsigned long usable_startpfn;
6439 unsigned long kernelcore_node, kernelcore_remaining;
6440 /* save the state before borrow the nodemask */
6441 nodemask_t saved_node_state = node_states[N_MEMORY];
6442 unsigned long totalpages = early_calculate_totalpages();
6443 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6444 struct memblock_region *r;
6446 /* Need to find movable_zone earlier when movable_node is specified. */
6447 find_usable_zone_for_movable();
6450 * If movable_node is specified, ignore kernelcore and movablecore
6453 if (movable_node_is_enabled()) {
6454 for_each_memblock(memory, r) {
6455 if (!memblock_is_hotpluggable(r))
6460 usable_startpfn = PFN_DOWN(r->base);
6461 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6462 min(usable_startpfn, zone_movable_pfn[nid]) :
6470 * If kernelcore=mirror is specified, ignore movablecore option
6472 if (mirrored_kernelcore) {
6473 bool mem_below_4gb_not_mirrored = false;
6475 for_each_memblock(memory, r) {
6476 if (memblock_is_mirror(r))
6481 usable_startpfn = memblock_region_memory_base_pfn(r);
6483 if (usable_startpfn < 0x100000) {
6484 mem_below_4gb_not_mirrored = true;
6488 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6489 min(usable_startpfn, zone_movable_pfn[nid]) :
6493 if (mem_below_4gb_not_mirrored)
6494 pr_warn("This configuration results in unmirrored kernel memory.");
6500 * If movablecore=nn[KMG] was specified, calculate what size of
6501 * kernelcore that corresponds so that memory usable for
6502 * any allocation type is evenly spread. If both kernelcore
6503 * and movablecore are specified, then the value of kernelcore
6504 * will be used for required_kernelcore if it's greater than
6505 * what movablecore would have allowed.
6507 if (required_movablecore) {
6508 unsigned long corepages;
6511 * Round-up so that ZONE_MOVABLE is at least as large as what
6512 * was requested by the user
6514 required_movablecore =
6515 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6516 required_movablecore = min(totalpages, required_movablecore);
6517 corepages = totalpages - required_movablecore;
6519 required_kernelcore = max(required_kernelcore, corepages);
6523 * If kernelcore was not specified or kernelcore size is larger
6524 * than totalpages, there is no ZONE_MOVABLE.
6526 if (!required_kernelcore || required_kernelcore >= totalpages)
6529 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6530 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6533 /* Spread kernelcore memory as evenly as possible throughout nodes */
6534 kernelcore_node = required_kernelcore / usable_nodes;
6535 for_each_node_state(nid, N_MEMORY) {
6536 unsigned long start_pfn, end_pfn;
6539 * Recalculate kernelcore_node if the division per node
6540 * now exceeds what is necessary to satisfy the requested
6541 * amount of memory for the kernel
6543 if (required_kernelcore < kernelcore_node)
6544 kernelcore_node = required_kernelcore / usable_nodes;
6547 * As the map is walked, we track how much memory is usable
6548 * by the kernel using kernelcore_remaining. When it is
6549 * 0, the rest of the node is usable by ZONE_MOVABLE
6551 kernelcore_remaining = kernelcore_node;
6553 /* Go through each range of PFNs within this node */
6554 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6555 unsigned long size_pages;
6557 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6558 if (start_pfn >= end_pfn)
6561 /* Account for what is only usable for kernelcore */
6562 if (start_pfn < usable_startpfn) {
6563 unsigned long kernel_pages;
6564 kernel_pages = min(end_pfn, usable_startpfn)
6567 kernelcore_remaining -= min(kernel_pages,
6568 kernelcore_remaining);
6569 required_kernelcore -= min(kernel_pages,
6570 required_kernelcore);
6572 /* Continue if range is now fully accounted */
6573 if (end_pfn <= usable_startpfn) {
6576 * Push zone_movable_pfn to the end so
6577 * that if we have to rebalance
6578 * kernelcore across nodes, we will
6579 * not double account here
6581 zone_movable_pfn[nid] = end_pfn;
6584 start_pfn = usable_startpfn;
6588 * The usable PFN range for ZONE_MOVABLE is from
6589 * start_pfn->end_pfn. Calculate size_pages as the
6590 * number of pages used as kernelcore
6592 size_pages = end_pfn - start_pfn;
6593 if (size_pages > kernelcore_remaining)
6594 size_pages = kernelcore_remaining;
6595 zone_movable_pfn[nid] = start_pfn + size_pages;
6598 * Some kernelcore has been met, update counts and
6599 * break if the kernelcore for this node has been
6602 required_kernelcore -= min(required_kernelcore,
6604 kernelcore_remaining -= size_pages;
6605 if (!kernelcore_remaining)
6611 * If there is still required_kernelcore, we do another pass with one
6612 * less node in the count. This will push zone_movable_pfn[nid] further
6613 * along on the nodes that still have memory until kernelcore is
6617 if (usable_nodes && required_kernelcore > usable_nodes)
6621 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6622 for (nid = 0; nid < MAX_NUMNODES; nid++)
6623 zone_movable_pfn[nid] =
6624 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6627 /* restore the node_state */
6628 node_states[N_MEMORY] = saved_node_state;
6631 /* Any regular or high memory on that node ? */
6632 static void check_for_memory(pg_data_t *pgdat, int nid)
6634 enum zone_type zone_type;
6636 if (N_MEMORY == N_NORMAL_MEMORY)
6639 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6640 struct zone *zone = &pgdat->node_zones[zone_type];
6641 if (populated_zone(zone)) {
6642 node_set_state(nid, N_HIGH_MEMORY);
6643 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6644 zone_type <= ZONE_NORMAL)
6645 node_set_state(nid, N_NORMAL_MEMORY);
6652 * free_area_init_nodes - Initialise all pg_data_t and zone data
6653 * @max_zone_pfn: an array of max PFNs for each zone
6655 * This will call free_area_init_node() for each active node in the system.
6656 * Using the page ranges provided by memblock_set_node(), the size of each
6657 * zone in each node and their holes is calculated. If the maximum PFN
6658 * between two adjacent zones match, it is assumed that the zone is empty.
6659 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6660 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6661 * starts where the previous one ended. For example, ZONE_DMA32 starts
6662 * at arch_max_dma_pfn.
6664 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6666 unsigned long start_pfn, end_pfn;
6669 /* Record where the zone boundaries are */
6670 memset(arch_zone_lowest_possible_pfn, 0,
6671 sizeof(arch_zone_lowest_possible_pfn));
6672 memset(arch_zone_highest_possible_pfn, 0,
6673 sizeof(arch_zone_highest_possible_pfn));
6675 start_pfn = find_min_pfn_with_active_regions();
6677 for (i = 0; i < MAX_NR_ZONES; i++) {
6678 if (i == ZONE_MOVABLE)
6681 end_pfn = max(max_zone_pfn[i], start_pfn);
6682 arch_zone_lowest_possible_pfn[i] = start_pfn;
6683 arch_zone_highest_possible_pfn[i] = end_pfn;
6685 start_pfn = end_pfn;
6688 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6689 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6690 find_zone_movable_pfns_for_nodes();
6692 /* Print out the zone ranges */
6693 pr_info("Zone ranges:\n");
6694 for (i = 0; i < MAX_NR_ZONES; i++) {
6695 if (i == ZONE_MOVABLE)
6697 pr_info(" %-8s ", zone_names[i]);
6698 if (arch_zone_lowest_possible_pfn[i] ==
6699 arch_zone_highest_possible_pfn[i])
6702 pr_cont("[mem %#018Lx-%#018Lx]\n",
6703 (u64)arch_zone_lowest_possible_pfn[i]
6705 ((u64)arch_zone_highest_possible_pfn[i]
6706 << PAGE_SHIFT) - 1);
6709 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6710 pr_info("Movable zone start for each node\n");
6711 for (i = 0; i < MAX_NUMNODES; i++) {
6712 if (zone_movable_pfn[i])
6713 pr_info(" Node %d: %#018Lx\n", i,
6714 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6717 /* Print out the early node map */
6718 pr_info("Early memory node ranges\n");
6719 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6720 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6721 (u64)start_pfn << PAGE_SHIFT,
6722 ((u64)end_pfn << PAGE_SHIFT) - 1);
6724 /* Initialise every node */
6725 mminit_verify_pageflags_layout();
6726 setup_nr_node_ids();
6727 for_each_online_node(nid) {
6728 pg_data_t *pgdat = NODE_DATA(nid);
6729 free_area_init_node(nid, NULL,
6730 find_min_pfn_for_node(nid), NULL);
6732 /* Any memory on that node */
6733 if (pgdat->node_present_pages)
6734 node_set_state(nid, N_MEMORY);
6735 check_for_memory(pgdat, nid);
6739 static int __init cmdline_parse_core(char *p, unsigned long *core)
6741 unsigned long long coremem;
6745 coremem = memparse(p, &p);
6746 *core = coremem >> PAGE_SHIFT;
6748 /* Paranoid check that UL is enough for the coremem value */
6749 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6755 * kernelcore=size sets the amount of memory for use for allocations that
6756 * cannot be reclaimed or migrated.
6758 static int __init cmdline_parse_kernelcore(char *p)
6760 /* parse kernelcore=mirror */
6761 if (parse_option_str(p, "mirror")) {
6762 mirrored_kernelcore = true;
6766 return cmdline_parse_core(p, &required_kernelcore);
6770 * movablecore=size sets the amount of memory for use for allocations that
6771 * can be reclaimed or migrated.
6773 static int __init cmdline_parse_movablecore(char *p)
6775 return cmdline_parse_core(p, &required_movablecore);
6778 early_param("kernelcore", cmdline_parse_kernelcore);
6779 early_param("movablecore", cmdline_parse_movablecore);
6781 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6783 void adjust_managed_page_count(struct page *page, long count)
6785 spin_lock(&managed_page_count_lock);
6786 page_zone(page)->managed_pages += count;
6787 totalram_pages += count;
6788 #ifdef CONFIG_HIGHMEM
6789 if (PageHighMem(page))
6790 totalhigh_pages += count;
6792 spin_unlock(&managed_page_count_lock);
6794 EXPORT_SYMBOL(adjust_managed_page_count);
6796 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6799 unsigned long pages = 0;
6801 start = (void *)PAGE_ALIGN((unsigned long)start);
6802 end = (void *)((unsigned long)end & PAGE_MASK);
6803 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6804 if ((unsigned int)poison <= 0xFF)
6805 memset(pos, poison, PAGE_SIZE);
6806 free_reserved_page(virt_to_page(pos));
6810 pr_info("Freeing %s memory: %ldK\n",
6811 s, pages << (PAGE_SHIFT - 10));
6815 EXPORT_SYMBOL(free_reserved_area);
6817 #ifdef CONFIG_HIGHMEM
6818 void free_highmem_page(struct page *page)
6820 __free_reserved_page(page);
6822 page_zone(page)->managed_pages++;
6828 void __init mem_init_print_info(const char *str)
6830 unsigned long physpages, codesize, datasize, rosize, bss_size;
6831 unsigned long init_code_size, init_data_size;
6833 physpages = get_num_physpages();
6834 codesize = _etext - _stext;
6835 datasize = _edata - _sdata;
6836 rosize = __end_rodata - __start_rodata;
6837 bss_size = __bss_stop - __bss_start;
6838 init_data_size = __init_end - __init_begin;
6839 init_code_size = _einittext - _sinittext;
6842 * Detect special cases and adjust section sizes accordingly:
6843 * 1) .init.* may be embedded into .data sections
6844 * 2) .init.text.* may be out of [__init_begin, __init_end],
6845 * please refer to arch/tile/kernel/vmlinux.lds.S.
6846 * 3) .rodata.* may be embedded into .text or .data sections.
6848 #define adj_init_size(start, end, size, pos, adj) \
6850 if (start <= pos && pos < end && size > adj) \
6854 adj_init_size(__init_begin, __init_end, init_data_size,
6855 _sinittext, init_code_size);
6856 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6857 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6858 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6859 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6861 #undef adj_init_size
6863 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6864 #ifdef CONFIG_HIGHMEM
6868 nr_free_pages() << (PAGE_SHIFT - 10),
6869 physpages << (PAGE_SHIFT - 10),
6870 codesize >> 10, datasize >> 10, rosize >> 10,
6871 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6872 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6873 totalcma_pages << (PAGE_SHIFT - 10),
6874 #ifdef CONFIG_HIGHMEM
6875 totalhigh_pages << (PAGE_SHIFT - 10),
6877 str ? ", " : "", str ? str : "");
6881 * set_dma_reserve - set the specified number of pages reserved in the first zone
6882 * @new_dma_reserve: The number of pages to mark reserved
6884 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6885 * In the DMA zone, a significant percentage may be consumed by kernel image
6886 * and other unfreeable allocations which can skew the watermarks badly. This
6887 * function may optionally be used to account for unfreeable pages in the
6888 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6889 * smaller per-cpu batchsize.
6891 void __init set_dma_reserve(unsigned long new_dma_reserve)
6893 dma_reserve = new_dma_reserve;
6896 void __init free_area_init(unsigned long *zones_size)
6898 free_area_init_node(0, zones_size,
6899 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6902 static int page_alloc_cpu_dead(unsigned int cpu)
6905 lru_add_drain_cpu(cpu);
6909 * Spill the event counters of the dead processor
6910 * into the current processors event counters.
6911 * This artificially elevates the count of the current
6914 vm_events_fold_cpu(cpu);
6917 * Zero the differential counters of the dead processor
6918 * so that the vm statistics are consistent.
6920 * This is only okay since the processor is dead and cannot
6921 * race with what we are doing.
6923 cpu_vm_stats_fold(cpu);
6927 void __init page_alloc_init(void)
6931 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6932 "mm/page_alloc:dead", NULL,
6933 page_alloc_cpu_dead);
6938 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6939 * or min_free_kbytes changes.
6941 static void calculate_totalreserve_pages(void)
6943 struct pglist_data *pgdat;
6944 unsigned long reserve_pages = 0;
6945 enum zone_type i, j;
6947 for_each_online_pgdat(pgdat) {
6949 pgdat->totalreserve_pages = 0;
6951 for (i = 0; i < MAX_NR_ZONES; i++) {
6952 struct zone *zone = pgdat->node_zones + i;
6955 /* Find valid and maximum lowmem_reserve in the zone */
6956 for (j = i; j < MAX_NR_ZONES; j++) {
6957 if (zone->lowmem_reserve[j] > max)
6958 max = zone->lowmem_reserve[j];
6961 /* we treat the high watermark as reserved pages. */
6962 max += high_wmark_pages(zone);
6964 if (max > zone->managed_pages)
6965 max = zone->managed_pages;
6967 pgdat->totalreserve_pages += max;
6969 reserve_pages += max;
6972 totalreserve_pages = reserve_pages;
6976 * setup_per_zone_lowmem_reserve - called whenever
6977 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6978 * has a correct pages reserved value, so an adequate number of
6979 * pages are left in the zone after a successful __alloc_pages().
6981 static void setup_per_zone_lowmem_reserve(void)
6983 struct pglist_data *pgdat;
6984 enum zone_type j, idx;
6986 for_each_online_pgdat(pgdat) {
6987 for (j = 0; j < MAX_NR_ZONES; j++) {
6988 struct zone *zone = pgdat->node_zones + j;
6989 unsigned long managed_pages = zone->managed_pages;
6991 zone->lowmem_reserve[j] = 0;
6995 struct zone *lower_zone;
6999 if (sysctl_lowmem_reserve_ratio[idx] < 1)
7000 sysctl_lowmem_reserve_ratio[idx] = 1;
7002 lower_zone = pgdat->node_zones + idx;
7003 lower_zone->lowmem_reserve[j] = managed_pages /
7004 sysctl_lowmem_reserve_ratio[idx];
7005 managed_pages += lower_zone->managed_pages;
7010 /* update totalreserve_pages */
7011 calculate_totalreserve_pages();
7014 static void __setup_per_zone_wmarks(void)
7016 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7017 unsigned long lowmem_pages = 0;
7019 unsigned long flags;
7021 /* Calculate total number of !ZONE_HIGHMEM pages */
7022 for_each_zone(zone) {
7023 if (!is_highmem(zone))
7024 lowmem_pages += zone->managed_pages;
7027 for_each_zone(zone) {
7030 spin_lock_irqsave(&zone->lock, flags);
7031 tmp = (u64)pages_min * zone->managed_pages;
7032 do_div(tmp, lowmem_pages);
7033 if (is_highmem(zone)) {
7035 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7036 * need highmem pages, so cap pages_min to a small
7039 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7040 * deltas control asynch page reclaim, and so should
7041 * not be capped for highmem.
7043 unsigned long min_pages;
7045 min_pages = zone->managed_pages / 1024;
7046 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7047 zone->watermark[WMARK_MIN] = min_pages;
7050 * If it's a lowmem zone, reserve a number of pages
7051 * proportionate to the zone's size.
7053 zone->watermark[WMARK_MIN] = tmp;
7057 * Set the kswapd watermarks distance according to the
7058 * scale factor in proportion to available memory, but
7059 * ensure a minimum size on small systems.
7061 tmp = max_t(u64, tmp >> 2,
7062 mult_frac(zone->managed_pages,
7063 watermark_scale_factor, 10000));
7065 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7066 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7068 spin_unlock_irqrestore(&zone->lock, flags);
7071 /* update totalreserve_pages */
7072 calculate_totalreserve_pages();
7076 * setup_per_zone_wmarks - called when min_free_kbytes changes
7077 * or when memory is hot-{added|removed}
7079 * Ensures that the watermark[min,low,high] values for each zone are set
7080 * correctly with respect to min_free_kbytes.
7082 void setup_per_zone_wmarks(void)
7084 mutex_lock(&zonelists_mutex);
7085 __setup_per_zone_wmarks();
7086 mutex_unlock(&zonelists_mutex);
7090 * Initialise min_free_kbytes.
7092 * For small machines we want it small (128k min). For large machines
7093 * we want it large (64MB max). But it is not linear, because network
7094 * bandwidth does not increase linearly with machine size. We use
7096 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7097 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7113 int __meminit init_per_zone_wmark_min(void)
7115 unsigned long lowmem_kbytes;
7116 int new_min_free_kbytes;
7118 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7119 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7121 if (new_min_free_kbytes > user_min_free_kbytes) {
7122 min_free_kbytes = new_min_free_kbytes;
7123 if (min_free_kbytes < 128)
7124 min_free_kbytes = 128;
7125 if (min_free_kbytes > 65536)
7126 min_free_kbytes = 65536;
7128 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7129 new_min_free_kbytes, user_min_free_kbytes);
7131 setup_per_zone_wmarks();
7132 refresh_zone_stat_thresholds();
7133 setup_per_zone_lowmem_reserve();
7136 setup_min_unmapped_ratio();
7137 setup_min_slab_ratio();
7142 core_initcall(init_per_zone_wmark_min)
7145 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7146 * that we can call two helper functions whenever min_free_kbytes
7149 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7150 void __user *buffer, size_t *length, loff_t *ppos)
7154 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7159 user_min_free_kbytes = min_free_kbytes;
7160 setup_per_zone_wmarks();
7165 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7166 void __user *buffer, size_t *length, loff_t *ppos)
7170 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7175 setup_per_zone_wmarks();
7181 static void setup_min_unmapped_ratio(void)
7186 for_each_online_pgdat(pgdat)
7187 pgdat->min_unmapped_pages = 0;
7190 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7191 sysctl_min_unmapped_ratio) / 100;
7195 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7196 void __user *buffer, size_t *length, loff_t *ppos)
7200 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7204 setup_min_unmapped_ratio();
7209 static void setup_min_slab_ratio(void)
7214 for_each_online_pgdat(pgdat)
7215 pgdat->min_slab_pages = 0;
7218 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7219 sysctl_min_slab_ratio) / 100;
7222 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7223 void __user *buffer, size_t *length, loff_t *ppos)
7227 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7231 setup_min_slab_ratio();
7238 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7239 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7240 * whenever sysctl_lowmem_reserve_ratio changes.
7242 * The reserve ratio obviously has absolutely no relation with the
7243 * minimum watermarks. The lowmem reserve ratio can only make sense
7244 * if in function of the boot time zone sizes.
7246 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7247 void __user *buffer, size_t *length, loff_t *ppos)
7249 proc_dointvec_minmax(table, write, buffer, length, ppos);
7250 setup_per_zone_lowmem_reserve();
7255 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7256 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7257 * pagelist can have before it gets flushed back to buddy allocator.
7259 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7260 void __user *buffer, size_t *length, loff_t *ppos)
7263 int old_percpu_pagelist_fraction;
7266 mutex_lock(&pcp_batch_high_lock);
7267 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7269 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7270 if (!write || ret < 0)
7273 /* Sanity checking to avoid pcp imbalance */
7274 if (percpu_pagelist_fraction &&
7275 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7276 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7282 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7285 for_each_populated_zone(zone) {
7288 for_each_possible_cpu(cpu)
7289 pageset_set_high_and_batch(zone,
7290 per_cpu_ptr(zone->pageset, cpu));
7293 mutex_unlock(&pcp_batch_high_lock);
7298 int hashdist = HASHDIST_DEFAULT;
7300 static int __init set_hashdist(char *str)
7304 hashdist = simple_strtoul(str, &str, 0);
7307 __setup("hashdist=", set_hashdist);
7310 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7312 * Returns the number of pages that arch has reserved but
7313 * is not known to alloc_large_system_hash().
7315 static unsigned long __init arch_reserved_kernel_pages(void)
7322 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7323 * machines. As memory size is increased the scale is also increased but at
7324 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7325 * quadruples the scale is increased by one, which means the size of hash table
7326 * only doubles, instead of quadrupling as well.
7327 * Because 32-bit systems cannot have large physical memory, where this scaling
7328 * makes sense, it is disabled on such platforms.
7330 #if __BITS_PER_LONG > 32
7331 #define ADAPT_SCALE_BASE (64ul << 30)
7332 #define ADAPT_SCALE_SHIFT 2
7333 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7337 * allocate a large system hash table from bootmem
7338 * - it is assumed that the hash table must contain an exact power-of-2
7339 * quantity of entries
7340 * - limit is the number of hash buckets, not the total allocation size
7342 void *__init alloc_large_system_hash(const char *tablename,
7343 unsigned long bucketsize,
7344 unsigned long numentries,
7347 unsigned int *_hash_shift,
7348 unsigned int *_hash_mask,
7349 unsigned long low_limit,
7350 unsigned long high_limit)
7352 unsigned long long max = high_limit;
7353 unsigned long log2qty, size;
7357 /* allow the kernel cmdline to have a say */
7359 /* round applicable memory size up to nearest megabyte */
7360 numentries = nr_kernel_pages;
7361 numentries -= arch_reserved_kernel_pages();
7363 /* It isn't necessary when PAGE_SIZE >= 1MB */
7364 if (PAGE_SHIFT < 20)
7365 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7367 #if __BITS_PER_LONG > 32
7369 unsigned long adapt;
7371 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7372 adapt <<= ADAPT_SCALE_SHIFT)
7377 /* limit to 1 bucket per 2^scale bytes of low memory */
7378 if (scale > PAGE_SHIFT)
7379 numentries >>= (scale - PAGE_SHIFT);
7381 numentries <<= (PAGE_SHIFT - scale);
7383 /* Make sure we've got at least a 0-order allocation.. */
7384 if (unlikely(flags & HASH_SMALL)) {
7385 /* Makes no sense without HASH_EARLY */
7386 WARN_ON(!(flags & HASH_EARLY));
7387 if (!(numentries >> *_hash_shift)) {
7388 numentries = 1UL << *_hash_shift;
7389 BUG_ON(!numentries);
7391 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7392 numentries = PAGE_SIZE / bucketsize;
7394 numentries = roundup_pow_of_two(numentries);
7396 /* limit allocation size to 1/16 total memory by default */
7398 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7399 do_div(max, bucketsize);
7401 max = min(max, 0x80000000ULL);
7403 if (numentries < low_limit)
7404 numentries = low_limit;
7405 if (numentries > max)
7408 log2qty = ilog2(numentries);
7411 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7412 * currently not used when HASH_EARLY is specified.
7414 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7416 size = bucketsize << log2qty;
7417 if (flags & HASH_EARLY)
7418 table = memblock_virt_alloc_nopanic(size, 0);
7420 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7423 * If bucketsize is not a power-of-two, we may free
7424 * some pages at the end of hash table which
7425 * alloc_pages_exact() automatically does
7427 if (get_order(size) < MAX_ORDER) {
7428 table = alloc_pages_exact(size, gfp_flags);
7429 kmemleak_alloc(table, size, 1, gfp_flags);
7432 } while (!table && size > PAGE_SIZE && --log2qty);
7435 panic("Failed to allocate %s hash table\n", tablename);
7437 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7438 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7441 *_hash_shift = log2qty;
7443 *_hash_mask = (1 << log2qty) - 1;
7449 * This function checks whether pageblock includes unmovable pages or not.
7450 * If @count is not zero, it is okay to include less @count unmovable pages
7452 * PageLRU check without isolation or lru_lock could race so that
7453 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7454 * check without lock_page also may miss some movable non-lru pages at
7455 * race condition. So you can't expect this function should be exact.
7457 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7458 bool skip_hwpoisoned_pages)
7460 unsigned long pfn, iter, found;
7464 * For avoiding noise data, lru_add_drain_all() should be called
7465 * If ZONE_MOVABLE, the zone never contains unmovable pages
7467 if (zone_idx(zone) == ZONE_MOVABLE)
7469 mt = get_pageblock_migratetype(page);
7470 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7473 pfn = page_to_pfn(page);
7474 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7475 unsigned long check = pfn + iter;
7477 if (!pfn_valid_within(check))
7480 page = pfn_to_page(check);
7483 * Hugepages are not in LRU lists, but they're movable.
7484 * We need not scan over tail pages bacause we don't
7485 * handle each tail page individually in migration.
7487 if (PageHuge(page)) {
7488 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7493 * We can't use page_count without pin a page
7494 * because another CPU can free compound page.
7495 * This check already skips compound tails of THP
7496 * because their page->_refcount is zero at all time.
7498 if (!page_ref_count(page)) {
7499 if (PageBuddy(page))
7500 iter += (1 << page_order(page)) - 1;
7505 * The HWPoisoned page may be not in buddy system, and
7506 * page_count() is not 0.
7508 if (skip_hwpoisoned_pages && PageHWPoison(page))
7511 if (__PageMovable(page))
7517 * If there are RECLAIMABLE pages, we need to check
7518 * it. But now, memory offline itself doesn't call
7519 * shrink_node_slabs() and it still to be fixed.
7522 * If the page is not RAM, page_count()should be 0.
7523 * we don't need more check. This is an _used_ not-movable page.
7525 * The problematic thing here is PG_reserved pages. PG_reserved
7526 * is set to both of a memory hole page and a _used_ kernel
7535 bool is_pageblock_removable_nolock(struct page *page)
7541 * We have to be careful here because we are iterating over memory
7542 * sections which are not zone aware so we might end up outside of
7543 * the zone but still within the section.
7544 * We have to take care about the node as well. If the node is offline
7545 * its NODE_DATA will be NULL - see page_zone.
7547 if (!node_online(page_to_nid(page)))
7550 zone = page_zone(page);
7551 pfn = page_to_pfn(page);
7552 if (!zone_spans_pfn(zone, pfn))
7555 return !has_unmovable_pages(zone, page, 0, true);
7558 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7560 static unsigned long pfn_max_align_down(unsigned long pfn)
7562 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7563 pageblock_nr_pages) - 1);
7566 static unsigned long pfn_max_align_up(unsigned long pfn)
7568 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7569 pageblock_nr_pages));
7572 /* [start, end) must belong to a single zone. */
7573 static int __alloc_contig_migrate_range(struct compact_control *cc,
7574 unsigned long start, unsigned long end)
7576 /* This function is based on compact_zone() from compaction.c. */
7577 unsigned long nr_reclaimed;
7578 unsigned long pfn = start;
7579 unsigned int tries = 0;
7584 while (pfn < end || !list_empty(&cc->migratepages)) {
7585 if (fatal_signal_pending(current)) {
7590 if (list_empty(&cc->migratepages)) {
7591 cc->nr_migratepages = 0;
7592 pfn = isolate_migratepages_range(cc, pfn, end);
7598 } else if (++tries == 5) {
7599 ret = ret < 0 ? ret : -EBUSY;
7603 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7605 cc->nr_migratepages -= nr_reclaimed;
7607 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7608 NULL, 0, cc->mode, MR_CMA);
7611 putback_movable_pages(&cc->migratepages);
7618 * alloc_contig_range() -- tries to allocate given range of pages
7619 * @start: start PFN to allocate
7620 * @end: one-past-the-last PFN to allocate
7621 * @migratetype: migratetype of the underlaying pageblocks (either
7622 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7623 * in range must have the same migratetype and it must
7624 * be either of the two.
7625 * @gfp_mask: GFP mask to use during compaction
7627 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7628 * aligned, however it's the caller's responsibility to guarantee that
7629 * we are the only thread that changes migrate type of pageblocks the
7632 * The PFN range must belong to a single zone.
7634 * Returns zero on success or negative error code. On success all
7635 * pages which PFN is in [start, end) are allocated for the caller and
7636 * need to be freed with free_contig_range().
7638 int alloc_contig_range(unsigned long start, unsigned long end,
7639 unsigned migratetype, gfp_t gfp_mask)
7641 unsigned long outer_start, outer_end;
7645 struct compact_control cc = {
7646 .nr_migratepages = 0,
7648 .zone = page_zone(pfn_to_page(start)),
7649 .mode = MIGRATE_SYNC,
7650 .ignore_skip_hint = true,
7651 .gfp_mask = current_gfp_context(gfp_mask),
7653 INIT_LIST_HEAD(&cc.migratepages);
7656 * What we do here is we mark all pageblocks in range as
7657 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7658 * have different sizes, and due to the way page allocator
7659 * work, we align the range to biggest of the two pages so
7660 * that page allocator won't try to merge buddies from
7661 * different pageblocks and change MIGRATE_ISOLATE to some
7662 * other migration type.
7664 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7665 * migrate the pages from an unaligned range (ie. pages that
7666 * we are interested in). This will put all the pages in
7667 * range back to page allocator as MIGRATE_ISOLATE.
7669 * When this is done, we take the pages in range from page
7670 * allocator removing them from the buddy system. This way
7671 * page allocator will never consider using them.
7673 * This lets us mark the pageblocks back as
7674 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7675 * aligned range but not in the unaligned, original range are
7676 * put back to page allocator so that buddy can use them.
7679 ret = start_isolate_page_range(pfn_max_align_down(start),
7680 pfn_max_align_up(end), migratetype,
7686 * In case of -EBUSY, we'd like to know which page causes problem.
7687 * So, just fall through. We will check it in test_pages_isolated().
7689 ret = __alloc_contig_migrate_range(&cc, start, end);
7690 if (ret && ret != -EBUSY)
7694 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7695 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7696 * more, all pages in [start, end) are free in page allocator.
7697 * What we are going to do is to allocate all pages from
7698 * [start, end) (that is remove them from page allocator).
7700 * The only problem is that pages at the beginning and at the
7701 * end of interesting range may be not aligned with pages that
7702 * page allocator holds, ie. they can be part of higher order
7703 * pages. Because of this, we reserve the bigger range and
7704 * once this is done free the pages we are not interested in.
7706 * We don't have to hold zone->lock here because the pages are
7707 * isolated thus they won't get removed from buddy.
7710 lru_add_drain_all();
7711 drain_all_pages(cc.zone);
7714 outer_start = start;
7715 while (!PageBuddy(pfn_to_page(outer_start))) {
7716 if (++order >= MAX_ORDER) {
7717 outer_start = start;
7720 outer_start &= ~0UL << order;
7723 if (outer_start != start) {
7724 order = page_order(pfn_to_page(outer_start));
7727 * outer_start page could be small order buddy page and
7728 * it doesn't include start page. Adjust outer_start
7729 * in this case to report failed page properly
7730 * on tracepoint in test_pages_isolated()
7732 if (outer_start + (1UL << order) <= start)
7733 outer_start = start;
7736 /* Make sure the range is really isolated. */
7737 if (test_pages_isolated(outer_start, end, false)) {
7738 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7739 __func__, outer_start, end);
7744 /* Grab isolated pages from freelists. */
7745 outer_end = isolate_freepages_range(&cc, outer_start, end);
7751 /* Free head and tail (if any) */
7752 if (start != outer_start)
7753 free_contig_range(outer_start, start - outer_start);
7754 if (end != outer_end)
7755 free_contig_range(end, outer_end - end);
7758 undo_isolate_page_range(pfn_max_align_down(start),
7759 pfn_max_align_up(end), migratetype);
7763 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7765 unsigned int count = 0;
7767 for (; nr_pages--; pfn++) {
7768 struct page *page = pfn_to_page(pfn);
7770 count += page_count(page) != 1;
7773 WARN(count != 0, "%d pages are still in use!\n", count);
7777 #ifdef CONFIG_MEMORY_HOTPLUG
7779 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7780 * page high values need to be recalulated.
7782 void __meminit zone_pcp_update(struct zone *zone)
7785 mutex_lock(&pcp_batch_high_lock);
7786 for_each_possible_cpu(cpu)
7787 pageset_set_high_and_batch(zone,
7788 per_cpu_ptr(zone->pageset, cpu));
7789 mutex_unlock(&pcp_batch_high_lock);
7793 void zone_pcp_reset(struct zone *zone)
7795 unsigned long flags;
7797 struct per_cpu_pageset *pset;
7799 /* avoid races with drain_pages() */
7800 local_irq_save(flags);
7801 if (zone->pageset != &boot_pageset) {
7802 for_each_online_cpu(cpu) {
7803 pset = per_cpu_ptr(zone->pageset, cpu);
7804 drain_zonestat(zone, pset);
7806 free_percpu(zone->pageset);
7807 zone->pageset = &boot_pageset;
7809 local_irq_restore(flags);
7812 #ifdef CONFIG_MEMORY_HOTREMOVE
7814 * All pages in the range must be in a single zone and isolated
7815 * before calling this.
7818 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7822 unsigned int order, i;
7824 unsigned long flags;
7825 /* find the first valid pfn */
7826 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7831 offline_mem_sections(pfn, end_pfn);
7832 zone = page_zone(pfn_to_page(pfn));
7833 spin_lock_irqsave(&zone->lock, flags);
7835 while (pfn < end_pfn) {
7836 if (!pfn_valid(pfn)) {
7840 page = pfn_to_page(pfn);
7842 * The HWPoisoned page may be not in buddy system, and
7843 * page_count() is not 0.
7845 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7847 SetPageReserved(page);
7851 BUG_ON(page_count(page));
7852 BUG_ON(!PageBuddy(page));
7853 order = page_order(page);
7854 #ifdef CONFIG_DEBUG_VM
7855 pr_info("remove from free list %lx %d %lx\n",
7856 pfn, 1 << order, end_pfn);
7858 list_del(&page->lru);
7859 rmv_page_order(page);
7860 zone->free_area[order].nr_free--;
7861 for (i = 0; i < (1 << order); i++)
7862 SetPageReserved((page+i));
7863 pfn += (1 << order);
7865 spin_unlock_irqrestore(&zone->lock, flags);
7869 bool is_free_buddy_page(struct page *page)
7871 struct zone *zone = page_zone(page);
7872 unsigned long pfn = page_to_pfn(page);
7873 unsigned long flags;
7876 spin_lock_irqsave(&zone->lock, flags);
7877 for (order = 0; order < MAX_ORDER; order++) {
7878 struct page *page_head = page - (pfn & ((1 << order) - 1));
7880 if (PageBuddy(page_head) && page_order(page_head) >= order)
7883 spin_unlock_irqrestore(&zone->lock, flags);
7885 return order < MAX_ORDER;