2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
87 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
88 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
89 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
90 * defined in <linux/topology.h>.
92 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
93 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
94 int _node_numa_mem_[MAX_NUMNODES];
97 /* work_structs for global per-cpu drains */
98 DEFINE_MUTEX(pcpu_drain_mutex);
99 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
101 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
102 volatile unsigned long latent_entropy __latent_entropy;
103 EXPORT_SYMBOL(latent_entropy);
107 * Array of node states.
109 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
110 [N_POSSIBLE] = NODE_MASK_ALL,
111 [N_ONLINE] = { { [0] = 1UL } },
113 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
114 #ifdef CONFIG_HIGHMEM
115 [N_HIGH_MEMORY] = { { [0] = 1UL } },
117 [N_MEMORY] = { { [0] = 1UL } },
118 [N_CPU] = { { [0] = 1UL } },
121 EXPORT_SYMBOL(node_states);
123 /* Protect totalram_pages and zone->managed_pages */
124 static DEFINE_SPINLOCK(managed_page_count_lock);
126 unsigned long totalram_pages __read_mostly;
127 unsigned long totalreserve_pages __read_mostly;
128 unsigned long totalcma_pages __read_mostly;
130 int percpu_pagelist_fraction;
131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page *page)
146 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
148 page->index = migratetype;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with pm_mutex held (gfp_allowed_mask also should
157 * only be modified with pm_mutex held, unless the suspend/hibernate code is
158 * guaranteed not to run in parallel with that modification).
161 static gfp_t saved_gfp_mask;
163 void pm_restore_gfp_mask(void)
165 WARN_ON(!mutex_is_locked(&pm_mutex));
166 if (saved_gfp_mask) {
167 gfp_allowed_mask = saved_gfp_mask;
172 void pm_restrict_gfp_mask(void)
174 WARN_ON(!mutex_is_locked(&pm_mutex));
175 WARN_ON(saved_gfp_mask);
176 saved_gfp_mask = gfp_allowed_mask;
177 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
180 bool pm_suspended_storage(void)
182 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
186 #endif /* CONFIG_PM_SLEEP */
188 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
189 unsigned int pageblock_order __read_mostly;
192 static void __free_pages_ok(struct page *page, unsigned int order);
195 * results with 256, 32 in the lowmem_reserve sysctl:
196 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
197 * 1G machine -> (16M dma, 784M normal, 224M high)
198 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
199 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
200 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
202 * TBD: should special case ZONE_DMA32 machines here - in those we normally
203 * don't need any ZONE_NORMAL reservation
205 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
206 #ifdef CONFIG_ZONE_DMA
209 #ifdef CONFIG_ZONE_DMA32
212 #ifdef CONFIG_HIGHMEM
218 EXPORT_SYMBOL(totalram_pages);
220 static char * const zone_names[MAX_NR_ZONES] = {
221 #ifdef CONFIG_ZONE_DMA
224 #ifdef CONFIG_ZONE_DMA32
228 #ifdef CONFIG_HIGHMEM
232 #ifdef CONFIG_ZONE_DEVICE
237 char * const migratetype_names[MIGRATE_TYPES] = {
245 #ifdef CONFIG_MEMORY_ISOLATION
250 compound_page_dtor * const compound_page_dtors[] = {
253 #ifdef CONFIG_HUGETLB_PAGE
256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
261 int min_free_kbytes = 1024;
262 int user_min_free_kbytes = -1;
263 int watermark_scale_factor = 10;
265 static unsigned long __meminitdata nr_kernel_pages;
266 static unsigned long __meminitdata nr_all_pages;
267 static unsigned long __meminitdata dma_reserve;
269 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
270 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
271 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
272 static unsigned long __initdata required_kernelcore;
273 static unsigned long __initdata required_movablecore;
274 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
275 static bool mirrored_kernelcore;
277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
279 EXPORT_SYMBOL(movable_zone);
280 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
283 int nr_node_ids __read_mostly = MAX_NUMNODES;
284 int nr_online_nodes __read_mostly = 1;
285 EXPORT_SYMBOL(nr_node_ids);
286 EXPORT_SYMBOL(nr_online_nodes);
289 int page_group_by_mobility_disabled __read_mostly;
291 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
292 static inline void reset_deferred_meminit(pg_data_t *pgdat)
294 unsigned long max_initialise;
295 unsigned long reserved_lowmem;
298 * Initialise at least 2G of a node but also take into account that
299 * two large system hashes that can take up 1GB for 0.25TB/node.
301 max_initialise = max(2UL << (30 - PAGE_SHIFT),
302 (pgdat->node_spanned_pages >> 8));
305 * Compensate the all the memblock reservations (e.g. crash kernel)
306 * from the initial estimation to make sure we will initialize enough
309 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
310 pgdat->node_start_pfn + max_initialise);
311 max_initialise += reserved_lowmem;
313 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
314 pgdat->first_deferred_pfn = ULONG_MAX;
317 /* Returns true if the struct page for the pfn is uninitialised */
318 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
320 int nid = early_pfn_to_nid(pfn);
322 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
329 * Returns false when the remaining initialisation should be deferred until
330 * later in the boot cycle when it can be parallelised.
332 static inline bool update_defer_init(pg_data_t *pgdat,
333 unsigned long pfn, unsigned long zone_end,
334 unsigned long *nr_initialised)
336 /* Always populate low zones for address-contrained allocations */
337 if (zone_end < pgdat_end_pfn(pgdat))
340 if ((*nr_initialised > pgdat->static_init_size) &&
341 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
342 pgdat->first_deferred_pfn = pfn;
349 static inline void reset_deferred_meminit(pg_data_t *pgdat)
353 static inline bool early_page_uninitialised(unsigned long pfn)
358 static inline bool update_defer_init(pg_data_t *pgdat,
359 unsigned long pfn, unsigned long zone_end,
360 unsigned long *nr_initialised)
366 /* Return a pointer to the bitmap storing bits affecting a block of pages */
367 static inline unsigned long *get_pageblock_bitmap(struct page *page,
370 #ifdef CONFIG_SPARSEMEM
371 return __pfn_to_section(pfn)->pageblock_flags;
373 return page_zone(page)->pageblock_flags;
374 #endif /* CONFIG_SPARSEMEM */
377 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
379 #ifdef CONFIG_SPARSEMEM
380 pfn &= (PAGES_PER_SECTION-1);
381 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
383 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
384 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
385 #endif /* CONFIG_SPARSEMEM */
389 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
390 * @page: The page within the block of interest
391 * @pfn: The target page frame number
392 * @end_bitidx: The last bit of interest to retrieve
393 * @mask: mask of bits that the caller is interested in
395 * Return: pageblock_bits flags
397 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
399 unsigned long end_bitidx,
402 unsigned long *bitmap;
403 unsigned long bitidx, word_bitidx;
406 bitmap = get_pageblock_bitmap(page, pfn);
407 bitidx = pfn_to_bitidx(page, pfn);
408 word_bitidx = bitidx / BITS_PER_LONG;
409 bitidx &= (BITS_PER_LONG-1);
411 word = bitmap[word_bitidx];
412 bitidx += end_bitidx;
413 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
416 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
417 unsigned long end_bitidx,
420 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
423 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
425 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
429 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
430 * @page: The page within the block of interest
431 * @flags: The flags to set
432 * @pfn: The target page frame number
433 * @end_bitidx: The last bit of interest
434 * @mask: mask of bits that the caller is interested in
436 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
438 unsigned long end_bitidx,
441 unsigned long *bitmap;
442 unsigned long bitidx, word_bitidx;
443 unsigned long old_word, word;
445 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
447 bitmap = get_pageblock_bitmap(page, pfn);
448 bitidx = pfn_to_bitidx(page, pfn);
449 word_bitidx = bitidx / BITS_PER_LONG;
450 bitidx &= (BITS_PER_LONG-1);
452 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
454 bitidx += end_bitidx;
455 mask <<= (BITS_PER_LONG - bitidx - 1);
456 flags <<= (BITS_PER_LONG - bitidx - 1);
458 word = READ_ONCE(bitmap[word_bitidx]);
460 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
461 if (word == old_word)
467 void set_pageblock_migratetype(struct page *page, int migratetype)
469 if (unlikely(page_group_by_mobility_disabled &&
470 migratetype < MIGRATE_PCPTYPES))
471 migratetype = MIGRATE_UNMOVABLE;
473 set_pageblock_flags_group(page, (unsigned long)migratetype,
474 PB_migrate, PB_migrate_end);
477 #ifdef CONFIG_DEBUG_VM
478 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
482 unsigned long pfn = page_to_pfn(page);
483 unsigned long sp, start_pfn;
486 seq = zone_span_seqbegin(zone);
487 start_pfn = zone->zone_start_pfn;
488 sp = zone->spanned_pages;
489 if (!zone_spans_pfn(zone, pfn))
491 } while (zone_span_seqretry(zone, seq));
494 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
495 pfn, zone_to_nid(zone), zone->name,
496 start_pfn, start_pfn + sp);
501 static int page_is_consistent(struct zone *zone, struct page *page)
503 if (!pfn_valid_within(page_to_pfn(page)))
505 if (zone != page_zone(page))
511 * Temporary debugging check for pages not lying within a given zone.
513 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
515 if (page_outside_zone_boundaries(zone, page))
517 if (!page_is_consistent(zone, page))
523 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
529 static void bad_page(struct page *page, const char *reason,
530 unsigned long bad_flags)
532 static unsigned long resume;
533 static unsigned long nr_shown;
534 static unsigned long nr_unshown;
537 * Allow a burst of 60 reports, then keep quiet for that minute;
538 * or allow a steady drip of one report per second.
540 if (nr_shown == 60) {
541 if (time_before(jiffies, resume)) {
547 "BUG: Bad page state: %lu messages suppressed\n",
554 resume = jiffies + 60 * HZ;
556 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
557 current->comm, page_to_pfn(page));
558 __dump_page(page, reason);
559 bad_flags &= page->flags;
561 pr_alert("bad because of flags: %#lx(%pGp)\n",
562 bad_flags, &bad_flags);
563 dump_page_owner(page);
568 /* Leave bad fields for debug, except PageBuddy could make trouble */
569 page_mapcount_reset(page); /* remove PageBuddy */
570 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
574 * Higher-order pages are called "compound pages". They are structured thusly:
576 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
578 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
579 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
581 * The first tail page's ->compound_dtor holds the offset in array of compound
582 * page destructors. See compound_page_dtors.
584 * The first tail page's ->compound_order holds the order of allocation.
585 * This usage means that zero-order pages may not be compound.
588 void free_compound_page(struct page *page)
590 __free_pages_ok(page, compound_order(page));
593 void prep_compound_page(struct page *page, unsigned int order)
596 int nr_pages = 1 << order;
598 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
599 set_compound_order(page, order);
601 for (i = 1; i < nr_pages; i++) {
602 struct page *p = page + i;
603 set_page_count(p, 0);
604 p->mapping = TAIL_MAPPING;
605 set_compound_head(p, page);
607 atomic_set(compound_mapcount_ptr(page), -1);
610 #ifdef CONFIG_DEBUG_PAGEALLOC
611 unsigned int _debug_guardpage_minorder;
612 bool _debug_pagealloc_enabled __read_mostly
613 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
614 EXPORT_SYMBOL(_debug_pagealloc_enabled);
615 bool _debug_guardpage_enabled __read_mostly;
617 static int __init early_debug_pagealloc(char *buf)
621 return kstrtobool(buf, &_debug_pagealloc_enabled);
623 early_param("debug_pagealloc", early_debug_pagealloc);
625 static bool need_debug_guardpage(void)
627 /* If we don't use debug_pagealloc, we don't need guard page */
628 if (!debug_pagealloc_enabled())
631 if (!debug_guardpage_minorder())
637 static void init_debug_guardpage(void)
639 if (!debug_pagealloc_enabled())
642 if (!debug_guardpage_minorder())
645 _debug_guardpage_enabled = true;
648 struct page_ext_operations debug_guardpage_ops = {
649 .need = need_debug_guardpage,
650 .init = init_debug_guardpage,
653 static int __init debug_guardpage_minorder_setup(char *buf)
657 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
658 pr_err("Bad debug_guardpage_minorder value\n");
661 _debug_guardpage_minorder = res;
662 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
665 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
667 static inline bool set_page_guard(struct zone *zone, struct page *page,
668 unsigned int order, int migratetype)
670 struct page_ext *page_ext;
672 if (!debug_guardpage_enabled())
675 if (order >= debug_guardpage_minorder())
678 page_ext = lookup_page_ext(page);
679 if (unlikely(!page_ext))
682 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
684 INIT_LIST_HEAD(&page->lru);
685 set_page_private(page, order);
686 /* Guard pages are not available for any usage */
687 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
692 static inline void clear_page_guard(struct zone *zone, struct page *page,
693 unsigned int order, int migratetype)
695 struct page_ext *page_ext;
697 if (!debug_guardpage_enabled())
700 page_ext = lookup_page_ext(page);
701 if (unlikely(!page_ext))
704 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
706 set_page_private(page, 0);
707 if (!is_migrate_isolate(migratetype))
708 __mod_zone_freepage_state(zone, (1 << order), migratetype);
711 struct page_ext_operations debug_guardpage_ops;
712 static inline bool set_page_guard(struct zone *zone, struct page *page,
713 unsigned int order, int migratetype) { return false; }
714 static inline void clear_page_guard(struct zone *zone, struct page *page,
715 unsigned int order, int migratetype) {}
718 static inline void set_page_order(struct page *page, unsigned int order)
720 set_page_private(page, order);
721 __SetPageBuddy(page);
724 static inline void rmv_page_order(struct page *page)
726 __ClearPageBuddy(page);
727 set_page_private(page, 0);
731 * This function checks whether a page is free && is the buddy
732 * we can do coalesce a page and its buddy if
733 * (a) the buddy is not in a hole (check before calling!) &&
734 * (b) the buddy is in the buddy system &&
735 * (c) a page and its buddy have the same order &&
736 * (d) a page and its buddy are in the same zone.
738 * For recording whether a page is in the buddy system, we set ->_mapcount
739 * PAGE_BUDDY_MAPCOUNT_VALUE.
740 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
741 * serialized by zone->lock.
743 * For recording page's order, we use page_private(page).
745 static inline int page_is_buddy(struct page *page, struct page *buddy,
748 if (page_is_guard(buddy) && page_order(buddy) == order) {
749 if (page_zone_id(page) != page_zone_id(buddy))
752 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
757 if (PageBuddy(buddy) && page_order(buddy) == order) {
759 * zone check is done late to avoid uselessly
760 * calculating zone/node ids for pages that could
763 if (page_zone_id(page) != page_zone_id(buddy))
766 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
774 * Freeing function for a buddy system allocator.
776 * The concept of a buddy system is to maintain direct-mapped table
777 * (containing bit values) for memory blocks of various "orders".
778 * The bottom level table contains the map for the smallest allocatable
779 * units of memory (here, pages), and each level above it describes
780 * pairs of units from the levels below, hence, "buddies".
781 * At a high level, all that happens here is marking the table entry
782 * at the bottom level available, and propagating the changes upward
783 * as necessary, plus some accounting needed to play nicely with other
784 * parts of the VM system.
785 * At each level, we keep a list of pages, which are heads of continuous
786 * free pages of length of (1 << order) and marked with _mapcount
787 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
789 * So when we are allocating or freeing one, we can derive the state of the
790 * other. That is, if we allocate a small block, and both were
791 * free, the remainder of the region must be split into blocks.
792 * If a block is freed, and its buddy is also free, then this
793 * triggers coalescing into a block of larger size.
798 static inline void __free_one_page(struct page *page,
800 struct zone *zone, unsigned int order,
803 unsigned long combined_pfn;
804 unsigned long uninitialized_var(buddy_pfn);
806 unsigned int max_order;
808 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
810 VM_BUG_ON(!zone_is_initialized(zone));
811 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
813 VM_BUG_ON(migratetype == -1);
814 if (likely(!is_migrate_isolate(migratetype)))
815 __mod_zone_freepage_state(zone, 1 << order, migratetype);
817 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
818 VM_BUG_ON_PAGE(bad_range(zone, page), page);
821 while (order < max_order - 1) {
822 buddy_pfn = __find_buddy_pfn(pfn, order);
823 buddy = page + (buddy_pfn - pfn);
825 if (!pfn_valid_within(buddy_pfn))
827 if (!page_is_buddy(page, buddy, order))
830 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
831 * merge with it and move up one order.
833 if (page_is_guard(buddy)) {
834 clear_page_guard(zone, buddy, order, migratetype);
836 list_del(&buddy->lru);
837 zone->free_area[order].nr_free--;
838 rmv_page_order(buddy);
840 combined_pfn = buddy_pfn & pfn;
841 page = page + (combined_pfn - pfn);
845 if (max_order < MAX_ORDER) {
846 /* If we are here, it means order is >= pageblock_order.
847 * We want to prevent merge between freepages on isolate
848 * pageblock and normal pageblock. Without this, pageblock
849 * isolation could cause incorrect freepage or CMA accounting.
851 * We don't want to hit this code for the more frequent
854 if (unlikely(has_isolate_pageblock(zone))) {
857 buddy_pfn = __find_buddy_pfn(pfn, order);
858 buddy = page + (buddy_pfn - pfn);
859 buddy_mt = get_pageblock_migratetype(buddy);
861 if (migratetype != buddy_mt
862 && (is_migrate_isolate(migratetype) ||
863 is_migrate_isolate(buddy_mt)))
867 goto continue_merging;
871 set_page_order(page, order);
874 * If this is not the largest possible page, check if the buddy
875 * of the next-highest order is free. If it is, it's possible
876 * that pages are being freed that will coalesce soon. In case,
877 * that is happening, add the free page to the tail of the list
878 * so it's less likely to be used soon and more likely to be merged
879 * as a higher order page
881 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
882 struct page *higher_page, *higher_buddy;
883 combined_pfn = buddy_pfn & pfn;
884 higher_page = page + (combined_pfn - pfn);
885 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
886 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
887 if (pfn_valid_within(buddy_pfn) &&
888 page_is_buddy(higher_page, higher_buddy, order + 1)) {
889 list_add_tail(&page->lru,
890 &zone->free_area[order].free_list[migratetype]);
895 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
897 zone->free_area[order].nr_free++;
901 * A bad page could be due to a number of fields. Instead of multiple branches,
902 * try and check multiple fields with one check. The caller must do a detailed
903 * check if necessary.
905 static inline bool page_expected_state(struct page *page,
906 unsigned long check_flags)
908 if (unlikely(atomic_read(&page->_mapcount) != -1))
911 if (unlikely((unsigned long)page->mapping |
912 page_ref_count(page) |
914 (unsigned long)page->mem_cgroup |
916 (page->flags & check_flags)))
922 static void free_pages_check_bad(struct page *page)
924 const char *bad_reason;
925 unsigned long bad_flags;
930 if (unlikely(atomic_read(&page->_mapcount) != -1))
931 bad_reason = "nonzero mapcount";
932 if (unlikely(page->mapping != NULL))
933 bad_reason = "non-NULL mapping";
934 if (unlikely(page_ref_count(page) != 0))
935 bad_reason = "nonzero _refcount";
936 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
937 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
938 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
941 if (unlikely(page->mem_cgroup))
942 bad_reason = "page still charged to cgroup";
944 bad_page(page, bad_reason, bad_flags);
947 static inline int free_pages_check(struct page *page)
949 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
952 /* Something has gone sideways, find it */
953 free_pages_check_bad(page);
957 static int free_tail_pages_check(struct page *head_page, struct page *page)
962 * We rely page->lru.next never has bit 0 set, unless the page
963 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
965 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
967 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
971 switch (page - head_page) {
973 /* the first tail page: ->mapping is compound_mapcount() */
974 if (unlikely(compound_mapcount(page))) {
975 bad_page(page, "nonzero compound_mapcount", 0);
981 * the second tail page: ->mapping is
982 * page_deferred_list().next -- ignore value.
986 if (page->mapping != TAIL_MAPPING) {
987 bad_page(page, "corrupted mapping in tail page", 0);
992 if (unlikely(!PageTail(page))) {
993 bad_page(page, "PageTail not set", 0);
996 if (unlikely(compound_head(page) != head_page)) {
997 bad_page(page, "compound_head not consistent", 0);
1002 page->mapping = NULL;
1003 clear_compound_head(page);
1007 static __always_inline bool free_pages_prepare(struct page *page,
1008 unsigned int order, bool check_free)
1012 VM_BUG_ON_PAGE(PageTail(page), page);
1014 trace_mm_page_free(page, order);
1017 * Check tail pages before head page information is cleared to
1018 * avoid checking PageCompound for order-0 pages.
1020 if (unlikely(order)) {
1021 bool compound = PageCompound(page);
1024 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1027 ClearPageDoubleMap(page);
1028 for (i = 1; i < (1 << order); i++) {
1030 bad += free_tail_pages_check(page, page + i);
1031 if (unlikely(free_pages_check(page + i))) {
1035 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1038 if (PageMappingFlags(page))
1039 page->mapping = NULL;
1040 if (memcg_kmem_enabled() && PageKmemcg(page))
1041 memcg_kmem_uncharge(page, order);
1043 bad += free_pages_check(page);
1047 page_cpupid_reset_last(page);
1048 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 reset_page_owner(page, order);
1051 if (!PageHighMem(page)) {
1052 debug_check_no_locks_freed(page_address(page),
1053 PAGE_SIZE << order);
1054 debug_check_no_obj_freed(page_address(page),
1055 PAGE_SIZE << order);
1057 arch_free_page(page, order);
1058 kernel_poison_pages(page, 1 << order, 0);
1059 kernel_map_pages(page, 1 << order, 0);
1060 kasan_free_pages(page, order);
1065 #ifdef CONFIG_DEBUG_VM
1066 static inline bool free_pcp_prepare(struct page *page)
1068 return free_pages_prepare(page, 0, true);
1071 static inline bool bulkfree_pcp_prepare(struct page *page)
1076 static bool free_pcp_prepare(struct page *page)
1078 return free_pages_prepare(page, 0, false);
1081 static bool bulkfree_pcp_prepare(struct page *page)
1083 return free_pages_check(page);
1085 #endif /* CONFIG_DEBUG_VM */
1088 * Frees a number of pages from the PCP lists
1089 * Assumes all pages on list are in same zone, and of same order.
1090 * count is the number of pages to free.
1092 * If the zone was previously in an "all pages pinned" state then look to
1093 * see if this freeing clears that state.
1095 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1096 * pinned" detection logic.
1098 static void free_pcppages_bulk(struct zone *zone, int count,
1099 struct per_cpu_pages *pcp)
1101 int migratetype = 0;
1103 bool isolated_pageblocks;
1105 spin_lock(&zone->lock);
1106 isolated_pageblocks = has_isolate_pageblock(zone);
1110 struct list_head *list;
1113 * Remove pages from lists in a round-robin fashion. A
1114 * batch_free count is maintained that is incremented when an
1115 * empty list is encountered. This is so more pages are freed
1116 * off fuller lists instead of spinning excessively around empty
1121 if (++migratetype == MIGRATE_PCPTYPES)
1123 list = &pcp->lists[migratetype];
1124 } while (list_empty(list));
1126 /* This is the only non-empty list. Free them all. */
1127 if (batch_free == MIGRATE_PCPTYPES)
1131 int mt; /* migratetype of the to-be-freed page */
1133 page = list_last_entry(list, struct page, lru);
1134 /* must delete as __free_one_page list manipulates */
1135 list_del(&page->lru);
1137 mt = get_pcppage_migratetype(page);
1138 /* MIGRATE_ISOLATE page should not go to pcplists */
1139 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1140 /* Pageblock could have been isolated meanwhile */
1141 if (unlikely(isolated_pageblocks))
1142 mt = get_pageblock_migratetype(page);
1144 if (bulkfree_pcp_prepare(page))
1147 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1148 trace_mm_page_pcpu_drain(page, 0, mt);
1149 } while (--count && --batch_free && !list_empty(list));
1151 spin_unlock(&zone->lock);
1154 static void free_one_page(struct zone *zone,
1155 struct page *page, unsigned long pfn,
1159 spin_lock(&zone->lock);
1160 if (unlikely(has_isolate_pageblock(zone) ||
1161 is_migrate_isolate(migratetype))) {
1162 migratetype = get_pfnblock_migratetype(page, pfn);
1164 __free_one_page(page, pfn, zone, order, migratetype);
1165 spin_unlock(&zone->lock);
1168 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1169 unsigned long zone, int nid)
1171 mm_zero_struct_page(page);
1172 set_page_links(page, zone, nid, pfn);
1173 init_page_count(page);
1174 page_mapcount_reset(page);
1175 page_cpupid_reset_last(page);
1177 INIT_LIST_HEAD(&page->lru);
1178 #ifdef WANT_PAGE_VIRTUAL
1179 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1180 if (!is_highmem_idx(zone))
1181 set_page_address(page, __va(pfn << PAGE_SHIFT));
1185 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1188 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1191 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1192 static void __meminit init_reserved_page(unsigned long pfn)
1197 if (!early_page_uninitialised(pfn))
1200 nid = early_pfn_to_nid(pfn);
1201 pgdat = NODE_DATA(nid);
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 struct zone *zone = &pgdat->node_zones[zid];
1206 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1209 __init_single_pfn(pfn, zid, nid);
1212 static inline void init_reserved_page(unsigned long pfn)
1215 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1218 * Initialised pages do not have PageReserved set. This function is
1219 * called for each range allocated by the bootmem allocator and
1220 * marks the pages PageReserved. The remaining valid pages are later
1221 * sent to the buddy page allocator.
1223 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1225 unsigned long start_pfn = PFN_DOWN(start);
1226 unsigned long end_pfn = PFN_UP(end);
1228 for (; start_pfn < end_pfn; start_pfn++) {
1229 if (pfn_valid(start_pfn)) {
1230 struct page *page = pfn_to_page(start_pfn);
1232 init_reserved_page(start_pfn);
1234 /* Avoid false-positive PageTail() */
1235 INIT_LIST_HEAD(&page->lru);
1237 SetPageReserved(page);
1242 static void __free_pages_ok(struct page *page, unsigned int order)
1244 unsigned long flags;
1246 unsigned long pfn = page_to_pfn(page);
1248 if (!free_pages_prepare(page, order, true))
1251 migratetype = get_pfnblock_migratetype(page, pfn);
1252 local_irq_save(flags);
1253 __count_vm_events(PGFREE, 1 << order);
1254 free_one_page(page_zone(page), page, pfn, order, migratetype);
1255 local_irq_restore(flags);
1258 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1260 unsigned int nr_pages = 1 << order;
1261 struct page *p = page;
1265 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1267 __ClearPageReserved(p);
1268 set_page_count(p, 0);
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1273 page_zone(page)->managed_pages += nr_pages;
1274 set_page_refcounted(page);
1275 __free_pages(page, order);
1278 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1279 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1281 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1283 int __meminit early_pfn_to_nid(unsigned long pfn)
1285 static DEFINE_SPINLOCK(early_pfn_lock);
1288 spin_lock(&early_pfn_lock);
1289 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1291 nid = first_online_node;
1292 spin_unlock(&early_pfn_lock);
1298 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1299 static inline bool __meminit __maybe_unused
1300 meminit_pfn_in_nid(unsigned long pfn, int node,
1301 struct mminit_pfnnid_cache *state)
1305 nid = __early_pfn_to_nid(pfn, state);
1306 if (nid >= 0 && nid != node)
1311 /* Only safe to use early in boot when initialisation is single-threaded */
1312 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1319 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1323 static inline bool __meminit __maybe_unused
1324 meminit_pfn_in_nid(unsigned long pfn, int node,
1325 struct mminit_pfnnid_cache *state)
1332 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1335 if (early_page_uninitialised(pfn))
1337 return __free_pages_boot_core(page, order);
1341 * Check that the whole (or subset of) a pageblock given by the interval of
1342 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1343 * with the migration of free compaction scanner. The scanners then need to
1344 * use only pfn_valid_within() check for arches that allow holes within
1347 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1349 * It's possible on some configurations to have a setup like node0 node1 node0
1350 * i.e. it's possible that all pages within a zones range of pages do not
1351 * belong to a single zone. We assume that a border between node0 and node1
1352 * can occur within a single pageblock, but not a node0 node1 node0
1353 * interleaving within a single pageblock. It is therefore sufficient to check
1354 * the first and last page of a pageblock and avoid checking each individual
1355 * page in a pageblock.
1357 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1358 unsigned long end_pfn, struct zone *zone)
1360 struct page *start_page;
1361 struct page *end_page;
1363 /* end_pfn is one past the range we are checking */
1366 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1369 start_page = pfn_to_online_page(start_pfn);
1373 if (page_zone(start_page) != zone)
1376 end_page = pfn_to_page(end_pfn);
1378 /* This gives a shorter code than deriving page_zone(end_page) */
1379 if (page_zone_id(start_page) != page_zone_id(end_page))
1385 void set_zone_contiguous(struct zone *zone)
1387 unsigned long block_start_pfn = zone->zone_start_pfn;
1388 unsigned long block_end_pfn;
1390 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1391 for (; block_start_pfn < zone_end_pfn(zone);
1392 block_start_pfn = block_end_pfn,
1393 block_end_pfn += pageblock_nr_pages) {
1395 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1397 if (!__pageblock_pfn_to_page(block_start_pfn,
1398 block_end_pfn, zone))
1402 /* We confirm that there is no hole */
1403 zone->contiguous = true;
1406 void clear_zone_contiguous(struct zone *zone)
1408 zone->contiguous = false;
1411 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1412 static void __init deferred_free_range(unsigned long pfn,
1413 unsigned long nr_pages)
1421 page = pfn_to_page(pfn);
1423 /* Free a large naturally-aligned chunk if possible */
1424 if (nr_pages == pageblock_nr_pages &&
1425 (pfn & (pageblock_nr_pages - 1)) == 0) {
1426 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1427 __free_pages_boot_core(page, pageblock_order);
1431 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1432 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1433 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1434 __free_pages_boot_core(page, 0);
1438 /* Completion tracking for deferred_init_memmap() threads */
1439 static atomic_t pgdat_init_n_undone __initdata;
1440 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1442 static inline void __init pgdat_init_report_one_done(void)
1444 if (atomic_dec_and_test(&pgdat_init_n_undone))
1445 complete(&pgdat_init_all_done_comp);
1449 * Helper for deferred_init_range, free the given range, reset the counters, and
1450 * return number of pages freed.
1452 static inline unsigned long __init __def_free(unsigned long *nr_free,
1453 unsigned long *free_base_pfn,
1456 unsigned long nr = *nr_free;
1458 deferred_free_range(*free_base_pfn, nr);
1466 static unsigned long __init deferred_init_range(int nid, int zid,
1467 unsigned long start_pfn,
1468 unsigned long end_pfn)
1470 struct mminit_pfnnid_cache nid_init_state = { };
1471 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1472 unsigned long free_base_pfn = 0;
1473 unsigned long nr_pages = 0;
1474 unsigned long nr_free = 0;
1475 struct page *page = NULL;
1479 * First we check if pfn is valid on architectures where it is possible
1480 * to have holes within pageblock_nr_pages. On systems where it is not
1481 * possible, this function is optimized out.
1483 * Then, we check if a current large page is valid by only checking the
1484 * validity of the head pfn.
1486 * meminit_pfn_in_nid is checked on systems where pfns can interleave
1487 * within a node: a pfn is between start and end of a node, but does not
1488 * belong to this memory node.
1490 * Finally, we minimize pfn page lookups and scheduler checks by
1491 * performing it only once every pageblock_nr_pages.
1493 * We do it in two loops: first we initialize struct page, than free to
1494 * buddy allocator, becuse while we are freeing pages we can access
1495 * pages that are ahead (computing buddy page in __free_one_page()).
1497 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1498 if (!pfn_valid_within(pfn))
1500 if ((pfn & nr_pgmask) || pfn_valid(pfn)) {
1501 if (meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1502 if (page && (pfn & nr_pgmask))
1505 page = pfn_to_page(pfn);
1506 __init_single_page(page, pfn, zid, nid);
1513 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1514 if (!pfn_valid_within(pfn)) {
1515 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1516 } else if (!(pfn & nr_pgmask) && !pfn_valid(pfn)) {
1517 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1518 } else if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1519 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1520 } else if (page && (pfn & nr_pgmask)) {
1524 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1525 page = pfn_to_page(pfn);
1526 free_base_pfn = pfn;
1531 /* Free the last block of pages to allocator */
1532 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1537 /* Initialise remaining memory on a node */
1538 static int __init deferred_init_memmap(void *data)
1540 pg_data_t *pgdat = data;
1541 int nid = pgdat->node_id;
1542 unsigned long start = jiffies;
1543 unsigned long nr_pages = 0;
1544 unsigned long spfn, epfn;
1545 phys_addr_t spa, epa;
1548 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1549 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1552 if (first_init_pfn == ULONG_MAX) {
1553 pgdat_init_report_one_done();
1557 /* Bind memory initialisation thread to a local node if possible */
1558 if (!cpumask_empty(cpumask))
1559 set_cpus_allowed_ptr(current, cpumask);
1561 /* Sanity check boundaries */
1562 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1563 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1564 pgdat->first_deferred_pfn = ULONG_MAX;
1566 /* Only the highest zone is deferred so find it */
1567 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1568 zone = pgdat->node_zones + zid;
1569 if (first_init_pfn < zone_end_pfn(zone))
1572 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1574 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1575 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1576 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1577 nr_pages += deferred_init_range(nid, zid, spfn, epfn);
1580 /* Sanity check that the next zone really is unpopulated */
1581 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1583 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1584 jiffies_to_msecs(jiffies - start));
1586 pgdat_init_report_one_done();
1589 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1591 void __init page_alloc_init_late(void)
1595 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1598 /* There will be num_node_state(N_MEMORY) threads */
1599 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1600 for_each_node_state(nid, N_MEMORY) {
1601 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1604 /* Block until all are initialised */
1605 wait_for_completion(&pgdat_init_all_done_comp);
1607 /* Reinit limits that are based on free pages after the kernel is up */
1608 files_maxfiles_init();
1610 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1611 /* Discard memblock private memory */
1615 for_each_populated_zone(zone)
1616 set_zone_contiguous(zone);
1620 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1621 void __init init_cma_reserved_pageblock(struct page *page)
1623 unsigned i = pageblock_nr_pages;
1624 struct page *p = page;
1627 __ClearPageReserved(p);
1628 set_page_count(p, 0);
1631 set_pageblock_migratetype(page, MIGRATE_CMA);
1633 if (pageblock_order >= MAX_ORDER) {
1634 i = pageblock_nr_pages;
1637 set_page_refcounted(p);
1638 __free_pages(p, MAX_ORDER - 1);
1639 p += MAX_ORDER_NR_PAGES;
1640 } while (i -= MAX_ORDER_NR_PAGES);
1642 set_page_refcounted(page);
1643 __free_pages(page, pageblock_order);
1646 adjust_managed_page_count(page, pageblock_nr_pages);
1651 * The order of subdivision here is critical for the IO subsystem.
1652 * Please do not alter this order without good reasons and regression
1653 * testing. Specifically, as large blocks of memory are subdivided,
1654 * the order in which smaller blocks are delivered depends on the order
1655 * they're subdivided in this function. This is the primary factor
1656 * influencing the order in which pages are delivered to the IO
1657 * subsystem according to empirical testing, and this is also justified
1658 * by considering the behavior of a buddy system containing a single
1659 * large block of memory acted on by a series of small allocations.
1660 * This behavior is a critical factor in sglist merging's success.
1664 static inline void expand(struct zone *zone, struct page *page,
1665 int low, int high, struct free_area *area,
1668 unsigned long size = 1 << high;
1670 while (high > low) {
1674 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1677 * Mark as guard pages (or page), that will allow to
1678 * merge back to allocator when buddy will be freed.
1679 * Corresponding page table entries will not be touched,
1680 * pages will stay not present in virtual address space
1682 if (set_page_guard(zone, &page[size], high, migratetype))
1685 list_add(&page[size].lru, &area->free_list[migratetype]);
1687 set_page_order(&page[size], high);
1691 static void check_new_page_bad(struct page *page)
1693 const char *bad_reason = NULL;
1694 unsigned long bad_flags = 0;
1696 if (unlikely(atomic_read(&page->_mapcount) != -1))
1697 bad_reason = "nonzero mapcount";
1698 if (unlikely(page->mapping != NULL))
1699 bad_reason = "non-NULL mapping";
1700 if (unlikely(page_ref_count(page) != 0))
1701 bad_reason = "nonzero _count";
1702 if (unlikely(page->flags & __PG_HWPOISON)) {
1703 bad_reason = "HWPoisoned (hardware-corrupted)";
1704 bad_flags = __PG_HWPOISON;
1705 /* Don't complain about hwpoisoned pages */
1706 page_mapcount_reset(page); /* remove PageBuddy */
1709 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1710 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1711 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1714 if (unlikely(page->mem_cgroup))
1715 bad_reason = "page still charged to cgroup";
1717 bad_page(page, bad_reason, bad_flags);
1721 * This page is about to be returned from the page allocator
1723 static inline int check_new_page(struct page *page)
1725 if (likely(page_expected_state(page,
1726 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1729 check_new_page_bad(page);
1733 static inline bool free_pages_prezeroed(void)
1735 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1736 page_poisoning_enabled();
1739 #ifdef CONFIG_DEBUG_VM
1740 static bool check_pcp_refill(struct page *page)
1745 static bool check_new_pcp(struct page *page)
1747 return check_new_page(page);
1750 static bool check_pcp_refill(struct page *page)
1752 return check_new_page(page);
1754 static bool check_new_pcp(struct page *page)
1758 #endif /* CONFIG_DEBUG_VM */
1760 static bool check_new_pages(struct page *page, unsigned int order)
1763 for (i = 0; i < (1 << order); i++) {
1764 struct page *p = page + i;
1766 if (unlikely(check_new_page(p)))
1773 inline void post_alloc_hook(struct page *page, unsigned int order,
1776 set_page_private(page, 0);
1777 set_page_refcounted(page);
1779 arch_alloc_page(page, order);
1780 kernel_map_pages(page, 1 << order, 1);
1781 kernel_poison_pages(page, 1 << order, 1);
1782 kasan_alloc_pages(page, order);
1783 set_page_owner(page, order, gfp_flags);
1786 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1787 unsigned int alloc_flags)
1791 post_alloc_hook(page, order, gfp_flags);
1793 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1794 for (i = 0; i < (1 << order); i++)
1795 clear_highpage(page + i);
1797 if (order && (gfp_flags & __GFP_COMP))
1798 prep_compound_page(page, order);
1801 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1802 * allocate the page. The expectation is that the caller is taking
1803 * steps that will free more memory. The caller should avoid the page
1804 * being used for !PFMEMALLOC purposes.
1806 if (alloc_flags & ALLOC_NO_WATERMARKS)
1807 set_page_pfmemalloc(page);
1809 clear_page_pfmemalloc(page);
1813 * Go through the free lists for the given migratetype and remove
1814 * the smallest available page from the freelists
1816 static __always_inline
1817 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1820 unsigned int current_order;
1821 struct free_area *area;
1824 /* Find a page of the appropriate size in the preferred list */
1825 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1826 area = &(zone->free_area[current_order]);
1827 page = list_first_entry_or_null(&area->free_list[migratetype],
1831 list_del(&page->lru);
1832 rmv_page_order(page);
1834 expand(zone, page, order, current_order, area, migratetype);
1835 set_pcppage_migratetype(page, migratetype);
1844 * This array describes the order lists are fallen back to when
1845 * the free lists for the desirable migrate type are depleted
1847 static int fallbacks[MIGRATE_TYPES][4] = {
1848 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1849 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1850 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1852 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1854 #ifdef CONFIG_MEMORY_ISOLATION
1855 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1860 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1863 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1866 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1867 unsigned int order) { return NULL; }
1871 * Move the free pages in a range to the free lists of the requested type.
1872 * Note that start_page and end_pages are not aligned on a pageblock
1873 * boundary. If alignment is required, use move_freepages_block()
1875 static int move_freepages(struct zone *zone,
1876 struct page *start_page, struct page *end_page,
1877 int migratetype, int *num_movable)
1881 int pages_moved = 0;
1883 #ifndef CONFIG_HOLES_IN_ZONE
1885 * page_zone is not safe to call in this context when
1886 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1887 * anyway as we check zone boundaries in move_freepages_block().
1888 * Remove at a later date when no bug reports exist related to
1889 * grouping pages by mobility
1891 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1897 for (page = start_page; page <= end_page;) {
1898 if (!pfn_valid_within(page_to_pfn(page))) {
1903 /* Make sure we are not inadvertently changing nodes */
1904 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1906 if (!PageBuddy(page)) {
1908 * We assume that pages that could be isolated for
1909 * migration are movable. But we don't actually try
1910 * isolating, as that would be expensive.
1913 (PageLRU(page) || __PageMovable(page)))
1920 order = page_order(page);
1921 list_move(&page->lru,
1922 &zone->free_area[order].free_list[migratetype]);
1924 pages_moved += 1 << order;
1930 int move_freepages_block(struct zone *zone, struct page *page,
1931 int migratetype, int *num_movable)
1933 unsigned long start_pfn, end_pfn;
1934 struct page *start_page, *end_page;
1936 start_pfn = page_to_pfn(page);
1937 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1938 start_page = pfn_to_page(start_pfn);
1939 end_page = start_page + pageblock_nr_pages - 1;
1940 end_pfn = start_pfn + pageblock_nr_pages - 1;
1942 /* Do not cross zone boundaries */
1943 if (!zone_spans_pfn(zone, start_pfn))
1945 if (!zone_spans_pfn(zone, end_pfn))
1948 return move_freepages(zone, start_page, end_page, migratetype,
1952 static void change_pageblock_range(struct page *pageblock_page,
1953 int start_order, int migratetype)
1955 int nr_pageblocks = 1 << (start_order - pageblock_order);
1957 while (nr_pageblocks--) {
1958 set_pageblock_migratetype(pageblock_page, migratetype);
1959 pageblock_page += pageblock_nr_pages;
1964 * When we are falling back to another migratetype during allocation, try to
1965 * steal extra free pages from the same pageblocks to satisfy further
1966 * allocations, instead of polluting multiple pageblocks.
1968 * If we are stealing a relatively large buddy page, it is likely there will
1969 * be more free pages in the pageblock, so try to steal them all. For
1970 * reclaimable and unmovable allocations, we steal regardless of page size,
1971 * as fragmentation caused by those allocations polluting movable pageblocks
1972 * is worse than movable allocations stealing from unmovable and reclaimable
1975 static bool can_steal_fallback(unsigned int order, int start_mt)
1978 * Leaving this order check is intended, although there is
1979 * relaxed order check in next check. The reason is that
1980 * we can actually steal whole pageblock if this condition met,
1981 * but, below check doesn't guarantee it and that is just heuristic
1982 * so could be changed anytime.
1984 if (order >= pageblock_order)
1987 if (order >= pageblock_order / 2 ||
1988 start_mt == MIGRATE_RECLAIMABLE ||
1989 start_mt == MIGRATE_UNMOVABLE ||
1990 page_group_by_mobility_disabled)
1997 * This function implements actual steal behaviour. If order is large enough,
1998 * we can steal whole pageblock. If not, we first move freepages in this
1999 * pageblock to our migratetype and determine how many already-allocated pages
2000 * are there in the pageblock with a compatible migratetype. If at least half
2001 * of pages are free or compatible, we can change migratetype of the pageblock
2002 * itself, so pages freed in the future will be put on the correct free list.
2004 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2005 int start_type, bool whole_block)
2007 unsigned int current_order = page_order(page);
2008 struct free_area *area;
2009 int free_pages, movable_pages, alike_pages;
2012 old_block_type = get_pageblock_migratetype(page);
2015 * This can happen due to races and we want to prevent broken
2016 * highatomic accounting.
2018 if (is_migrate_highatomic(old_block_type))
2021 /* Take ownership for orders >= pageblock_order */
2022 if (current_order >= pageblock_order) {
2023 change_pageblock_range(page, current_order, start_type);
2027 /* We are not allowed to try stealing from the whole block */
2031 free_pages = move_freepages_block(zone, page, start_type,
2034 * Determine how many pages are compatible with our allocation.
2035 * For movable allocation, it's the number of movable pages which
2036 * we just obtained. For other types it's a bit more tricky.
2038 if (start_type == MIGRATE_MOVABLE) {
2039 alike_pages = movable_pages;
2042 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2043 * to MOVABLE pageblock, consider all non-movable pages as
2044 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2045 * vice versa, be conservative since we can't distinguish the
2046 * exact migratetype of non-movable pages.
2048 if (old_block_type == MIGRATE_MOVABLE)
2049 alike_pages = pageblock_nr_pages
2050 - (free_pages + movable_pages);
2055 /* moving whole block can fail due to zone boundary conditions */
2060 * If a sufficient number of pages in the block are either free or of
2061 * comparable migratability as our allocation, claim the whole block.
2063 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2064 page_group_by_mobility_disabled)
2065 set_pageblock_migratetype(page, start_type);
2070 area = &zone->free_area[current_order];
2071 list_move(&page->lru, &area->free_list[start_type]);
2075 * Check whether there is a suitable fallback freepage with requested order.
2076 * If only_stealable is true, this function returns fallback_mt only if
2077 * we can steal other freepages all together. This would help to reduce
2078 * fragmentation due to mixed migratetype pages in one pageblock.
2080 int find_suitable_fallback(struct free_area *area, unsigned int order,
2081 int migratetype, bool only_stealable, bool *can_steal)
2086 if (area->nr_free == 0)
2091 fallback_mt = fallbacks[migratetype][i];
2092 if (fallback_mt == MIGRATE_TYPES)
2095 if (list_empty(&area->free_list[fallback_mt]))
2098 if (can_steal_fallback(order, migratetype))
2101 if (!only_stealable)
2112 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2113 * there are no empty page blocks that contain a page with a suitable order
2115 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2116 unsigned int alloc_order)
2119 unsigned long max_managed, flags;
2122 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2123 * Check is race-prone but harmless.
2125 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2126 if (zone->nr_reserved_highatomic >= max_managed)
2129 spin_lock_irqsave(&zone->lock, flags);
2131 /* Recheck the nr_reserved_highatomic limit under the lock */
2132 if (zone->nr_reserved_highatomic >= max_managed)
2136 mt = get_pageblock_migratetype(page);
2137 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2138 && !is_migrate_cma(mt)) {
2139 zone->nr_reserved_highatomic += pageblock_nr_pages;
2140 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2141 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2145 spin_unlock_irqrestore(&zone->lock, flags);
2149 * Used when an allocation is about to fail under memory pressure. This
2150 * potentially hurts the reliability of high-order allocations when under
2151 * intense memory pressure but failed atomic allocations should be easier
2152 * to recover from than an OOM.
2154 * If @force is true, try to unreserve a pageblock even though highatomic
2155 * pageblock is exhausted.
2157 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2160 struct zonelist *zonelist = ac->zonelist;
2161 unsigned long flags;
2168 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2171 * Preserve at least one pageblock unless memory pressure
2174 if (!force && zone->nr_reserved_highatomic <=
2178 spin_lock_irqsave(&zone->lock, flags);
2179 for (order = 0; order < MAX_ORDER; order++) {
2180 struct free_area *area = &(zone->free_area[order]);
2182 page = list_first_entry_or_null(
2183 &area->free_list[MIGRATE_HIGHATOMIC],
2189 * In page freeing path, migratetype change is racy so
2190 * we can counter several free pages in a pageblock
2191 * in this loop althoug we changed the pageblock type
2192 * from highatomic to ac->migratetype. So we should
2193 * adjust the count once.
2195 if (is_migrate_highatomic_page(page)) {
2197 * It should never happen but changes to
2198 * locking could inadvertently allow a per-cpu
2199 * drain to add pages to MIGRATE_HIGHATOMIC
2200 * while unreserving so be safe and watch for
2203 zone->nr_reserved_highatomic -= min(
2205 zone->nr_reserved_highatomic);
2209 * Convert to ac->migratetype and avoid the normal
2210 * pageblock stealing heuristics. Minimally, the caller
2211 * is doing the work and needs the pages. More
2212 * importantly, if the block was always converted to
2213 * MIGRATE_UNMOVABLE or another type then the number
2214 * of pageblocks that cannot be completely freed
2217 set_pageblock_migratetype(page, ac->migratetype);
2218 ret = move_freepages_block(zone, page, ac->migratetype,
2221 spin_unlock_irqrestore(&zone->lock, flags);
2225 spin_unlock_irqrestore(&zone->lock, flags);
2232 * Try finding a free buddy page on the fallback list and put it on the free
2233 * list of requested migratetype, possibly along with other pages from the same
2234 * block, depending on fragmentation avoidance heuristics. Returns true if
2235 * fallback was found so that __rmqueue_smallest() can grab it.
2237 * The use of signed ints for order and current_order is a deliberate
2238 * deviation from the rest of this file, to make the for loop
2239 * condition simpler.
2241 static __always_inline bool
2242 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2244 struct free_area *area;
2251 * Find the largest available free page in the other list. This roughly
2252 * approximates finding the pageblock with the most free pages, which
2253 * would be too costly to do exactly.
2255 for (current_order = MAX_ORDER - 1; current_order >= order;
2257 area = &(zone->free_area[current_order]);
2258 fallback_mt = find_suitable_fallback(area, current_order,
2259 start_migratetype, false, &can_steal);
2260 if (fallback_mt == -1)
2264 * We cannot steal all free pages from the pageblock and the
2265 * requested migratetype is movable. In that case it's better to
2266 * steal and split the smallest available page instead of the
2267 * largest available page, because even if the next movable
2268 * allocation falls back into a different pageblock than this
2269 * one, it won't cause permanent fragmentation.
2271 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2272 && current_order > order)
2281 for (current_order = order; current_order < MAX_ORDER;
2283 area = &(zone->free_area[current_order]);
2284 fallback_mt = find_suitable_fallback(area, current_order,
2285 start_migratetype, false, &can_steal);
2286 if (fallback_mt != -1)
2291 * This should not happen - we already found a suitable fallback
2292 * when looking for the largest page.
2294 VM_BUG_ON(current_order == MAX_ORDER);
2297 page = list_first_entry(&area->free_list[fallback_mt],
2300 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2302 trace_mm_page_alloc_extfrag(page, order, current_order,
2303 start_migratetype, fallback_mt);
2310 * Do the hard work of removing an element from the buddy allocator.
2311 * Call me with the zone->lock already held.
2313 static __always_inline struct page *
2314 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2319 page = __rmqueue_smallest(zone, order, migratetype);
2320 if (unlikely(!page)) {
2321 if (migratetype == MIGRATE_MOVABLE)
2322 page = __rmqueue_cma_fallback(zone, order);
2324 if (!page && __rmqueue_fallback(zone, order, migratetype))
2328 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2333 * Obtain a specified number of elements from the buddy allocator, all under
2334 * a single hold of the lock, for efficiency. Add them to the supplied list.
2335 * Returns the number of new pages which were placed at *list.
2337 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2338 unsigned long count, struct list_head *list,
2343 spin_lock(&zone->lock);
2344 for (i = 0; i < count; ++i) {
2345 struct page *page = __rmqueue(zone, order, migratetype);
2346 if (unlikely(page == NULL))
2349 if (unlikely(check_pcp_refill(page)))
2353 * Split buddy pages returned by expand() are received here in
2354 * physical page order. The page is added to the tail of
2355 * caller's list. From the callers perspective, the linked list
2356 * is ordered by page number under some conditions. This is
2357 * useful for IO devices that can forward direction from the
2358 * head, thus also in the physical page order. This is useful
2359 * for IO devices that can merge IO requests if the physical
2360 * pages are ordered properly.
2362 list_add_tail(&page->lru, list);
2364 if (is_migrate_cma(get_pcppage_migratetype(page)))
2365 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2370 * i pages were removed from the buddy list even if some leak due
2371 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2372 * on i. Do not confuse with 'alloced' which is the number of
2373 * pages added to the pcp list.
2375 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2376 spin_unlock(&zone->lock);
2382 * Called from the vmstat counter updater to drain pagesets of this
2383 * currently executing processor on remote nodes after they have
2386 * Note that this function must be called with the thread pinned to
2387 * a single processor.
2389 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2391 unsigned long flags;
2392 int to_drain, batch;
2394 local_irq_save(flags);
2395 batch = READ_ONCE(pcp->batch);
2396 to_drain = min(pcp->count, batch);
2398 free_pcppages_bulk(zone, to_drain, pcp);
2399 pcp->count -= to_drain;
2401 local_irq_restore(flags);
2406 * Drain pcplists of the indicated processor and zone.
2408 * The processor must either be the current processor and the
2409 * thread pinned to the current processor or a processor that
2412 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2414 unsigned long flags;
2415 struct per_cpu_pageset *pset;
2416 struct per_cpu_pages *pcp;
2418 local_irq_save(flags);
2419 pset = per_cpu_ptr(zone->pageset, cpu);
2423 free_pcppages_bulk(zone, pcp->count, pcp);
2426 local_irq_restore(flags);
2430 * Drain pcplists of all zones on the indicated processor.
2432 * The processor must either be the current processor and the
2433 * thread pinned to the current processor or a processor that
2436 static void drain_pages(unsigned int cpu)
2440 for_each_populated_zone(zone) {
2441 drain_pages_zone(cpu, zone);
2446 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2448 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2449 * the single zone's pages.
2451 void drain_local_pages(struct zone *zone)
2453 int cpu = smp_processor_id();
2456 drain_pages_zone(cpu, zone);
2461 static void drain_local_pages_wq(struct work_struct *work)
2464 * drain_all_pages doesn't use proper cpu hotplug protection so
2465 * we can race with cpu offline when the WQ can move this from
2466 * a cpu pinned worker to an unbound one. We can operate on a different
2467 * cpu which is allright but we also have to make sure to not move to
2471 drain_local_pages(NULL);
2476 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2478 * When zone parameter is non-NULL, spill just the single zone's pages.
2480 * Note that this can be extremely slow as the draining happens in a workqueue.
2482 void drain_all_pages(struct zone *zone)
2487 * Allocate in the BSS so we wont require allocation in
2488 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2490 static cpumask_t cpus_with_pcps;
2493 * Make sure nobody triggers this path before mm_percpu_wq is fully
2496 if (WARN_ON_ONCE(!mm_percpu_wq))
2499 /* Workqueues cannot recurse */
2500 if (current->flags & PF_WQ_WORKER)
2504 * Do not drain if one is already in progress unless it's specific to
2505 * a zone. Such callers are primarily CMA and memory hotplug and need
2506 * the drain to be complete when the call returns.
2508 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2511 mutex_lock(&pcpu_drain_mutex);
2515 * We don't care about racing with CPU hotplug event
2516 * as offline notification will cause the notified
2517 * cpu to drain that CPU pcps and on_each_cpu_mask
2518 * disables preemption as part of its processing
2520 for_each_online_cpu(cpu) {
2521 struct per_cpu_pageset *pcp;
2523 bool has_pcps = false;
2526 pcp = per_cpu_ptr(zone->pageset, cpu);
2530 for_each_populated_zone(z) {
2531 pcp = per_cpu_ptr(z->pageset, cpu);
2532 if (pcp->pcp.count) {
2540 cpumask_set_cpu(cpu, &cpus_with_pcps);
2542 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2545 for_each_cpu(cpu, &cpus_with_pcps) {
2546 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2547 INIT_WORK(work, drain_local_pages_wq);
2548 queue_work_on(cpu, mm_percpu_wq, work);
2550 for_each_cpu(cpu, &cpus_with_pcps)
2551 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2553 mutex_unlock(&pcpu_drain_mutex);
2556 #ifdef CONFIG_HIBERNATION
2559 * Touch the watchdog for every WD_PAGE_COUNT pages.
2561 #define WD_PAGE_COUNT (128*1024)
2563 void mark_free_pages(struct zone *zone)
2565 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2566 unsigned long flags;
2567 unsigned int order, t;
2570 if (zone_is_empty(zone))
2573 spin_lock_irqsave(&zone->lock, flags);
2575 max_zone_pfn = zone_end_pfn(zone);
2576 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2577 if (pfn_valid(pfn)) {
2578 page = pfn_to_page(pfn);
2580 if (!--page_count) {
2581 touch_nmi_watchdog();
2582 page_count = WD_PAGE_COUNT;
2585 if (page_zone(page) != zone)
2588 if (!swsusp_page_is_forbidden(page))
2589 swsusp_unset_page_free(page);
2592 for_each_migratetype_order(order, t) {
2593 list_for_each_entry(page,
2594 &zone->free_area[order].free_list[t], lru) {
2597 pfn = page_to_pfn(page);
2598 for (i = 0; i < (1UL << order); i++) {
2599 if (!--page_count) {
2600 touch_nmi_watchdog();
2601 page_count = WD_PAGE_COUNT;
2603 swsusp_set_page_free(pfn_to_page(pfn + i));
2607 spin_unlock_irqrestore(&zone->lock, flags);
2609 #endif /* CONFIG_PM */
2611 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2615 if (!free_pcp_prepare(page))
2618 migratetype = get_pfnblock_migratetype(page, pfn);
2619 set_pcppage_migratetype(page, migratetype);
2623 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2625 struct zone *zone = page_zone(page);
2626 struct per_cpu_pages *pcp;
2629 migratetype = get_pcppage_migratetype(page);
2630 __count_vm_event(PGFREE);
2633 * We only track unmovable, reclaimable and movable on pcp lists.
2634 * Free ISOLATE pages back to the allocator because they are being
2635 * offlined but treat HIGHATOMIC as movable pages so we can get those
2636 * areas back if necessary. Otherwise, we may have to free
2637 * excessively into the page allocator
2639 if (migratetype >= MIGRATE_PCPTYPES) {
2640 if (unlikely(is_migrate_isolate(migratetype))) {
2641 free_one_page(zone, page, pfn, 0, migratetype);
2644 migratetype = MIGRATE_MOVABLE;
2647 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2648 list_add(&page->lru, &pcp->lists[migratetype]);
2650 if (pcp->count >= pcp->high) {
2651 unsigned long batch = READ_ONCE(pcp->batch);
2652 free_pcppages_bulk(zone, batch, pcp);
2653 pcp->count -= batch;
2658 * Free a 0-order page
2660 void free_unref_page(struct page *page)
2662 unsigned long flags;
2663 unsigned long pfn = page_to_pfn(page);
2665 if (!free_unref_page_prepare(page, pfn))
2668 local_irq_save(flags);
2669 free_unref_page_commit(page, pfn);
2670 local_irq_restore(flags);
2674 * Free a list of 0-order pages
2676 void free_unref_page_list(struct list_head *list)
2678 struct page *page, *next;
2679 unsigned long flags, pfn;
2681 /* Prepare pages for freeing */
2682 list_for_each_entry_safe(page, next, list, lru) {
2683 pfn = page_to_pfn(page);
2684 if (!free_unref_page_prepare(page, pfn))
2685 list_del(&page->lru);
2686 set_page_private(page, pfn);
2689 local_irq_save(flags);
2690 list_for_each_entry_safe(page, next, list, lru) {
2691 unsigned long pfn = page_private(page);
2693 set_page_private(page, 0);
2694 trace_mm_page_free_batched(page);
2695 free_unref_page_commit(page, pfn);
2697 local_irq_restore(flags);
2701 * split_page takes a non-compound higher-order page, and splits it into
2702 * n (1<<order) sub-pages: page[0..n]
2703 * Each sub-page must be freed individually.
2705 * Note: this is probably too low level an operation for use in drivers.
2706 * Please consult with lkml before using this in your driver.
2708 void split_page(struct page *page, unsigned int order)
2712 VM_BUG_ON_PAGE(PageCompound(page), page);
2713 VM_BUG_ON_PAGE(!page_count(page), page);
2715 for (i = 1; i < (1 << order); i++)
2716 set_page_refcounted(page + i);
2717 split_page_owner(page, order);
2719 EXPORT_SYMBOL_GPL(split_page);
2721 int __isolate_free_page(struct page *page, unsigned int order)
2723 unsigned long watermark;
2727 BUG_ON(!PageBuddy(page));
2729 zone = page_zone(page);
2730 mt = get_pageblock_migratetype(page);
2732 if (!is_migrate_isolate(mt)) {
2734 * Obey watermarks as if the page was being allocated. We can
2735 * emulate a high-order watermark check with a raised order-0
2736 * watermark, because we already know our high-order page
2739 watermark = min_wmark_pages(zone) + (1UL << order);
2740 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2743 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2746 /* Remove page from free list */
2747 list_del(&page->lru);
2748 zone->free_area[order].nr_free--;
2749 rmv_page_order(page);
2752 * Set the pageblock if the isolated page is at least half of a
2755 if (order >= pageblock_order - 1) {
2756 struct page *endpage = page + (1 << order) - 1;
2757 for (; page < endpage; page += pageblock_nr_pages) {
2758 int mt = get_pageblock_migratetype(page);
2759 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2760 && !is_migrate_highatomic(mt))
2761 set_pageblock_migratetype(page,
2767 return 1UL << order;
2771 * Update NUMA hit/miss statistics
2773 * Must be called with interrupts disabled.
2775 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2778 enum numa_stat_item local_stat = NUMA_LOCAL;
2780 if (z->node != numa_node_id())
2781 local_stat = NUMA_OTHER;
2783 if (z->node == preferred_zone->node)
2784 __inc_numa_state(z, NUMA_HIT);
2786 __inc_numa_state(z, NUMA_MISS);
2787 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2789 __inc_numa_state(z, local_stat);
2793 /* Remove page from the per-cpu list, caller must protect the list */
2794 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2795 struct per_cpu_pages *pcp,
2796 struct list_head *list)
2801 if (list_empty(list)) {
2802 pcp->count += rmqueue_bulk(zone, 0,
2805 if (unlikely(list_empty(list)))
2809 page = list_first_entry(list, struct page, lru);
2810 list_del(&page->lru);
2812 } while (check_new_pcp(page));
2817 /* Lock and remove page from the per-cpu list */
2818 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2819 struct zone *zone, unsigned int order,
2820 gfp_t gfp_flags, int migratetype)
2822 struct per_cpu_pages *pcp;
2823 struct list_head *list;
2825 unsigned long flags;
2827 local_irq_save(flags);
2828 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2829 list = &pcp->lists[migratetype];
2830 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2832 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2833 zone_statistics(preferred_zone, zone);
2835 local_irq_restore(flags);
2840 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2843 struct page *rmqueue(struct zone *preferred_zone,
2844 struct zone *zone, unsigned int order,
2845 gfp_t gfp_flags, unsigned int alloc_flags,
2848 unsigned long flags;
2851 if (likely(order == 0)) {
2852 page = rmqueue_pcplist(preferred_zone, zone, order,
2853 gfp_flags, migratetype);
2858 * We most definitely don't want callers attempting to
2859 * allocate greater than order-1 page units with __GFP_NOFAIL.
2861 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2862 spin_lock_irqsave(&zone->lock, flags);
2866 if (alloc_flags & ALLOC_HARDER) {
2867 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2869 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2872 page = __rmqueue(zone, order, migratetype);
2873 } while (page && check_new_pages(page, order));
2874 spin_unlock(&zone->lock);
2877 __mod_zone_freepage_state(zone, -(1 << order),
2878 get_pcppage_migratetype(page));
2880 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2881 zone_statistics(preferred_zone, zone);
2882 local_irq_restore(flags);
2885 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2889 local_irq_restore(flags);
2893 #ifdef CONFIG_FAIL_PAGE_ALLOC
2896 struct fault_attr attr;
2898 bool ignore_gfp_highmem;
2899 bool ignore_gfp_reclaim;
2901 } fail_page_alloc = {
2902 .attr = FAULT_ATTR_INITIALIZER,
2903 .ignore_gfp_reclaim = true,
2904 .ignore_gfp_highmem = true,
2908 static int __init setup_fail_page_alloc(char *str)
2910 return setup_fault_attr(&fail_page_alloc.attr, str);
2912 __setup("fail_page_alloc=", setup_fail_page_alloc);
2914 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2916 if (order < fail_page_alloc.min_order)
2918 if (gfp_mask & __GFP_NOFAIL)
2920 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2922 if (fail_page_alloc.ignore_gfp_reclaim &&
2923 (gfp_mask & __GFP_DIRECT_RECLAIM))
2926 return should_fail(&fail_page_alloc.attr, 1 << order);
2929 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2931 static int __init fail_page_alloc_debugfs(void)
2933 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2936 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2937 &fail_page_alloc.attr);
2939 return PTR_ERR(dir);
2941 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2942 &fail_page_alloc.ignore_gfp_reclaim))
2944 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2945 &fail_page_alloc.ignore_gfp_highmem))
2947 if (!debugfs_create_u32("min-order", mode, dir,
2948 &fail_page_alloc.min_order))
2953 debugfs_remove_recursive(dir);
2958 late_initcall(fail_page_alloc_debugfs);
2960 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2962 #else /* CONFIG_FAIL_PAGE_ALLOC */
2964 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2969 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2972 * Return true if free base pages are above 'mark'. For high-order checks it
2973 * will return true of the order-0 watermark is reached and there is at least
2974 * one free page of a suitable size. Checking now avoids taking the zone lock
2975 * to check in the allocation paths if no pages are free.
2977 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2978 int classzone_idx, unsigned int alloc_flags,
2983 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2985 /* free_pages may go negative - that's OK */
2986 free_pages -= (1 << order) - 1;
2988 if (alloc_flags & ALLOC_HIGH)
2992 * If the caller does not have rights to ALLOC_HARDER then subtract
2993 * the high-atomic reserves. This will over-estimate the size of the
2994 * atomic reserve but it avoids a search.
2996 if (likely(!alloc_harder)) {
2997 free_pages -= z->nr_reserved_highatomic;
3000 * OOM victims can try even harder than normal ALLOC_HARDER
3001 * users on the grounds that it's definitely going to be in
3002 * the exit path shortly and free memory. Any allocation it
3003 * makes during the free path will be small and short-lived.
3005 if (alloc_flags & ALLOC_OOM)
3013 /* If allocation can't use CMA areas don't use free CMA pages */
3014 if (!(alloc_flags & ALLOC_CMA))
3015 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3019 * Check watermarks for an order-0 allocation request. If these
3020 * are not met, then a high-order request also cannot go ahead
3021 * even if a suitable page happened to be free.
3023 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3026 /* If this is an order-0 request then the watermark is fine */
3030 /* For a high-order request, check at least one suitable page is free */
3031 for (o = order; o < MAX_ORDER; o++) {
3032 struct free_area *area = &z->free_area[o];
3041 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3042 if (!list_empty(&area->free_list[mt]))
3047 if ((alloc_flags & ALLOC_CMA) &&
3048 !list_empty(&area->free_list[MIGRATE_CMA])) {
3056 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3057 int classzone_idx, unsigned int alloc_flags)
3059 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3060 zone_page_state(z, NR_FREE_PAGES));
3063 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3064 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3066 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3070 /* If allocation can't use CMA areas don't use free CMA pages */
3071 if (!(alloc_flags & ALLOC_CMA))
3072 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3076 * Fast check for order-0 only. If this fails then the reserves
3077 * need to be calculated. There is a corner case where the check
3078 * passes but only the high-order atomic reserve are free. If
3079 * the caller is !atomic then it'll uselessly search the free
3080 * list. That corner case is then slower but it is harmless.
3082 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3085 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3089 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3090 unsigned long mark, int classzone_idx)
3092 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3094 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3095 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3097 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3102 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3104 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3107 #else /* CONFIG_NUMA */
3108 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3112 #endif /* CONFIG_NUMA */
3115 * get_page_from_freelist goes through the zonelist trying to allocate
3118 static struct page *
3119 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3120 const struct alloc_context *ac)
3122 struct zoneref *z = ac->preferred_zoneref;
3124 struct pglist_data *last_pgdat_dirty_limit = NULL;
3127 * Scan zonelist, looking for a zone with enough free.
3128 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3130 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3135 if (cpusets_enabled() &&
3136 (alloc_flags & ALLOC_CPUSET) &&
3137 !__cpuset_zone_allowed(zone, gfp_mask))
3140 * When allocating a page cache page for writing, we
3141 * want to get it from a node that is within its dirty
3142 * limit, such that no single node holds more than its
3143 * proportional share of globally allowed dirty pages.
3144 * The dirty limits take into account the node's
3145 * lowmem reserves and high watermark so that kswapd
3146 * should be able to balance it without having to
3147 * write pages from its LRU list.
3149 * XXX: For now, allow allocations to potentially
3150 * exceed the per-node dirty limit in the slowpath
3151 * (spread_dirty_pages unset) before going into reclaim,
3152 * which is important when on a NUMA setup the allowed
3153 * nodes are together not big enough to reach the
3154 * global limit. The proper fix for these situations
3155 * will require awareness of nodes in the
3156 * dirty-throttling and the flusher threads.
3158 if (ac->spread_dirty_pages) {
3159 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3162 if (!node_dirty_ok(zone->zone_pgdat)) {
3163 last_pgdat_dirty_limit = zone->zone_pgdat;
3168 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3169 if (!zone_watermark_fast(zone, order, mark,
3170 ac_classzone_idx(ac), alloc_flags)) {
3173 /* Checked here to keep the fast path fast */
3174 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3175 if (alloc_flags & ALLOC_NO_WATERMARKS)
3178 if (node_reclaim_mode == 0 ||
3179 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3182 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3184 case NODE_RECLAIM_NOSCAN:
3187 case NODE_RECLAIM_FULL:
3188 /* scanned but unreclaimable */
3191 /* did we reclaim enough */
3192 if (zone_watermark_ok(zone, order, mark,
3193 ac_classzone_idx(ac), alloc_flags))
3201 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3202 gfp_mask, alloc_flags, ac->migratetype);
3204 prep_new_page(page, order, gfp_mask, alloc_flags);
3207 * If this is a high-order atomic allocation then check
3208 * if the pageblock should be reserved for the future
3210 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3211 reserve_highatomic_pageblock(page, zone, order);
3221 * Large machines with many possible nodes should not always dump per-node
3222 * meminfo in irq context.
3224 static inline bool should_suppress_show_mem(void)
3229 ret = in_interrupt();
3234 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3236 unsigned int filter = SHOW_MEM_FILTER_NODES;
3237 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3239 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3243 * This documents exceptions given to allocations in certain
3244 * contexts that are allowed to allocate outside current's set
3247 if (!(gfp_mask & __GFP_NOMEMALLOC))
3248 if (tsk_is_oom_victim(current) ||
3249 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3250 filter &= ~SHOW_MEM_FILTER_NODES;
3251 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3252 filter &= ~SHOW_MEM_FILTER_NODES;
3254 show_mem(filter, nodemask);
3257 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3259 struct va_format vaf;
3261 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3262 DEFAULT_RATELIMIT_BURST);
3264 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3267 pr_warn("%s: ", current->comm);
3269 va_start(args, fmt);
3272 pr_cont("%pV", &vaf);
3275 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3277 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3279 pr_cont("(null)\n");
3281 cpuset_print_current_mems_allowed();
3284 warn_alloc_show_mem(gfp_mask, nodemask);
3287 static inline struct page *
3288 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3289 unsigned int alloc_flags,
3290 const struct alloc_context *ac)
3294 page = get_page_from_freelist(gfp_mask, order,
3295 alloc_flags|ALLOC_CPUSET, ac);
3297 * fallback to ignore cpuset restriction if our nodes
3301 page = get_page_from_freelist(gfp_mask, order,
3307 static inline struct page *
3308 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3309 const struct alloc_context *ac, unsigned long *did_some_progress)
3311 struct oom_control oc = {
3312 .zonelist = ac->zonelist,
3313 .nodemask = ac->nodemask,
3315 .gfp_mask = gfp_mask,
3320 *did_some_progress = 0;
3323 * Acquire the oom lock. If that fails, somebody else is
3324 * making progress for us.
3326 if (!mutex_trylock(&oom_lock)) {
3327 *did_some_progress = 1;
3328 schedule_timeout_uninterruptible(1);
3333 * Go through the zonelist yet one more time, keep very high watermark
3334 * here, this is only to catch a parallel oom killing, we must fail if
3335 * we're still under heavy pressure. But make sure that this reclaim
3336 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3337 * allocation which will never fail due to oom_lock already held.
3339 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3340 ~__GFP_DIRECT_RECLAIM, order,
3341 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3345 /* Coredumps can quickly deplete all memory reserves */
3346 if (current->flags & PF_DUMPCORE)
3348 /* The OOM killer will not help higher order allocs */
3349 if (order > PAGE_ALLOC_COSTLY_ORDER)
3352 * We have already exhausted all our reclaim opportunities without any
3353 * success so it is time to admit defeat. We will skip the OOM killer
3354 * because it is very likely that the caller has a more reasonable
3355 * fallback than shooting a random task.
3357 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3359 /* The OOM killer does not needlessly kill tasks for lowmem */
3360 if (ac->high_zoneidx < ZONE_NORMAL)
3362 if (pm_suspended_storage())
3365 * XXX: GFP_NOFS allocations should rather fail than rely on
3366 * other request to make a forward progress.
3367 * We are in an unfortunate situation where out_of_memory cannot
3368 * do much for this context but let's try it to at least get
3369 * access to memory reserved if the current task is killed (see
3370 * out_of_memory). Once filesystems are ready to handle allocation
3371 * failures more gracefully we should just bail out here.
3374 /* The OOM killer may not free memory on a specific node */
3375 if (gfp_mask & __GFP_THISNODE)
3378 /* Exhausted what can be done so it's blamo time */
3379 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3380 *did_some_progress = 1;
3383 * Help non-failing allocations by giving them access to memory
3386 if (gfp_mask & __GFP_NOFAIL)
3387 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3388 ALLOC_NO_WATERMARKS, ac);
3391 mutex_unlock(&oom_lock);
3396 * Maximum number of compaction retries wit a progress before OOM
3397 * killer is consider as the only way to move forward.
3399 #define MAX_COMPACT_RETRIES 16
3401 #ifdef CONFIG_COMPACTION
3402 /* Try memory compaction for high-order allocations before reclaim */
3403 static struct page *
3404 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3405 unsigned int alloc_flags, const struct alloc_context *ac,
3406 enum compact_priority prio, enum compact_result *compact_result)
3409 unsigned int noreclaim_flag;
3414 noreclaim_flag = memalloc_noreclaim_save();
3415 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3417 memalloc_noreclaim_restore(noreclaim_flag);
3419 if (*compact_result <= COMPACT_INACTIVE)
3423 * At least in one zone compaction wasn't deferred or skipped, so let's
3424 * count a compaction stall
3426 count_vm_event(COMPACTSTALL);
3428 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3431 struct zone *zone = page_zone(page);
3433 zone->compact_blockskip_flush = false;
3434 compaction_defer_reset(zone, order, true);
3435 count_vm_event(COMPACTSUCCESS);
3440 * It's bad if compaction run occurs and fails. The most likely reason
3441 * is that pages exist, but not enough to satisfy watermarks.
3443 count_vm_event(COMPACTFAIL);
3451 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3452 enum compact_result compact_result,
3453 enum compact_priority *compact_priority,
3454 int *compaction_retries)
3456 int max_retries = MAX_COMPACT_RETRIES;
3459 int retries = *compaction_retries;
3460 enum compact_priority priority = *compact_priority;
3465 if (compaction_made_progress(compact_result))
3466 (*compaction_retries)++;
3469 * compaction considers all the zone as desperately out of memory
3470 * so it doesn't really make much sense to retry except when the
3471 * failure could be caused by insufficient priority
3473 if (compaction_failed(compact_result))
3474 goto check_priority;
3477 * make sure the compaction wasn't deferred or didn't bail out early
3478 * due to locks contention before we declare that we should give up.
3479 * But do not retry if the given zonelist is not suitable for
3482 if (compaction_withdrawn(compact_result)) {
3483 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3488 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3489 * costly ones because they are de facto nofail and invoke OOM
3490 * killer to move on while costly can fail and users are ready
3491 * to cope with that. 1/4 retries is rather arbitrary but we
3492 * would need much more detailed feedback from compaction to
3493 * make a better decision.
3495 if (order > PAGE_ALLOC_COSTLY_ORDER)
3497 if (*compaction_retries <= max_retries) {
3503 * Make sure there are attempts at the highest priority if we exhausted
3504 * all retries or failed at the lower priorities.
3507 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3508 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3510 if (*compact_priority > min_priority) {
3511 (*compact_priority)--;
3512 *compaction_retries = 0;
3516 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3520 static inline struct page *
3521 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3522 unsigned int alloc_flags, const struct alloc_context *ac,
3523 enum compact_priority prio, enum compact_result *compact_result)
3525 *compact_result = COMPACT_SKIPPED;
3530 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3531 enum compact_result compact_result,
3532 enum compact_priority *compact_priority,
3533 int *compaction_retries)
3538 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3542 * There are setups with compaction disabled which would prefer to loop
3543 * inside the allocator rather than hit the oom killer prematurely.
3544 * Let's give them a good hope and keep retrying while the order-0
3545 * watermarks are OK.
3547 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3549 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3550 ac_classzone_idx(ac), alloc_flags))
3555 #endif /* CONFIG_COMPACTION */
3557 #ifdef CONFIG_LOCKDEP
3558 struct lockdep_map __fs_reclaim_map =
3559 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3561 static bool __need_fs_reclaim(gfp_t gfp_mask)
3563 gfp_mask = current_gfp_context(gfp_mask);
3565 /* no reclaim without waiting on it */
3566 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3569 /* this guy won't enter reclaim */
3570 if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
3573 /* We're only interested __GFP_FS allocations for now */
3574 if (!(gfp_mask & __GFP_FS))
3577 if (gfp_mask & __GFP_NOLOCKDEP)
3583 void fs_reclaim_acquire(gfp_t gfp_mask)
3585 if (__need_fs_reclaim(gfp_mask))
3586 lock_map_acquire(&__fs_reclaim_map);
3588 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3590 void fs_reclaim_release(gfp_t gfp_mask)
3592 if (__need_fs_reclaim(gfp_mask))
3593 lock_map_release(&__fs_reclaim_map);
3595 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3598 /* Perform direct synchronous page reclaim */
3600 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3601 const struct alloc_context *ac)
3603 struct reclaim_state reclaim_state;
3605 unsigned int noreclaim_flag;
3609 /* We now go into synchronous reclaim */
3610 cpuset_memory_pressure_bump();
3611 noreclaim_flag = memalloc_noreclaim_save();
3612 fs_reclaim_acquire(gfp_mask);
3613 reclaim_state.reclaimed_slab = 0;
3614 current->reclaim_state = &reclaim_state;
3616 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3619 current->reclaim_state = NULL;
3620 fs_reclaim_release(gfp_mask);
3621 memalloc_noreclaim_restore(noreclaim_flag);
3628 /* The really slow allocator path where we enter direct reclaim */
3629 static inline struct page *
3630 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3631 unsigned int alloc_flags, const struct alloc_context *ac,
3632 unsigned long *did_some_progress)
3634 struct page *page = NULL;
3635 bool drained = false;
3637 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3638 if (unlikely(!(*did_some_progress)))
3642 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3645 * If an allocation failed after direct reclaim, it could be because
3646 * pages are pinned on the per-cpu lists or in high alloc reserves.
3647 * Shrink them them and try again
3649 if (!page && !drained) {
3650 unreserve_highatomic_pageblock(ac, false);
3651 drain_all_pages(NULL);
3659 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3663 pg_data_t *last_pgdat = NULL;
3665 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3666 ac->high_zoneidx, ac->nodemask) {
3667 if (last_pgdat != zone->zone_pgdat)
3668 wakeup_kswapd(zone, order, ac->high_zoneidx);
3669 last_pgdat = zone->zone_pgdat;
3673 static inline unsigned int
3674 gfp_to_alloc_flags(gfp_t gfp_mask)
3676 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3678 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3679 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3682 * The caller may dip into page reserves a bit more if the caller
3683 * cannot run direct reclaim, or if the caller has realtime scheduling
3684 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3685 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3687 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3689 if (gfp_mask & __GFP_ATOMIC) {
3691 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3692 * if it can't schedule.
3694 if (!(gfp_mask & __GFP_NOMEMALLOC))
3695 alloc_flags |= ALLOC_HARDER;
3697 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3698 * comment for __cpuset_node_allowed().
3700 alloc_flags &= ~ALLOC_CPUSET;
3701 } else if (unlikely(rt_task(current)) && !in_interrupt())
3702 alloc_flags |= ALLOC_HARDER;
3705 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3706 alloc_flags |= ALLOC_CMA;
3711 static bool oom_reserves_allowed(struct task_struct *tsk)
3713 if (!tsk_is_oom_victim(tsk))
3717 * !MMU doesn't have oom reaper so give access to memory reserves
3718 * only to the thread with TIF_MEMDIE set
3720 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3727 * Distinguish requests which really need access to full memory
3728 * reserves from oom victims which can live with a portion of it
3730 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3732 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3734 if (gfp_mask & __GFP_MEMALLOC)
3735 return ALLOC_NO_WATERMARKS;
3736 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3737 return ALLOC_NO_WATERMARKS;
3738 if (!in_interrupt()) {
3739 if (current->flags & PF_MEMALLOC)
3740 return ALLOC_NO_WATERMARKS;
3741 else if (oom_reserves_allowed(current))
3748 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3750 return !!__gfp_pfmemalloc_flags(gfp_mask);
3754 * Checks whether it makes sense to retry the reclaim to make a forward progress
3755 * for the given allocation request.
3757 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3758 * without success, or when we couldn't even meet the watermark if we
3759 * reclaimed all remaining pages on the LRU lists.
3761 * Returns true if a retry is viable or false to enter the oom path.
3764 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3765 struct alloc_context *ac, int alloc_flags,
3766 bool did_some_progress, int *no_progress_loops)
3772 * Costly allocations might have made a progress but this doesn't mean
3773 * their order will become available due to high fragmentation so
3774 * always increment the no progress counter for them
3776 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3777 *no_progress_loops = 0;
3779 (*no_progress_loops)++;
3782 * Make sure we converge to OOM if we cannot make any progress
3783 * several times in the row.
3785 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3786 /* Before OOM, exhaust highatomic_reserve */
3787 return unreserve_highatomic_pageblock(ac, true);
3791 * Keep reclaiming pages while there is a chance this will lead
3792 * somewhere. If none of the target zones can satisfy our allocation
3793 * request even if all reclaimable pages are considered then we are
3794 * screwed and have to go OOM.
3796 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3798 unsigned long available;
3799 unsigned long reclaimable;
3800 unsigned long min_wmark = min_wmark_pages(zone);
3803 available = reclaimable = zone_reclaimable_pages(zone);
3804 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3807 * Would the allocation succeed if we reclaimed all
3808 * reclaimable pages?
3810 wmark = __zone_watermark_ok(zone, order, min_wmark,
3811 ac_classzone_idx(ac), alloc_flags, available);
3812 trace_reclaim_retry_zone(z, order, reclaimable,
3813 available, min_wmark, *no_progress_loops, wmark);
3816 * If we didn't make any progress and have a lot of
3817 * dirty + writeback pages then we should wait for
3818 * an IO to complete to slow down the reclaim and
3819 * prevent from pre mature OOM
3821 if (!did_some_progress) {
3822 unsigned long write_pending;
3824 write_pending = zone_page_state_snapshot(zone,
3825 NR_ZONE_WRITE_PENDING);
3827 if (2 * write_pending > reclaimable) {
3828 congestion_wait(BLK_RW_ASYNC, HZ/10);
3834 * Memory allocation/reclaim might be called from a WQ
3835 * context and the current implementation of the WQ
3836 * concurrency control doesn't recognize that
3837 * a particular WQ is congested if the worker thread is
3838 * looping without ever sleeping. Therefore we have to
3839 * do a short sleep here rather than calling
3842 if (current->flags & PF_WQ_WORKER)
3843 schedule_timeout_uninterruptible(1);
3855 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3858 * It's possible that cpuset's mems_allowed and the nodemask from
3859 * mempolicy don't intersect. This should be normally dealt with by
3860 * policy_nodemask(), but it's possible to race with cpuset update in
3861 * such a way the check therein was true, and then it became false
3862 * before we got our cpuset_mems_cookie here.
3863 * This assumes that for all allocations, ac->nodemask can come only
3864 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3865 * when it does not intersect with the cpuset restrictions) or the
3866 * caller can deal with a violated nodemask.
3868 if (cpusets_enabled() && ac->nodemask &&
3869 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3870 ac->nodemask = NULL;
3875 * When updating a task's mems_allowed or mempolicy nodemask, it is
3876 * possible to race with parallel threads in such a way that our
3877 * allocation can fail while the mask is being updated. If we are about
3878 * to fail, check if the cpuset changed during allocation and if so,
3881 if (read_mems_allowed_retry(cpuset_mems_cookie))
3887 static inline struct page *
3888 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3889 struct alloc_context *ac)
3891 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3892 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3893 struct page *page = NULL;
3894 unsigned int alloc_flags;
3895 unsigned long did_some_progress;
3896 enum compact_priority compact_priority;
3897 enum compact_result compact_result;
3898 int compaction_retries;
3899 int no_progress_loops;
3900 unsigned long alloc_start = jiffies;
3901 unsigned int stall_timeout = 10 * HZ;
3902 unsigned int cpuset_mems_cookie;
3906 * In the slowpath, we sanity check order to avoid ever trying to
3907 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3908 * be using allocators in order of preference for an area that is
3911 if (order >= MAX_ORDER) {
3912 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3917 * We also sanity check to catch abuse of atomic reserves being used by
3918 * callers that are not in atomic context.
3920 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3921 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3922 gfp_mask &= ~__GFP_ATOMIC;
3925 compaction_retries = 0;
3926 no_progress_loops = 0;
3927 compact_priority = DEF_COMPACT_PRIORITY;
3928 cpuset_mems_cookie = read_mems_allowed_begin();
3931 * The fast path uses conservative alloc_flags to succeed only until
3932 * kswapd needs to be woken up, and to avoid the cost of setting up
3933 * alloc_flags precisely. So we do that now.
3935 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3938 * We need to recalculate the starting point for the zonelist iterator
3939 * because we might have used different nodemask in the fast path, or
3940 * there was a cpuset modification and we are retrying - otherwise we
3941 * could end up iterating over non-eligible zones endlessly.
3943 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3944 ac->high_zoneidx, ac->nodemask);
3945 if (!ac->preferred_zoneref->zone)
3948 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3949 wake_all_kswapds(order, ac);
3952 * The adjusted alloc_flags might result in immediate success, so try
3955 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3960 * For costly allocations, try direct compaction first, as it's likely
3961 * that we have enough base pages and don't need to reclaim. For non-
3962 * movable high-order allocations, do that as well, as compaction will
3963 * try prevent permanent fragmentation by migrating from blocks of the
3965 * Don't try this for allocations that are allowed to ignore
3966 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3968 if (can_direct_reclaim &&
3970 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3971 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3972 page = __alloc_pages_direct_compact(gfp_mask, order,
3974 INIT_COMPACT_PRIORITY,
3980 * Checks for costly allocations with __GFP_NORETRY, which
3981 * includes THP page fault allocations
3983 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3985 * If compaction is deferred for high-order allocations,
3986 * it is because sync compaction recently failed. If
3987 * this is the case and the caller requested a THP
3988 * allocation, we do not want to heavily disrupt the
3989 * system, so we fail the allocation instead of entering
3992 if (compact_result == COMPACT_DEFERRED)
3996 * Looks like reclaim/compaction is worth trying, but
3997 * sync compaction could be very expensive, so keep
3998 * using async compaction.
4000 compact_priority = INIT_COMPACT_PRIORITY;
4005 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4006 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4007 wake_all_kswapds(order, ac);
4009 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4011 alloc_flags = reserve_flags;
4014 * Reset the zonelist iterators if memory policies can be ignored.
4015 * These allocations are high priority and system rather than user
4018 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4019 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4020 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4021 ac->high_zoneidx, ac->nodemask);
4024 /* Attempt with potentially adjusted zonelist and alloc_flags */
4025 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4029 /* Caller is not willing to reclaim, we can't balance anything */
4030 if (!can_direct_reclaim)
4033 /* Make sure we know about allocations which stall for too long */
4034 if (time_after(jiffies, alloc_start + stall_timeout)) {
4035 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
4036 "page allocation stalls for %ums, order:%u",
4037 jiffies_to_msecs(jiffies-alloc_start), order);
4038 stall_timeout += 10 * HZ;
4041 /* Avoid recursion of direct reclaim */
4042 if (current->flags & PF_MEMALLOC)
4045 /* Try direct reclaim and then allocating */
4046 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4047 &did_some_progress);
4051 /* Try direct compaction and then allocating */
4052 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4053 compact_priority, &compact_result);
4057 /* Do not loop if specifically requested */
4058 if (gfp_mask & __GFP_NORETRY)
4062 * Do not retry costly high order allocations unless they are
4063 * __GFP_RETRY_MAYFAIL
4065 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4068 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4069 did_some_progress > 0, &no_progress_loops))
4073 * It doesn't make any sense to retry for the compaction if the order-0
4074 * reclaim is not able to make any progress because the current
4075 * implementation of the compaction depends on the sufficient amount
4076 * of free memory (see __compaction_suitable)
4078 if (did_some_progress > 0 &&
4079 should_compact_retry(ac, order, alloc_flags,
4080 compact_result, &compact_priority,
4081 &compaction_retries))
4085 /* Deal with possible cpuset update races before we start OOM killing */
4086 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4089 /* Reclaim has failed us, start killing things */
4090 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4094 /* Avoid allocations with no watermarks from looping endlessly */
4095 if (tsk_is_oom_victim(current) &&
4096 (alloc_flags == ALLOC_OOM ||
4097 (gfp_mask & __GFP_NOMEMALLOC)))
4100 /* Retry as long as the OOM killer is making progress */
4101 if (did_some_progress) {
4102 no_progress_loops = 0;
4107 /* Deal with possible cpuset update races before we fail */
4108 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4112 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4115 if (gfp_mask & __GFP_NOFAIL) {
4117 * All existing users of the __GFP_NOFAIL are blockable, so warn
4118 * of any new users that actually require GFP_NOWAIT
4120 if (WARN_ON_ONCE(!can_direct_reclaim))
4124 * PF_MEMALLOC request from this context is rather bizarre
4125 * because we cannot reclaim anything and only can loop waiting
4126 * for somebody to do a work for us
4128 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4131 * non failing costly orders are a hard requirement which we
4132 * are not prepared for much so let's warn about these users
4133 * so that we can identify them and convert them to something
4136 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4139 * Help non-failing allocations by giving them access to memory
4140 * reserves but do not use ALLOC_NO_WATERMARKS because this
4141 * could deplete whole memory reserves which would just make
4142 * the situation worse
4144 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4152 warn_alloc(gfp_mask, ac->nodemask,
4153 "page allocation failure: order:%u", order);
4158 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4159 int preferred_nid, nodemask_t *nodemask,
4160 struct alloc_context *ac, gfp_t *alloc_mask,
4161 unsigned int *alloc_flags)
4163 ac->high_zoneidx = gfp_zone(gfp_mask);
4164 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4165 ac->nodemask = nodemask;
4166 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4168 if (cpusets_enabled()) {
4169 *alloc_mask |= __GFP_HARDWALL;
4171 ac->nodemask = &cpuset_current_mems_allowed;
4173 *alloc_flags |= ALLOC_CPUSET;
4176 fs_reclaim_acquire(gfp_mask);
4177 fs_reclaim_release(gfp_mask);
4179 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4181 if (should_fail_alloc_page(gfp_mask, order))
4184 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4185 *alloc_flags |= ALLOC_CMA;
4190 /* Determine whether to spread dirty pages and what the first usable zone */
4191 static inline void finalise_ac(gfp_t gfp_mask,
4192 unsigned int order, struct alloc_context *ac)
4194 /* Dirty zone balancing only done in the fast path */
4195 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4198 * The preferred zone is used for statistics but crucially it is
4199 * also used as the starting point for the zonelist iterator. It
4200 * may get reset for allocations that ignore memory policies.
4202 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4203 ac->high_zoneidx, ac->nodemask);
4207 * This is the 'heart' of the zoned buddy allocator.
4210 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4211 nodemask_t *nodemask)
4214 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4215 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4216 struct alloc_context ac = { };
4218 gfp_mask &= gfp_allowed_mask;
4219 alloc_mask = gfp_mask;
4220 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4223 finalise_ac(gfp_mask, order, &ac);
4225 /* First allocation attempt */
4226 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4231 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4232 * resp. GFP_NOIO which has to be inherited for all allocation requests
4233 * from a particular context which has been marked by
4234 * memalloc_no{fs,io}_{save,restore}.
4236 alloc_mask = current_gfp_context(gfp_mask);
4237 ac.spread_dirty_pages = false;
4240 * Restore the original nodemask if it was potentially replaced with
4241 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4243 if (unlikely(ac.nodemask != nodemask))
4244 ac.nodemask = nodemask;
4246 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4249 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4250 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4251 __free_pages(page, order);
4255 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4259 EXPORT_SYMBOL(__alloc_pages_nodemask);
4262 * Common helper functions.
4264 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4269 * __get_free_pages() returns a 32-bit address, which cannot represent
4272 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4274 page = alloc_pages(gfp_mask, order);
4277 return (unsigned long) page_address(page);
4279 EXPORT_SYMBOL(__get_free_pages);
4281 unsigned long get_zeroed_page(gfp_t gfp_mask)
4283 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4285 EXPORT_SYMBOL(get_zeroed_page);
4287 void __free_pages(struct page *page, unsigned int order)
4289 if (put_page_testzero(page)) {
4291 free_unref_page(page);
4293 __free_pages_ok(page, order);
4297 EXPORT_SYMBOL(__free_pages);
4299 void free_pages(unsigned long addr, unsigned int order)
4302 VM_BUG_ON(!virt_addr_valid((void *)addr));
4303 __free_pages(virt_to_page((void *)addr), order);
4307 EXPORT_SYMBOL(free_pages);
4311 * An arbitrary-length arbitrary-offset area of memory which resides
4312 * within a 0 or higher order page. Multiple fragments within that page
4313 * are individually refcounted, in the page's reference counter.
4315 * The page_frag functions below provide a simple allocation framework for
4316 * page fragments. This is used by the network stack and network device
4317 * drivers to provide a backing region of memory for use as either an
4318 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4320 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4323 struct page *page = NULL;
4324 gfp_t gfp = gfp_mask;
4326 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4327 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4329 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4330 PAGE_FRAG_CACHE_MAX_ORDER);
4331 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4333 if (unlikely(!page))
4334 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4336 nc->va = page ? page_address(page) : NULL;
4341 void __page_frag_cache_drain(struct page *page, unsigned int count)
4343 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4345 if (page_ref_sub_and_test(page, count)) {
4346 unsigned int order = compound_order(page);
4349 free_unref_page(page);
4351 __free_pages_ok(page, order);
4354 EXPORT_SYMBOL(__page_frag_cache_drain);
4356 void *page_frag_alloc(struct page_frag_cache *nc,
4357 unsigned int fragsz, gfp_t gfp_mask)
4359 unsigned int size = PAGE_SIZE;
4363 if (unlikely(!nc->va)) {
4365 page = __page_frag_cache_refill(nc, gfp_mask);
4369 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4370 /* if size can vary use size else just use PAGE_SIZE */
4373 /* Even if we own the page, we do not use atomic_set().
4374 * This would break get_page_unless_zero() users.
4376 page_ref_add(page, size - 1);
4378 /* reset page count bias and offset to start of new frag */
4379 nc->pfmemalloc = page_is_pfmemalloc(page);
4380 nc->pagecnt_bias = size;
4384 offset = nc->offset - fragsz;
4385 if (unlikely(offset < 0)) {
4386 page = virt_to_page(nc->va);
4388 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4391 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4392 /* if size can vary use size else just use PAGE_SIZE */
4395 /* OK, page count is 0, we can safely set it */
4396 set_page_count(page, size);
4398 /* reset page count bias and offset to start of new frag */
4399 nc->pagecnt_bias = size;
4400 offset = size - fragsz;
4404 nc->offset = offset;
4406 return nc->va + offset;
4408 EXPORT_SYMBOL(page_frag_alloc);
4411 * Frees a page fragment allocated out of either a compound or order 0 page.
4413 void page_frag_free(void *addr)
4415 struct page *page = virt_to_head_page(addr);
4417 if (unlikely(put_page_testzero(page)))
4418 __free_pages_ok(page, compound_order(page));
4420 EXPORT_SYMBOL(page_frag_free);
4422 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4426 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4427 unsigned long used = addr + PAGE_ALIGN(size);
4429 split_page(virt_to_page((void *)addr), order);
4430 while (used < alloc_end) {
4435 return (void *)addr;
4439 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4440 * @size: the number of bytes to allocate
4441 * @gfp_mask: GFP flags for the allocation
4443 * This function is similar to alloc_pages(), except that it allocates the
4444 * minimum number of pages to satisfy the request. alloc_pages() can only
4445 * allocate memory in power-of-two pages.
4447 * This function is also limited by MAX_ORDER.
4449 * Memory allocated by this function must be released by free_pages_exact().
4451 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4453 unsigned int order = get_order(size);
4456 addr = __get_free_pages(gfp_mask, order);
4457 return make_alloc_exact(addr, order, size);
4459 EXPORT_SYMBOL(alloc_pages_exact);
4462 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4464 * @nid: the preferred node ID where memory should be allocated
4465 * @size: the number of bytes to allocate
4466 * @gfp_mask: GFP flags for the allocation
4468 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4471 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4473 unsigned int order = get_order(size);
4474 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4477 return make_alloc_exact((unsigned long)page_address(p), order, size);
4481 * free_pages_exact - release memory allocated via alloc_pages_exact()
4482 * @virt: the value returned by alloc_pages_exact.
4483 * @size: size of allocation, same value as passed to alloc_pages_exact().
4485 * Release the memory allocated by a previous call to alloc_pages_exact.
4487 void free_pages_exact(void *virt, size_t size)
4489 unsigned long addr = (unsigned long)virt;
4490 unsigned long end = addr + PAGE_ALIGN(size);
4492 while (addr < end) {
4497 EXPORT_SYMBOL(free_pages_exact);
4500 * nr_free_zone_pages - count number of pages beyond high watermark
4501 * @offset: The zone index of the highest zone
4503 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4504 * high watermark within all zones at or below a given zone index. For each
4505 * zone, the number of pages is calculated as:
4507 * nr_free_zone_pages = managed_pages - high_pages
4509 static unsigned long nr_free_zone_pages(int offset)
4514 /* Just pick one node, since fallback list is circular */
4515 unsigned long sum = 0;
4517 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4519 for_each_zone_zonelist(zone, z, zonelist, offset) {
4520 unsigned long size = zone->managed_pages;
4521 unsigned long high = high_wmark_pages(zone);
4530 * nr_free_buffer_pages - count number of pages beyond high watermark
4532 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4533 * watermark within ZONE_DMA and ZONE_NORMAL.
4535 unsigned long nr_free_buffer_pages(void)
4537 return nr_free_zone_pages(gfp_zone(GFP_USER));
4539 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4542 * nr_free_pagecache_pages - count number of pages beyond high watermark
4544 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4545 * high watermark within all zones.
4547 unsigned long nr_free_pagecache_pages(void)
4549 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4552 static inline void show_node(struct zone *zone)
4554 if (IS_ENABLED(CONFIG_NUMA))
4555 printk("Node %d ", zone_to_nid(zone));
4558 long si_mem_available(void)
4561 unsigned long pagecache;
4562 unsigned long wmark_low = 0;
4563 unsigned long pages[NR_LRU_LISTS];
4567 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4568 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4571 wmark_low += zone->watermark[WMARK_LOW];
4574 * Estimate the amount of memory available for userspace allocations,
4575 * without causing swapping.
4577 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4580 * Not all the page cache can be freed, otherwise the system will
4581 * start swapping. Assume at least half of the page cache, or the
4582 * low watermark worth of cache, needs to stay.
4584 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4585 pagecache -= min(pagecache / 2, wmark_low);
4586 available += pagecache;
4589 * Part of the reclaimable slab consists of items that are in use,
4590 * and cannot be freed. Cap this estimate at the low watermark.
4592 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4593 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4600 EXPORT_SYMBOL_GPL(si_mem_available);
4602 void si_meminfo(struct sysinfo *val)
4604 val->totalram = totalram_pages;
4605 val->sharedram = global_node_page_state(NR_SHMEM);
4606 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4607 val->bufferram = nr_blockdev_pages();
4608 val->totalhigh = totalhigh_pages;
4609 val->freehigh = nr_free_highpages();
4610 val->mem_unit = PAGE_SIZE;
4613 EXPORT_SYMBOL(si_meminfo);
4616 void si_meminfo_node(struct sysinfo *val, int nid)
4618 int zone_type; /* needs to be signed */
4619 unsigned long managed_pages = 0;
4620 unsigned long managed_highpages = 0;
4621 unsigned long free_highpages = 0;
4622 pg_data_t *pgdat = NODE_DATA(nid);
4624 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4625 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4626 val->totalram = managed_pages;
4627 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4628 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4629 #ifdef CONFIG_HIGHMEM
4630 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4631 struct zone *zone = &pgdat->node_zones[zone_type];
4633 if (is_highmem(zone)) {
4634 managed_highpages += zone->managed_pages;
4635 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4638 val->totalhigh = managed_highpages;
4639 val->freehigh = free_highpages;
4641 val->totalhigh = managed_highpages;
4642 val->freehigh = free_highpages;
4644 val->mem_unit = PAGE_SIZE;
4649 * Determine whether the node should be displayed or not, depending on whether
4650 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4652 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4654 if (!(flags & SHOW_MEM_FILTER_NODES))
4658 * no node mask - aka implicit memory numa policy. Do not bother with
4659 * the synchronization - read_mems_allowed_begin - because we do not
4660 * have to be precise here.
4663 nodemask = &cpuset_current_mems_allowed;
4665 return !node_isset(nid, *nodemask);
4668 #define K(x) ((x) << (PAGE_SHIFT-10))
4670 static void show_migration_types(unsigned char type)
4672 static const char types[MIGRATE_TYPES] = {
4673 [MIGRATE_UNMOVABLE] = 'U',
4674 [MIGRATE_MOVABLE] = 'M',
4675 [MIGRATE_RECLAIMABLE] = 'E',
4676 [MIGRATE_HIGHATOMIC] = 'H',
4678 [MIGRATE_CMA] = 'C',
4680 #ifdef CONFIG_MEMORY_ISOLATION
4681 [MIGRATE_ISOLATE] = 'I',
4684 char tmp[MIGRATE_TYPES + 1];
4688 for (i = 0; i < MIGRATE_TYPES; i++) {
4689 if (type & (1 << i))
4694 printk(KERN_CONT "(%s) ", tmp);
4698 * Show free area list (used inside shift_scroll-lock stuff)
4699 * We also calculate the percentage fragmentation. We do this by counting the
4700 * memory on each free list with the exception of the first item on the list.
4703 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4706 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4708 unsigned long free_pcp = 0;
4713 for_each_populated_zone(zone) {
4714 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4717 for_each_online_cpu(cpu)
4718 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4721 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4722 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4723 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4724 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4725 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4726 " free:%lu free_pcp:%lu free_cma:%lu\n",
4727 global_node_page_state(NR_ACTIVE_ANON),
4728 global_node_page_state(NR_INACTIVE_ANON),
4729 global_node_page_state(NR_ISOLATED_ANON),
4730 global_node_page_state(NR_ACTIVE_FILE),
4731 global_node_page_state(NR_INACTIVE_FILE),
4732 global_node_page_state(NR_ISOLATED_FILE),
4733 global_node_page_state(NR_UNEVICTABLE),
4734 global_node_page_state(NR_FILE_DIRTY),
4735 global_node_page_state(NR_WRITEBACK),
4736 global_node_page_state(NR_UNSTABLE_NFS),
4737 global_node_page_state(NR_SLAB_RECLAIMABLE),
4738 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4739 global_node_page_state(NR_FILE_MAPPED),
4740 global_node_page_state(NR_SHMEM),
4741 global_zone_page_state(NR_PAGETABLE),
4742 global_zone_page_state(NR_BOUNCE),
4743 global_zone_page_state(NR_FREE_PAGES),
4745 global_zone_page_state(NR_FREE_CMA_PAGES));
4747 for_each_online_pgdat(pgdat) {
4748 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4752 " active_anon:%lukB"
4753 " inactive_anon:%lukB"
4754 " active_file:%lukB"
4755 " inactive_file:%lukB"
4756 " unevictable:%lukB"
4757 " isolated(anon):%lukB"
4758 " isolated(file):%lukB"
4763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4765 " shmem_pmdmapped: %lukB"
4768 " writeback_tmp:%lukB"
4770 " all_unreclaimable? %s"
4773 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4774 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4775 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4776 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4777 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4778 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4779 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4780 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4781 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4782 K(node_page_state(pgdat, NR_WRITEBACK)),
4783 K(node_page_state(pgdat, NR_SHMEM)),
4784 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4785 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4786 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4788 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4790 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4791 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4792 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4796 for_each_populated_zone(zone) {
4799 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4803 for_each_online_cpu(cpu)
4804 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4813 " active_anon:%lukB"
4814 " inactive_anon:%lukB"
4815 " active_file:%lukB"
4816 " inactive_file:%lukB"
4817 " unevictable:%lukB"
4818 " writepending:%lukB"
4822 " kernel_stack:%lukB"
4830 K(zone_page_state(zone, NR_FREE_PAGES)),
4831 K(min_wmark_pages(zone)),
4832 K(low_wmark_pages(zone)),
4833 K(high_wmark_pages(zone)),
4834 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4835 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4836 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4837 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4838 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4839 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4840 K(zone->present_pages),
4841 K(zone->managed_pages),
4842 K(zone_page_state(zone, NR_MLOCK)),
4843 zone_page_state(zone, NR_KERNEL_STACK_KB),
4844 K(zone_page_state(zone, NR_PAGETABLE)),
4845 K(zone_page_state(zone, NR_BOUNCE)),
4847 K(this_cpu_read(zone->pageset->pcp.count)),
4848 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4849 printk("lowmem_reserve[]:");
4850 for (i = 0; i < MAX_NR_ZONES; i++)
4851 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4852 printk(KERN_CONT "\n");
4855 for_each_populated_zone(zone) {
4857 unsigned long nr[MAX_ORDER], flags, total = 0;
4858 unsigned char types[MAX_ORDER];
4860 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4863 printk(KERN_CONT "%s: ", zone->name);
4865 spin_lock_irqsave(&zone->lock, flags);
4866 for (order = 0; order < MAX_ORDER; order++) {
4867 struct free_area *area = &zone->free_area[order];
4870 nr[order] = area->nr_free;
4871 total += nr[order] << order;
4874 for (type = 0; type < MIGRATE_TYPES; type++) {
4875 if (!list_empty(&area->free_list[type]))
4876 types[order] |= 1 << type;
4879 spin_unlock_irqrestore(&zone->lock, flags);
4880 for (order = 0; order < MAX_ORDER; order++) {
4881 printk(KERN_CONT "%lu*%lukB ",
4882 nr[order], K(1UL) << order);
4884 show_migration_types(types[order]);
4886 printk(KERN_CONT "= %lukB\n", K(total));
4889 hugetlb_show_meminfo();
4891 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4893 show_swap_cache_info();
4896 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4898 zoneref->zone = zone;
4899 zoneref->zone_idx = zone_idx(zone);
4903 * Builds allocation fallback zone lists.
4905 * Add all populated zones of a node to the zonelist.
4907 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4910 enum zone_type zone_type = MAX_NR_ZONES;
4915 zone = pgdat->node_zones + zone_type;
4916 if (managed_zone(zone)) {
4917 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4918 check_highest_zone(zone_type);
4920 } while (zone_type);
4927 static int __parse_numa_zonelist_order(char *s)
4930 * We used to support different zonlists modes but they turned
4931 * out to be just not useful. Let's keep the warning in place
4932 * if somebody still use the cmd line parameter so that we do
4933 * not fail it silently
4935 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4936 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4942 static __init int setup_numa_zonelist_order(char *s)
4947 return __parse_numa_zonelist_order(s);
4949 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4951 char numa_zonelist_order[] = "Node";
4954 * sysctl handler for numa_zonelist_order
4956 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4957 void __user *buffer, size_t *length,
4964 return proc_dostring(table, write, buffer, length, ppos);
4965 str = memdup_user_nul(buffer, 16);
4967 return PTR_ERR(str);
4969 ret = __parse_numa_zonelist_order(str);
4975 #define MAX_NODE_LOAD (nr_online_nodes)
4976 static int node_load[MAX_NUMNODES];
4979 * find_next_best_node - find the next node that should appear in a given node's fallback list
4980 * @node: node whose fallback list we're appending
4981 * @used_node_mask: nodemask_t of already used nodes
4983 * We use a number of factors to determine which is the next node that should
4984 * appear on a given node's fallback list. The node should not have appeared
4985 * already in @node's fallback list, and it should be the next closest node
4986 * according to the distance array (which contains arbitrary distance values
4987 * from each node to each node in the system), and should also prefer nodes
4988 * with no CPUs, since presumably they'll have very little allocation pressure
4989 * on them otherwise.
4990 * It returns -1 if no node is found.
4992 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4995 int min_val = INT_MAX;
4996 int best_node = NUMA_NO_NODE;
4997 const struct cpumask *tmp = cpumask_of_node(0);
4999 /* Use the local node if we haven't already */
5000 if (!node_isset(node, *used_node_mask)) {
5001 node_set(node, *used_node_mask);
5005 for_each_node_state(n, N_MEMORY) {
5007 /* Don't want a node to appear more than once */
5008 if (node_isset(n, *used_node_mask))
5011 /* Use the distance array to find the distance */
5012 val = node_distance(node, n);
5014 /* Penalize nodes under us ("prefer the next node") */
5017 /* Give preference to headless and unused nodes */
5018 tmp = cpumask_of_node(n);
5019 if (!cpumask_empty(tmp))
5020 val += PENALTY_FOR_NODE_WITH_CPUS;
5022 /* Slight preference for less loaded node */
5023 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5024 val += node_load[n];
5026 if (val < min_val) {
5033 node_set(best_node, *used_node_mask);
5040 * Build zonelists ordered by node and zones within node.
5041 * This results in maximum locality--normal zone overflows into local
5042 * DMA zone, if any--but risks exhausting DMA zone.
5044 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5047 struct zoneref *zonerefs;
5050 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5052 for (i = 0; i < nr_nodes; i++) {
5055 pg_data_t *node = NODE_DATA(node_order[i]);
5057 nr_zones = build_zonerefs_node(node, zonerefs);
5058 zonerefs += nr_zones;
5060 zonerefs->zone = NULL;
5061 zonerefs->zone_idx = 0;
5065 * Build gfp_thisnode zonelists
5067 static void build_thisnode_zonelists(pg_data_t *pgdat)
5069 struct zoneref *zonerefs;
5072 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5073 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5074 zonerefs += nr_zones;
5075 zonerefs->zone = NULL;
5076 zonerefs->zone_idx = 0;
5080 * Build zonelists ordered by zone and nodes within zones.
5081 * This results in conserving DMA zone[s] until all Normal memory is
5082 * exhausted, but results in overflowing to remote node while memory
5083 * may still exist in local DMA zone.
5086 static void build_zonelists(pg_data_t *pgdat)
5088 static int node_order[MAX_NUMNODES];
5089 int node, load, nr_nodes = 0;
5090 nodemask_t used_mask;
5091 int local_node, prev_node;
5093 /* NUMA-aware ordering of nodes */
5094 local_node = pgdat->node_id;
5095 load = nr_online_nodes;
5096 prev_node = local_node;
5097 nodes_clear(used_mask);
5099 memset(node_order, 0, sizeof(node_order));
5100 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5102 * We don't want to pressure a particular node.
5103 * So adding penalty to the first node in same
5104 * distance group to make it round-robin.
5106 if (node_distance(local_node, node) !=
5107 node_distance(local_node, prev_node))
5108 node_load[node] = load;
5110 node_order[nr_nodes++] = node;
5115 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5116 build_thisnode_zonelists(pgdat);
5119 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5121 * Return node id of node used for "local" allocations.
5122 * I.e., first node id of first zone in arg node's generic zonelist.
5123 * Used for initializing percpu 'numa_mem', which is used primarily
5124 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5126 int local_memory_node(int node)
5130 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5131 gfp_zone(GFP_KERNEL),
5133 return z->zone->node;
5137 static void setup_min_unmapped_ratio(void);
5138 static void setup_min_slab_ratio(void);
5139 #else /* CONFIG_NUMA */
5141 static void build_zonelists(pg_data_t *pgdat)
5143 int node, local_node;
5144 struct zoneref *zonerefs;
5147 local_node = pgdat->node_id;
5149 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5150 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5151 zonerefs += nr_zones;
5154 * Now we build the zonelist so that it contains the zones
5155 * of all the other nodes.
5156 * We don't want to pressure a particular node, so when
5157 * building the zones for node N, we make sure that the
5158 * zones coming right after the local ones are those from
5159 * node N+1 (modulo N)
5161 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5162 if (!node_online(node))
5164 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5165 zonerefs += nr_zones;
5167 for (node = 0; node < local_node; node++) {
5168 if (!node_online(node))
5170 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5171 zonerefs += nr_zones;
5174 zonerefs->zone = NULL;
5175 zonerefs->zone_idx = 0;
5178 #endif /* CONFIG_NUMA */
5181 * Boot pageset table. One per cpu which is going to be used for all
5182 * zones and all nodes. The parameters will be set in such a way
5183 * that an item put on a list will immediately be handed over to
5184 * the buddy list. This is safe since pageset manipulation is done
5185 * with interrupts disabled.
5187 * The boot_pagesets must be kept even after bootup is complete for
5188 * unused processors and/or zones. They do play a role for bootstrapping
5189 * hotplugged processors.
5191 * zoneinfo_show() and maybe other functions do
5192 * not check if the processor is online before following the pageset pointer.
5193 * Other parts of the kernel may not check if the zone is available.
5195 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5196 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5197 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5199 static void __build_all_zonelists(void *data)
5202 int __maybe_unused cpu;
5203 pg_data_t *self = data;
5204 static DEFINE_SPINLOCK(lock);
5209 memset(node_load, 0, sizeof(node_load));
5213 * This node is hotadded and no memory is yet present. So just
5214 * building zonelists is fine - no need to touch other nodes.
5216 if (self && !node_online(self->node_id)) {
5217 build_zonelists(self);
5219 for_each_online_node(nid) {
5220 pg_data_t *pgdat = NODE_DATA(nid);
5222 build_zonelists(pgdat);
5225 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5227 * We now know the "local memory node" for each node--
5228 * i.e., the node of the first zone in the generic zonelist.
5229 * Set up numa_mem percpu variable for on-line cpus. During
5230 * boot, only the boot cpu should be on-line; we'll init the
5231 * secondary cpus' numa_mem as they come on-line. During
5232 * node/memory hotplug, we'll fixup all on-line cpus.
5234 for_each_online_cpu(cpu)
5235 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5242 static noinline void __init
5243 build_all_zonelists_init(void)
5247 __build_all_zonelists(NULL);
5250 * Initialize the boot_pagesets that are going to be used
5251 * for bootstrapping processors. The real pagesets for
5252 * each zone will be allocated later when the per cpu
5253 * allocator is available.
5255 * boot_pagesets are used also for bootstrapping offline
5256 * cpus if the system is already booted because the pagesets
5257 * are needed to initialize allocators on a specific cpu too.
5258 * F.e. the percpu allocator needs the page allocator which
5259 * needs the percpu allocator in order to allocate its pagesets
5260 * (a chicken-egg dilemma).
5262 for_each_possible_cpu(cpu)
5263 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5265 mminit_verify_zonelist();
5266 cpuset_init_current_mems_allowed();
5270 * unless system_state == SYSTEM_BOOTING.
5272 * __ref due to call of __init annotated helper build_all_zonelists_init
5273 * [protected by SYSTEM_BOOTING].
5275 void __ref build_all_zonelists(pg_data_t *pgdat)
5277 if (system_state == SYSTEM_BOOTING) {
5278 build_all_zonelists_init();
5280 __build_all_zonelists(pgdat);
5281 /* cpuset refresh routine should be here */
5283 vm_total_pages = nr_free_pagecache_pages();
5285 * Disable grouping by mobility if the number of pages in the
5286 * system is too low to allow the mechanism to work. It would be
5287 * more accurate, but expensive to check per-zone. This check is
5288 * made on memory-hotadd so a system can start with mobility
5289 * disabled and enable it later
5291 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5292 page_group_by_mobility_disabled = 1;
5294 page_group_by_mobility_disabled = 0;
5296 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5298 page_group_by_mobility_disabled ? "off" : "on",
5301 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5306 * Initially all pages are reserved - free ones are freed
5307 * up by free_all_bootmem() once the early boot process is
5308 * done. Non-atomic initialization, single-pass.
5310 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5311 unsigned long start_pfn, enum memmap_context context)
5313 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5314 unsigned long end_pfn = start_pfn + size;
5315 pg_data_t *pgdat = NODE_DATA(nid);
5317 unsigned long nr_initialised = 0;
5318 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5319 struct memblock_region *r = NULL, *tmp;
5322 if (highest_memmap_pfn < end_pfn - 1)
5323 highest_memmap_pfn = end_pfn - 1;
5326 * Honor reservation requested by the driver for this ZONE_DEVICE
5329 if (altmap && start_pfn == altmap->base_pfn)
5330 start_pfn += altmap->reserve;
5332 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5334 * There can be holes in boot-time mem_map[]s handed to this
5335 * function. They do not exist on hotplugged memory.
5337 if (context != MEMMAP_EARLY)
5340 if (!early_pfn_valid(pfn)) {
5341 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5343 * Skip to the pfn preceding the next valid one (or
5344 * end_pfn), such that we hit a valid pfn (or end_pfn)
5345 * on our next iteration of the loop.
5347 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5351 if (!early_pfn_in_nid(pfn, nid))
5353 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5356 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5358 * Check given memblock attribute by firmware which can affect
5359 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5360 * mirrored, it's an overlapped memmap init. skip it.
5362 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5363 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5364 for_each_memblock(memory, tmp)
5365 if (pfn < memblock_region_memory_end_pfn(tmp))
5369 if (pfn >= memblock_region_memory_base_pfn(r) &&
5370 memblock_is_mirror(r)) {
5371 /* already initialized as NORMAL */
5372 pfn = memblock_region_memory_end_pfn(r);
5380 * Mark the block movable so that blocks are reserved for
5381 * movable at startup. This will force kernel allocations
5382 * to reserve their blocks rather than leaking throughout
5383 * the address space during boot when many long-lived
5384 * kernel allocations are made.
5386 * bitmap is created for zone's valid pfn range. but memmap
5387 * can be created for invalid pages (for alignment)
5388 * check here not to call set_pageblock_migratetype() against
5391 if (!(pfn & (pageblock_nr_pages - 1))) {
5392 struct page *page = pfn_to_page(pfn);
5394 __init_single_page(page, pfn, zone, nid);
5395 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5398 __init_single_pfn(pfn, zone, nid);
5403 static void __meminit zone_init_free_lists(struct zone *zone)
5405 unsigned int order, t;
5406 for_each_migratetype_order(order, t) {
5407 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5408 zone->free_area[order].nr_free = 0;
5412 #ifndef __HAVE_ARCH_MEMMAP_INIT
5413 #define memmap_init(size, nid, zone, start_pfn) \
5414 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5417 static int zone_batchsize(struct zone *zone)
5423 * The per-cpu-pages pools are set to around 1000th of the
5424 * size of the zone. But no more than 1/2 of a meg.
5426 * OK, so we don't know how big the cache is. So guess.
5428 batch = zone->managed_pages / 1024;
5429 if (batch * PAGE_SIZE > 512 * 1024)
5430 batch = (512 * 1024) / PAGE_SIZE;
5431 batch /= 4; /* We effectively *= 4 below */
5436 * Clamp the batch to a 2^n - 1 value. Having a power
5437 * of 2 value was found to be more likely to have
5438 * suboptimal cache aliasing properties in some cases.
5440 * For example if 2 tasks are alternately allocating
5441 * batches of pages, one task can end up with a lot
5442 * of pages of one half of the possible page colors
5443 * and the other with pages of the other colors.
5445 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5450 /* The deferral and batching of frees should be suppressed under NOMMU
5453 * The problem is that NOMMU needs to be able to allocate large chunks
5454 * of contiguous memory as there's no hardware page translation to
5455 * assemble apparent contiguous memory from discontiguous pages.
5457 * Queueing large contiguous runs of pages for batching, however,
5458 * causes the pages to actually be freed in smaller chunks. As there
5459 * can be a significant delay between the individual batches being
5460 * recycled, this leads to the once large chunks of space being
5461 * fragmented and becoming unavailable for high-order allocations.
5468 * pcp->high and pcp->batch values are related and dependent on one another:
5469 * ->batch must never be higher then ->high.
5470 * The following function updates them in a safe manner without read side
5473 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5474 * those fields changing asynchronously (acording the the above rule).
5476 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5477 * outside of boot time (or some other assurance that no concurrent updaters
5480 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5481 unsigned long batch)
5483 /* start with a fail safe value for batch */
5487 /* Update high, then batch, in order */
5494 /* a companion to pageset_set_high() */
5495 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5497 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5500 static void pageset_init(struct per_cpu_pageset *p)
5502 struct per_cpu_pages *pcp;
5505 memset(p, 0, sizeof(*p));
5509 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5510 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5513 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5516 pageset_set_batch(p, batch);
5520 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5521 * to the value high for the pageset p.
5523 static void pageset_set_high(struct per_cpu_pageset *p,
5526 unsigned long batch = max(1UL, high / 4);
5527 if ((high / 4) > (PAGE_SHIFT * 8))
5528 batch = PAGE_SHIFT * 8;
5530 pageset_update(&p->pcp, high, batch);
5533 static void pageset_set_high_and_batch(struct zone *zone,
5534 struct per_cpu_pageset *pcp)
5536 if (percpu_pagelist_fraction)
5537 pageset_set_high(pcp,
5538 (zone->managed_pages /
5539 percpu_pagelist_fraction));
5541 pageset_set_batch(pcp, zone_batchsize(zone));
5544 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5546 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5549 pageset_set_high_and_batch(zone, pcp);
5552 void __meminit setup_zone_pageset(struct zone *zone)
5555 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5556 for_each_possible_cpu(cpu)
5557 zone_pageset_init(zone, cpu);
5561 * Allocate per cpu pagesets and initialize them.
5562 * Before this call only boot pagesets were available.
5564 void __init setup_per_cpu_pageset(void)
5566 struct pglist_data *pgdat;
5569 for_each_populated_zone(zone)
5570 setup_zone_pageset(zone);
5572 for_each_online_pgdat(pgdat)
5573 pgdat->per_cpu_nodestats =
5574 alloc_percpu(struct per_cpu_nodestat);
5577 static __meminit void zone_pcp_init(struct zone *zone)
5580 * per cpu subsystem is not up at this point. The following code
5581 * relies on the ability of the linker to provide the
5582 * offset of a (static) per cpu variable into the per cpu area.
5584 zone->pageset = &boot_pageset;
5586 if (populated_zone(zone))
5587 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5588 zone->name, zone->present_pages,
5589 zone_batchsize(zone));
5592 void __meminit init_currently_empty_zone(struct zone *zone,
5593 unsigned long zone_start_pfn,
5596 struct pglist_data *pgdat = zone->zone_pgdat;
5598 pgdat->nr_zones = zone_idx(zone) + 1;
5600 zone->zone_start_pfn = zone_start_pfn;
5602 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5603 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5605 (unsigned long)zone_idx(zone),
5606 zone_start_pfn, (zone_start_pfn + size));
5608 zone_init_free_lists(zone);
5609 zone->initialized = 1;
5612 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5613 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5616 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5618 int __meminit __early_pfn_to_nid(unsigned long pfn,
5619 struct mminit_pfnnid_cache *state)
5621 unsigned long start_pfn, end_pfn;
5624 if (state->last_start <= pfn && pfn < state->last_end)
5625 return state->last_nid;
5627 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5629 state->last_start = start_pfn;
5630 state->last_end = end_pfn;
5631 state->last_nid = nid;
5636 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5639 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5640 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5641 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5643 * If an architecture guarantees that all ranges registered contain no holes
5644 * and may be freed, this this function may be used instead of calling
5645 * memblock_free_early_nid() manually.
5647 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5649 unsigned long start_pfn, end_pfn;
5652 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5653 start_pfn = min(start_pfn, max_low_pfn);
5654 end_pfn = min(end_pfn, max_low_pfn);
5656 if (start_pfn < end_pfn)
5657 memblock_free_early_nid(PFN_PHYS(start_pfn),
5658 (end_pfn - start_pfn) << PAGE_SHIFT,
5664 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5665 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5667 * If an architecture guarantees that all ranges registered contain no holes and may
5668 * be freed, this function may be used instead of calling memory_present() manually.
5670 void __init sparse_memory_present_with_active_regions(int nid)
5672 unsigned long start_pfn, end_pfn;
5675 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5676 memory_present(this_nid, start_pfn, end_pfn);
5680 * get_pfn_range_for_nid - Return the start and end page frames for a node
5681 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5682 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5683 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5685 * It returns the start and end page frame of a node based on information
5686 * provided by memblock_set_node(). If called for a node
5687 * with no available memory, a warning is printed and the start and end
5690 void __meminit get_pfn_range_for_nid(unsigned int nid,
5691 unsigned long *start_pfn, unsigned long *end_pfn)
5693 unsigned long this_start_pfn, this_end_pfn;
5699 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5700 *start_pfn = min(*start_pfn, this_start_pfn);
5701 *end_pfn = max(*end_pfn, this_end_pfn);
5704 if (*start_pfn == -1UL)
5709 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5710 * assumption is made that zones within a node are ordered in monotonic
5711 * increasing memory addresses so that the "highest" populated zone is used
5713 static void __init find_usable_zone_for_movable(void)
5716 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5717 if (zone_index == ZONE_MOVABLE)
5720 if (arch_zone_highest_possible_pfn[zone_index] >
5721 arch_zone_lowest_possible_pfn[zone_index])
5725 VM_BUG_ON(zone_index == -1);
5726 movable_zone = zone_index;
5730 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5731 * because it is sized independent of architecture. Unlike the other zones,
5732 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5733 * in each node depending on the size of each node and how evenly kernelcore
5734 * is distributed. This helper function adjusts the zone ranges
5735 * provided by the architecture for a given node by using the end of the
5736 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5737 * zones within a node are in order of monotonic increases memory addresses
5739 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5740 unsigned long zone_type,
5741 unsigned long node_start_pfn,
5742 unsigned long node_end_pfn,
5743 unsigned long *zone_start_pfn,
5744 unsigned long *zone_end_pfn)
5746 /* Only adjust if ZONE_MOVABLE is on this node */
5747 if (zone_movable_pfn[nid]) {
5748 /* Size ZONE_MOVABLE */
5749 if (zone_type == ZONE_MOVABLE) {
5750 *zone_start_pfn = zone_movable_pfn[nid];
5751 *zone_end_pfn = min(node_end_pfn,
5752 arch_zone_highest_possible_pfn[movable_zone]);
5754 /* Adjust for ZONE_MOVABLE starting within this range */
5755 } else if (!mirrored_kernelcore &&
5756 *zone_start_pfn < zone_movable_pfn[nid] &&
5757 *zone_end_pfn > zone_movable_pfn[nid]) {
5758 *zone_end_pfn = zone_movable_pfn[nid];
5760 /* Check if this whole range is within ZONE_MOVABLE */
5761 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5762 *zone_start_pfn = *zone_end_pfn;
5767 * Return the number of pages a zone spans in a node, including holes
5768 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5770 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5771 unsigned long zone_type,
5772 unsigned long node_start_pfn,
5773 unsigned long node_end_pfn,
5774 unsigned long *zone_start_pfn,
5775 unsigned long *zone_end_pfn,
5776 unsigned long *ignored)
5778 /* When hotadd a new node from cpu_up(), the node should be empty */
5779 if (!node_start_pfn && !node_end_pfn)
5782 /* Get the start and end of the zone */
5783 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5784 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5785 adjust_zone_range_for_zone_movable(nid, zone_type,
5786 node_start_pfn, node_end_pfn,
5787 zone_start_pfn, zone_end_pfn);
5789 /* Check that this node has pages within the zone's required range */
5790 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5793 /* Move the zone boundaries inside the node if necessary */
5794 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5795 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5797 /* Return the spanned pages */
5798 return *zone_end_pfn - *zone_start_pfn;
5802 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5803 * then all holes in the requested range will be accounted for.
5805 unsigned long __meminit __absent_pages_in_range(int nid,
5806 unsigned long range_start_pfn,
5807 unsigned long range_end_pfn)
5809 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5810 unsigned long start_pfn, end_pfn;
5813 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5814 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5815 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5816 nr_absent -= end_pfn - start_pfn;
5822 * absent_pages_in_range - Return number of page frames in holes within a range
5823 * @start_pfn: The start PFN to start searching for holes
5824 * @end_pfn: The end PFN to stop searching for holes
5826 * It returns the number of pages frames in memory holes within a range.
5828 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5829 unsigned long end_pfn)
5831 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5834 /* Return the number of page frames in holes in a zone on a node */
5835 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5836 unsigned long zone_type,
5837 unsigned long node_start_pfn,
5838 unsigned long node_end_pfn,
5839 unsigned long *ignored)
5841 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5842 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5843 unsigned long zone_start_pfn, zone_end_pfn;
5844 unsigned long nr_absent;
5846 /* When hotadd a new node from cpu_up(), the node should be empty */
5847 if (!node_start_pfn && !node_end_pfn)
5850 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5851 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5853 adjust_zone_range_for_zone_movable(nid, zone_type,
5854 node_start_pfn, node_end_pfn,
5855 &zone_start_pfn, &zone_end_pfn);
5856 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5859 * ZONE_MOVABLE handling.
5860 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5863 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5864 unsigned long start_pfn, end_pfn;
5865 struct memblock_region *r;
5867 for_each_memblock(memory, r) {
5868 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5869 zone_start_pfn, zone_end_pfn);
5870 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5871 zone_start_pfn, zone_end_pfn);
5873 if (zone_type == ZONE_MOVABLE &&
5874 memblock_is_mirror(r))
5875 nr_absent += end_pfn - start_pfn;
5877 if (zone_type == ZONE_NORMAL &&
5878 !memblock_is_mirror(r))
5879 nr_absent += end_pfn - start_pfn;
5886 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5887 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5888 unsigned long zone_type,
5889 unsigned long node_start_pfn,
5890 unsigned long node_end_pfn,
5891 unsigned long *zone_start_pfn,
5892 unsigned long *zone_end_pfn,
5893 unsigned long *zones_size)
5897 *zone_start_pfn = node_start_pfn;
5898 for (zone = 0; zone < zone_type; zone++)
5899 *zone_start_pfn += zones_size[zone];
5901 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5903 return zones_size[zone_type];
5906 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5907 unsigned long zone_type,
5908 unsigned long node_start_pfn,
5909 unsigned long node_end_pfn,
5910 unsigned long *zholes_size)
5915 return zholes_size[zone_type];
5918 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5920 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5921 unsigned long node_start_pfn,
5922 unsigned long node_end_pfn,
5923 unsigned long *zones_size,
5924 unsigned long *zholes_size)
5926 unsigned long realtotalpages = 0, totalpages = 0;
5929 for (i = 0; i < MAX_NR_ZONES; i++) {
5930 struct zone *zone = pgdat->node_zones + i;
5931 unsigned long zone_start_pfn, zone_end_pfn;
5932 unsigned long size, real_size;
5934 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5940 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5941 node_start_pfn, node_end_pfn,
5944 zone->zone_start_pfn = zone_start_pfn;
5946 zone->zone_start_pfn = 0;
5947 zone->spanned_pages = size;
5948 zone->present_pages = real_size;
5951 realtotalpages += real_size;
5954 pgdat->node_spanned_pages = totalpages;
5955 pgdat->node_present_pages = realtotalpages;
5956 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5960 #ifndef CONFIG_SPARSEMEM
5962 * Calculate the size of the zone->blockflags rounded to an unsigned long
5963 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5964 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5965 * round what is now in bits to nearest long in bits, then return it in
5968 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5970 unsigned long usemapsize;
5972 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5973 usemapsize = roundup(zonesize, pageblock_nr_pages);
5974 usemapsize = usemapsize >> pageblock_order;
5975 usemapsize *= NR_PAGEBLOCK_BITS;
5976 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5978 return usemapsize / 8;
5981 static void __init setup_usemap(struct pglist_data *pgdat,
5983 unsigned long zone_start_pfn,
5984 unsigned long zonesize)
5986 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5987 zone->pageblock_flags = NULL;
5989 zone->pageblock_flags =
5990 memblock_virt_alloc_node_nopanic(usemapsize,
5994 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5995 unsigned long zone_start_pfn, unsigned long zonesize) {}
5996 #endif /* CONFIG_SPARSEMEM */
5998 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6000 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6001 void __paginginit set_pageblock_order(void)
6005 /* Check that pageblock_nr_pages has not already been setup */
6006 if (pageblock_order)
6009 if (HPAGE_SHIFT > PAGE_SHIFT)
6010 order = HUGETLB_PAGE_ORDER;
6012 order = MAX_ORDER - 1;
6015 * Assume the largest contiguous order of interest is a huge page.
6016 * This value may be variable depending on boot parameters on IA64 and
6019 pageblock_order = order;
6021 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6024 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6025 * is unused as pageblock_order is set at compile-time. See
6026 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6029 void __paginginit set_pageblock_order(void)
6033 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6035 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6036 unsigned long present_pages)
6038 unsigned long pages = spanned_pages;
6041 * Provide a more accurate estimation if there are holes within
6042 * the zone and SPARSEMEM is in use. If there are holes within the
6043 * zone, each populated memory region may cost us one or two extra
6044 * memmap pages due to alignment because memmap pages for each
6045 * populated regions may not be naturally aligned on page boundary.
6046 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6048 if (spanned_pages > present_pages + (present_pages >> 4) &&
6049 IS_ENABLED(CONFIG_SPARSEMEM))
6050 pages = present_pages;
6052 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6056 * Set up the zone data structures:
6057 * - mark all pages reserved
6058 * - mark all memory queues empty
6059 * - clear the memory bitmaps
6061 * NOTE: pgdat should get zeroed by caller.
6063 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6066 int nid = pgdat->node_id;
6068 pgdat_resize_init(pgdat);
6069 #ifdef CONFIG_NUMA_BALANCING
6070 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6071 pgdat->numabalancing_migrate_nr_pages = 0;
6072 pgdat->numabalancing_migrate_next_window = jiffies;
6074 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6075 spin_lock_init(&pgdat->split_queue_lock);
6076 INIT_LIST_HEAD(&pgdat->split_queue);
6077 pgdat->split_queue_len = 0;
6079 init_waitqueue_head(&pgdat->kswapd_wait);
6080 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6081 #ifdef CONFIG_COMPACTION
6082 init_waitqueue_head(&pgdat->kcompactd_wait);
6084 pgdat_page_ext_init(pgdat);
6085 spin_lock_init(&pgdat->lru_lock);
6086 lruvec_init(node_lruvec(pgdat));
6088 pgdat->per_cpu_nodestats = &boot_nodestats;
6090 for (j = 0; j < MAX_NR_ZONES; j++) {
6091 struct zone *zone = pgdat->node_zones + j;
6092 unsigned long size, realsize, freesize, memmap_pages;
6093 unsigned long zone_start_pfn = zone->zone_start_pfn;
6095 size = zone->spanned_pages;
6096 realsize = freesize = zone->present_pages;
6099 * Adjust freesize so that it accounts for how much memory
6100 * is used by this zone for memmap. This affects the watermark
6101 * and per-cpu initialisations
6103 memmap_pages = calc_memmap_size(size, realsize);
6104 if (!is_highmem_idx(j)) {
6105 if (freesize >= memmap_pages) {
6106 freesize -= memmap_pages;
6109 " %s zone: %lu pages used for memmap\n",
6110 zone_names[j], memmap_pages);
6112 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6113 zone_names[j], memmap_pages, freesize);
6116 /* Account for reserved pages */
6117 if (j == 0 && freesize > dma_reserve) {
6118 freesize -= dma_reserve;
6119 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6120 zone_names[0], dma_reserve);
6123 if (!is_highmem_idx(j))
6124 nr_kernel_pages += freesize;
6125 /* Charge for highmem memmap if there are enough kernel pages */
6126 else if (nr_kernel_pages > memmap_pages * 2)
6127 nr_kernel_pages -= memmap_pages;
6128 nr_all_pages += freesize;
6131 * Set an approximate value for lowmem here, it will be adjusted
6132 * when the bootmem allocator frees pages into the buddy system.
6133 * And all highmem pages will be managed by the buddy system.
6135 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6139 zone->name = zone_names[j];
6140 zone->zone_pgdat = pgdat;
6141 spin_lock_init(&zone->lock);
6142 zone_seqlock_init(zone);
6143 zone_pcp_init(zone);
6148 set_pageblock_order();
6149 setup_usemap(pgdat, zone, zone_start_pfn, size);
6150 init_currently_empty_zone(zone, zone_start_pfn, size);
6151 memmap_init(size, nid, j, zone_start_pfn);
6155 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6157 unsigned long __maybe_unused start = 0;
6158 unsigned long __maybe_unused offset = 0;
6160 /* Skip empty nodes */
6161 if (!pgdat->node_spanned_pages)
6164 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6165 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6166 offset = pgdat->node_start_pfn - start;
6167 /* ia64 gets its own node_mem_map, before this, without bootmem */
6168 if (!pgdat->node_mem_map) {
6169 unsigned long size, end;
6173 * The zone's endpoints aren't required to be MAX_ORDER
6174 * aligned but the node_mem_map endpoints must be in order
6175 * for the buddy allocator to function correctly.
6177 end = pgdat_end_pfn(pgdat);
6178 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6179 size = (end - start) * sizeof(struct page);
6180 map = alloc_remap(pgdat->node_id, size);
6182 map = memblock_virt_alloc_node_nopanic(size,
6184 pgdat->node_mem_map = map + offset;
6186 #ifndef CONFIG_NEED_MULTIPLE_NODES
6188 * With no DISCONTIG, the global mem_map is just set as node 0's
6190 if (pgdat == NODE_DATA(0)) {
6191 mem_map = NODE_DATA(0)->node_mem_map;
6192 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6193 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6195 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6198 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6201 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6202 unsigned long node_start_pfn, unsigned long *zholes_size)
6204 pg_data_t *pgdat = NODE_DATA(nid);
6205 unsigned long start_pfn = 0;
6206 unsigned long end_pfn = 0;
6208 /* pg_data_t should be reset to zero when it's allocated */
6209 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6211 pgdat->node_id = nid;
6212 pgdat->node_start_pfn = node_start_pfn;
6213 pgdat->per_cpu_nodestats = NULL;
6214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6215 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6216 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6217 (u64)start_pfn << PAGE_SHIFT,
6218 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6220 start_pfn = node_start_pfn;
6222 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6223 zones_size, zholes_size);
6225 alloc_node_mem_map(pgdat);
6226 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6227 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6228 nid, (unsigned long)pgdat,
6229 (unsigned long)pgdat->node_mem_map);
6232 reset_deferred_meminit(pgdat);
6233 free_area_init_core(pgdat);
6236 #ifdef CONFIG_HAVE_MEMBLOCK
6238 * Only struct pages that are backed by physical memory are zeroed and
6239 * initialized by going through __init_single_page(). But, there are some
6240 * struct pages which are reserved in memblock allocator and their fields
6241 * may be accessed (for example page_to_pfn() on some configuration accesses
6242 * flags). We must explicitly zero those struct pages.
6244 void __paginginit zero_resv_unavail(void)
6246 phys_addr_t start, end;
6251 * Loop through ranges that are reserved, but do not have reported
6252 * physical memory backing.
6255 for_each_resv_unavail_range(i, &start, &end) {
6256 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6257 mm_zero_struct_page(pfn_to_page(pfn));
6263 * Struct pages that do not have backing memory. This could be because
6264 * firmware is using some of this memory, or for some other reasons.
6265 * Once memblock is changed so such behaviour is not allowed: i.e.
6266 * list of "reserved" memory must be a subset of list of "memory", then
6267 * this code can be removed.
6270 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6272 #endif /* CONFIG_HAVE_MEMBLOCK */
6274 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6276 #if MAX_NUMNODES > 1
6278 * Figure out the number of possible node ids.
6280 void __init setup_nr_node_ids(void)
6282 unsigned int highest;
6284 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6285 nr_node_ids = highest + 1;
6290 * node_map_pfn_alignment - determine the maximum internode alignment
6292 * This function should be called after node map is populated and sorted.
6293 * It calculates the maximum power of two alignment which can distinguish
6296 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6297 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6298 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6299 * shifted, 1GiB is enough and this function will indicate so.
6301 * This is used to test whether pfn -> nid mapping of the chosen memory
6302 * model has fine enough granularity to avoid incorrect mapping for the
6303 * populated node map.
6305 * Returns the determined alignment in pfn's. 0 if there is no alignment
6306 * requirement (single node).
6308 unsigned long __init node_map_pfn_alignment(void)
6310 unsigned long accl_mask = 0, last_end = 0;
6311 unsigned long start, end, mask;
6315 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6316 if (!start || last_nid < 0 || last_nid == nid) {
6323 * Start with a mask granular enough to pin-point to the
6324 * start pfn and tick off bits one-by-one until it becomes
6325 * too coarse to separate the current node from the last.
6327 mask = ~((1 << __ffs(start)) - 1);
6328 while (mask && last_end <= (start & (mask << 1)))
6331 /* accumulate all internode masks */
6335 /* convert mask to number of pages */
6336 return ~accl_mask + 1;
6339 /* Find the lowest pfn for a node */
6340 static unsigned long __init find_min_pfn_for_node(int nid)
6342 unsigned long min_pfn = ULONG_MAX;
6343 unsigned long start_pfn;
6346 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6347 min_pfn = min(min_pfn, start_pfn);
6349 if (min_pfn == ULONG_MAX) {
6350 pr_warn("Could not find start_pfn for node %d\n", nid);
6358 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6360 * It returns the minimum PFN based on information provided via
6361 * memblock_set_node().
6363 unsigned long __init find_min_pfn_with_active_regions(void)
6365 return find_min_pfn_for_node(MAX_NUMNODES);
6369 * early_calculate_totalpages()
6370 * Sum pages in active regions for movable zone.
6371 * Populate N_MEMORY for calculating usable_nodes.
6373 static unsigned long __init early_calculate_totalpages(void)
6375 unsigned long totalpages = 0;
6376 unsigned long start_pfn, end_pfn;
6379 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6380 unsigned long pages = end_pfn - start_pfn;
6382 totalpages += pages;
6384 node_set_state(nid, N_MEMORY);
6390 * Find the PFN the Movable zone begins in each node. Kernel memory
6391 * is spread evenly between nodes as long as the nodes have enough
6392 * memory. When they don't, some nodes will have more kernelcore than
6395 static void __init find_zone_movable_pfns_for_nodes(void)
6398 unsigned long usable_startpfn;
6399 unsigned long kernelcore_node, kernelcore_remaining;
6400 /* save the state before borrow the nodemask */
6401 nodemask_t saved_node_state = node_states[N_MEMORY];
6402 unsigned long totalpages = early_calculate_totalpages();
6403 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6404 struct memblock_region *r;
6406 /* Need to find movable_zone earlier when movable_node is specified. */
6407 find_usable_zone_for_movable();
6410 * If movable_node is specified, ignore kernelcore and movablecore
6413 if (movable_node_is_enabled()) {
6414 for_each_memblock(memory, r) {
6415 if (!memblock_is_hotpluggable(r))
6420 usable_startpfn = PFN_DOWN(r->base);
6421 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6422 min(usable_startpfn, zone_movable_pfn[nid]) :
6430 * If kernelcore=mirror is specified, ignore movablecore option
6432 if (mirrored_kernelcore) {
6433 bool mem_below_4gb_not_mirrored = false;
6435 for_each_memblock(memory, r) {
6436 if (memblock_is_mirror(r))
6441 usable_startpfn = memblock_region_memory_base_pfn(r);
6443 if (usable_startpfn < 0x100000) {
6444 mem_below_4gb_not_mirrored = true;
6448 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6449 min(usable_startpfn, zone_movable_pfn[nid]) :
6453 if (mem_below_4gb_not_mirrored)
6454 pr_warn("This configuration results in unmirrored kernel memory.");
6460 * If movablecore=nn[KMG] was specified, calculate what size of
6461 * kernelcore that corresponds so that memory usable for
6462 * any allocation type is evenly spread. If both kernelcore
6463 * and movablecore are specified, then the value of kernelcore
6464 * will be used for required_kernelcore if it's greater than
6465 * what movablecore would have allowed.
6467 if (required_movablecore) {
6468 unsigned long corepages;
6471 * Round-up so that ZONE_MOVABLE is at least as large as what
6472 * was requested by the user
6474 required_movablecore =
6475 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6476 required_movablecore = min(totalpages, required_movablecore);
6477 corepages = totalpages - required_movablecore;
6479 required_kernelcore = max(required_kernelcore, corepages);
6483 * If kernelcore was not specified or kernelcore size is larger
6484 * than totalpages, there is no ZONE_MOVABLE.
6486 if (!required_kernelcore || required_kernelcore >= totalpages)
6489 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6490 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6493 /* Spread kernelcore memory as evenly as possible throughout nodes */
6494 kernelcore_node = required_kernelcore / usable_nodes;
6495 for_each_node_state(nid, N_MEMORY) {
6496 unsigned long start_pfn, end_pfn;
6499 * Recalculate kernelcore_node if the division per node
6500 * now exceeds what is necessary to satisfy the requested
6501 * amount of memory for the kernel
6503 if (required_kernelcore < kernelcore_node)
6504 kernelcore_node = required_kernelcore / usable_nodes;
6507 * As the map is walked, we track how much memory is usable
6508 * by the kernel using kernelcore_remaining. When it is
6509 * 0, the rest of the node is usable by ZONE_MOVABLE
6511 kernelcore_remaining = kernelcore_node;
6513 /* Go through each range of PFNs within this node */
6514 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6515 unsigned long size_pages;
6517 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6518 if (start_pfn >= end_pfn)
6521 /* Account for what is only usable for kernelcore */
6522 if (start_pfn < usable_startpfn) {
6523 unsigned long kernel_pages;
6524 kernel_pages = min(end_pfn, usable_startpfn)
6527 kernelcore_remaining -= min(kernel_pages,
6528 kernelcore_remaining);
6529 required_kernelcore -= min(kernel_pages,
6530 required_kernelcore);
6532 /* Continue if range is now fully accounted */
6533 if (end_pfn <= usable_startpfn) {
6536 * Push zone_movable_pfn to the end so
6537 * that if we have to rebalance
6538 * kernelcore across nodes, we will
6539 * not double account here
6541 zone_movable_pfn[nid] = end_pfn;
6544 start_pfn = usable_startpfn;
6548 * The usable PFN range for ZONE_MOVABLE is from
6549 * start_pfn->end_pfn. Calculate size_pages as the
6550 * number of pages used as kernelcore
6552 size_pages = end_pfn - start_pfn;
6553 if (size_pages > kernelcore_remaining)
6554 size_pages = kernelcore_remaining;
6555 zone_movable_pfn[nid] = start_pfn + size_pages;
6558 * Some kernelcore has been met, update counts and
6559 * break if the kernelcore for this node has been
6562 required_kernelcore -= min(required_kernelcore,
6564 kernelcore_remaining -= size_pages;
6565 if (!kernelcore_remaining)
6571 * If there is still required_kernelcore, we do another pass with one
6572 * less node in the count. This will push zone_movable_pfn[nid] further
6573 * along on the nodes that still have memory until kernelcore is
6577 if (usable_nodes && required_kernelcore > usable_nodes)
6581 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6582 for (nid = 0; nid < MAX_NUMNODES; nid++)
6583 zone_movable_pfn[nid] =
6584 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6587 /* restore the node_state */
6588 node_states[N_MEMORY] = saved_node_state;
6591 /* Any regular or high memory on that node ? */
6592 static void check_for_memory(pg_data_t *pgdat, int nid)
6594 enum zone_type zone_type;
6596 if (N_MEMORY == N_NORMAL_MEMORY)
6599 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6600 struct zone *zone = &pgdat->node_zones[zone_type];
6601 if (populated_zone(zone)) {
6602 node_set_state(nid, N_HIGH_MEMORY);
6603 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6604 zone_type <= ZONE_NORMAL)
6605 node_set_state(nid, N_NORMAL_MEMORY);
6612 * free_area_init_nodes - Initialise all pg_data_t and zone data
6613 * @max_zone_pfn: an array of max PFNs for each zone
6615 * This will call free_area_init_node() for each active node in the system.
6616 * Using the page ranges provided by memblock_set_node(), the size of each
6617 * zone in each node and their holes is calculated. If the maximum PFN
6618 * between two adjacent zones match, it is assumed that the zone is empty.
6619 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6620 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6621 * starts where the previous one ended. For example, ZONE_DMA32 starts
6622 * at arch_max_dma_pfn.
6624 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6626 unsigned long start_pfn, end_pfn;
6629 /* Record where the zone boundaries are */
6630 memset(arch_zone_lowest_possible_pfn, 0,
6631 sizeof(arch_zone_lowest_possible_pfn));
6632 memset(arch_zone_highest_possible_pfn, 0,
6633 sizeof(arch_zone_highest_possible_pfn));
6635 start_pfn = find_min_pfn_with_active_regions();
6637 for (i = 0; i < MAX_NR_ZONES; i++) {
6638 if (i == ZONE_MOVABLE)
6641 end_pfn = max(max_zone_pfn[i], start_pfn);
6642 arch_zone_lowest_possible_pfn[i] = start_pfn;
6643 arch_zone_highest_possible_pfn[i] = end_pfn;
6645 start_pfn = end_pfn;
6648 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6649 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6650 find_zone_movable_pfns_for_nodes();
6652 /* Print out the zone ranges */
6653 pr_info("Zone ranges:\n");
6654 for (i = 0; i < MAX_NR_ZONES; i++) {
6655 if (i == ZONE_MOVABLE)
6657 pr_info(" %-8s ", zone_names[i]);
6658 if (arch_zone_lowest_possible_pfn[i] ==
6659 arch_zone_highest_possible_pfn[i])
6662 pr_cont("[mem %#018Lx-%#018Lx]\n",
6663 (u64)arch_zone_lowest_possible_pfn[i]
6665 ((u64)arch_zone_highest_possible_pfn[i]
6666 << PAGE_SHIFT) - 1);
6669 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6670 pr_info("Movable zone start for each node\n");
6671 for (i = 0; i < MAX_NUMNODES; i++) {
6672 if (zone_movable_pfn[i])
6673 pr_info(" Node %d: %#018Lx\n", i,
6674 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6677 /* Print out the early node map */
6678 pr_info("Early memory node ranges\n");
6679 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6680 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6681 (u64)start_pfn << PAGE_SHIFT,
6682 ((u64)end_pfn << PAGE_SHIFT) - 1);
6684 /* Initialise every node */
6685 mminit_verify_pageflags_layout();
6686 setup_nr_node_ids();
6687 for_each_online_node(nid) {
6688 pg_data_t *pgdat = NODE_DATA(nid);
6689 free_area_init_node(nid, NULL,
6690 find_min_pfn_for_node(nid), NULL);
6692 /* Any memory on that node */
6693 if (pgdat->node_present_pages)
6694 node_set_state(nid, N_MEMORY);
6695 check_for_memory(pgdat, nid);
6697 zero_resv_unavail();
6700 static int __init cmdline_parse_core(char *p, unsigned long *core)
6702 unsigned long long coremem;
6706 coremem = memparse(p, &p);
6707 *core = coremem >> PAGE_SHIFT;
6709 /* Paranoid check that UL is enough for the coremem value */
6710 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6716 * kernelcore=size sets the amount of memory for use for allocations that
6717 * cannot be reclaimed or migrated.
6719 static int __init cmdline_parse_kernelcore(char *p)
6721 /* parse kernelcore=mirror */
6722 if (parse_option_str(p, "mirror")) {
6723 mirrored_kernelcore = true;
6727 return cmdline_parse_core(p, &required_kernelcore);
6731 * movablecore=size sets the amount of memory for use for allocations that
6732 * can be reclaimed or migrated.
6734 static int __init cmdline_parse_movablecore(char *p)
6736 return cmdline_parse_core(p, &required_movablecore);
6739 early_param("kernelcore", cmdline_parse_kernelcore);
6740 early_param("movablecore", cmdline_parse_movablecore);
6742 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6744 void adjust_managed_page_count(struct page *page, long count)
6746 spin_lock(&managed_page_count_lock);
6747 page_zone(page)->managed_pages += count;
6748 totalram_pages += count;
6749 #ifdef CONFIG_HIGHMEM
6750 if (PageHighMem(page))
6751 totalhigh_pages += count;
6753 spin_unlock(&managed_page_count_lock);
6755 EXPORT_SYMBOL(adjust_managed_page_count);
6757 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6760 unsigned long pages = 0;
6762 start = (void *)PAGE_ALIGN((unsigned long)start);
6763 end = (void *)((unsigned long)end & PAGE_MASK);
6764 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6765 if ((unsigned int)poison <= 0xFF)
6766 memset(pos, poison, PAGE_SIZE);
6767 free_reserved_page(virt_to_page(pos));
6771 pr_info("Freeing %s memory: %ldK\n",
6772 s, pages << (PAGE_SHIFT - 10));
6776 EXPORT_SYMBOL(free_reserved_area);
6778 #ifdef CONFIG_HIGHMEM
6779 void free_highmem_page(struct page *page)
6781 __free_reserved_page(page);
6783 page_zone(page)->managed_pages++;
6789 void __init mem_init_print_info(const char *str)
6791 unsigned long physpages, codesize, datasize, rosize, bss_size;
6792 unsigned long init_code_size, init_data_size;
6794 physpages = get_num_physpages();
6795 codesize = _etext - _stext;
6796 datasize = _edata - _sdata;
6797 rosize = __end_rodata - __start_rodata;
6798 bss_size = __bss_stop - __bss_start;
6799 init_data_size = __init_end - __init_begin;
6800 init_code_size = _einittext - _sinittext;
6803 * Detect special cases and adjust section sizes accordingly:
6804 * 1) .init.* may be embedded into .data sections
6805 * 2) .init.text.* may be out of [__init_begin, __init_end],
6806 * please refer to arch/tile/kernel/vmlinux.lds.S.
6807 * 3) .rodata.* may be embedded into .text or .data sections.
6809 #define adj_init_size(start, end, size, pos, adj) \
6811 if (start <= pos && pos < end && size > adj) \
6815 adj_init_size(__init_begin, __init_end, init_data_size,
6816 _sinittext, init_code_size);
6817 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6818 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6819 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6820 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6822 #undef adj_init_size
6824 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6825 #ifdef CONFIG_HIGHMEM
6829 nr_free_pages() << (PAGE_SHIFT - 10),
6830 physpages << (PAGE_SHIFT - 10),
6831 codesize >> 10, datasize >> 10, rosize >> 10,
6832 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6833 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6834 totalcma_pages << (PAGE_SHIFT - 10),
6835 #ifdef CONFIG_HIGHMEM
6836 totalhigh_pages << (PAGE_SHIFT - 10),
6838 str ? ", " : "", str ? str : "");
6842 * set_dma_reserve - set the specified number of pages reserved in the first zone
6843 * @new_dma_reserve: The number of pages to mark reserved
6845 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6846 * In the DMA zone, a significant percentage may be consumed by kernel image
6847 * and other unfreeable allocations which can skew the watermarks badly. This
6848 * function may optionally be used to account for unfreeable pages in the
6849 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6850 * smaller per-cpu batchsize.
6852 void __init set_dma_reserve(unsigned long new_dma_reserve)
6854 dma_reserve = new_dma_reserve;
6857 void __init free_area_init(unsigned long *zones_size)
6859 free_area_init_node(0, zones_size,
6860 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6861 zero_resv_unavail();
6864 static int page_alloc_cpu_dead(unsigned int cpu)
6867 lru_add_drain_cpu(cpu);
6871 * Spill the event counters of the dead processor
6872 * into the current processors event counters.
6873 * This artificially elevates the count of the current
6876 vm_events_fold_cpu(cpu);
6879 * Zero the differential counters of the dead processor
6880 * so that the vm statistics are consistent.
6882 * This is only okay since the processor is dead and cannot
6883 * race with what we are doing.
6885 cpu_vm_stats_fold(cpu);
6889 void __init page_alloc_init(void)
6893 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6894 "mm/page_alloc:dead", NULL,
6895 page_alloc_cpu_dead);
6900 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6901 * or min_free_kbytes changes.
6903 static void calculate_totalreserve_pages(void)
6905 struct pglist_data *pgdat;
6906 unsigned long reserve_pages = 0;
6907 enum zone_type i, j;
6909 for_each_online_pgdat(pgdat) {
6911 pgdat->totalreserve_pages = 0;
6913 for (i = 0; i < MAX_NR_ZONES; i++) {
6914 struct zone *zone = pgdat->node_zones + i;
6917 /* Find valid and maximum lowmem_reserve in the zone */
6918 for (j = i; j < MAX_NR_ZONES; j++) {
6919 if (zone->lowmem_reserve[j] > max)
6920 max = zone->lowmem_reserve[j];
6923 /* we treat the high watermark as reserved pages. */
6924 max += high_wmark_pages(zone);
6926 if (max > zone->managed_pages)
6927 max = zone->managed_pages;
6929 pgdat->totalreserve_pages += max;
6931 reserve_pages += max;
6934 totalreserve_pages = reserve_pages;
6938 * setup_per_zone_lowmem_reserve - called whenever
6939 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6940 * has a correct pages reserved value, so an adequate number of
6941 * pages are left in the zone after a successful __alloc_pages().
6943 static void setup_per_zone_lowmem_reserve(void)
6945 struct pglist_data *pgdat;
6946 enum zone_type j, idx;
6948 for_each_online_pgdat(pgdat) {
6949 for (j = 0; j < MAX_NR_ZONES; j++) {
6950 struct zone *zone = pgdat->node_zones + j;
6951 unsigned long managed_pages = zone->managed_pages;
6953 zone->lowmem_reserve[j] = 0;
6957 struct zone *lower_zone;
6961 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6962 sysctl_lowmem_reserve_ratio[idx] = 1;
6964 lower_zone = pgdat->node_zones + idx;
6965 lower_zone->lowmem_reserve[j] = managed_pages /
6966 sysctl_lowmem_reserve_ratio[idx];
6967 managed_pages += lower_zone->managed_pages;
6972 /* update totalreserve_pages */
6973 calculate_totalreserve_pages();
6976 static void __setup_per_zone_wmarks(void)
6978 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6979 unsigned long lowmem_pages = 0;
6981 unsigned long flags;
6983 /* Calculate total number of !ZONE_HIGHMEM pages */
6984 for_each_zone(zone) {
6985 if (!is_highmem(zone))
6986 lowmem_pages += zone->managed_pages;
6989 for_each_zone(zone) {
6992 spin_lock_irqsave(&zone->lock, flags);
6993 tmp = (u64)pages_min * zone->managed_pages;
6994 do_div(tmp, lowmem_pages);
6995 if (is_highmem(zone)) {
6997 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6998 * need highmem pages, so cap pages_min to a small
7001 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7002 * deltas control asynch page reclaim, and so should
7003 * not be capped for highmem.
7005 unsigned long min_pages;
7007 min_pages = zone->managed_pages / 1024;
7008 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7009 zone->watermark[WMARK_MIN] = min_pages;
7012 * If it's a lowmem zone, reserve a number of pages
7013 * proportionate to the zone's size.
7015 zone->watermark[WMARK_MIN] = tmp;
7019 * Set the kswapd watermarks distance according to the
7020 * scale factor in proportion to available memory, but
7021 * ensure a minimum size on small systems.
7023 tmp = max_t(u64, tmp >> 2,
7024 mult_frac(zone->managed_pages,
7025 watermark_scale_factor, 10000));
7027 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7028 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7030 spin_unlock_irqrestore(&zone->lock, flags);
7033 /* update totalreserve_pages */
7034 calculate_totalreserve_pages();
7038 * setup_per_zone_wmarks - called when min_free_kbytes changes
7039 * or when memory is hot-{added|removed}
7041 * Ensures that the watermark[min,low,high] values for each zone are set
7042 * correctly with respect to min_free_kbytes.
7044 void setup_per_zone_wmarks(void)
7046 static DEFINE_SPINLOCK(lock);
7049 __setup_per_zone_wmarks();
7054 * Initialise min_free_kbytes.
7056 * For small machines we want it small (128k min). For large machines
7057 * we want it large (64MB max). But it is not linear, because network
7058 * bandwidth does not increase linearly with machine size. We use
7060 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7061 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7077 int __meminit init_per_zone_wmark_min(void)
7079 unsigned long lowmem_kbytes;
7080 int new_min_free_kbytes;
7082 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7083 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7085 if (new_min_free_kbytes > user_min_free_kbytes) {
7086 min_free_kbytes = new_min_free_kbytes;
7087 if (min_free_kbytes < 128)
7088 min_free_kbytes = 128;
7089 if (min_free_kbytes > 65536)
7090 min_free_kbytes = 65536;
7092 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7093 new_min_free_kbytes, user_min_free_kbytes);
7095 setup_per_zone_wmarks();
7096 refresh_zone_stat_thresholds();
7097 setup_per_zone_lowmem_reserve();
7100 setup_min_unmapped_ratio();
7101 setup_min_slab_ratio();
7106 core_initcall(init_per_zone_wmark_min)
7109 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7110 * that we can call two helper functions whenever min_free_kbytes
7113 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7114 void __user *buffer, size_t *length, loff_t *ppos)
7118 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7123 user_min_free_kbytes = min_free_kbytes;
7124 setup_per_zone_wmarks();
7129 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7130 void __user *buffer, size_t *length, loff_t *ppos)
7134 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7139 setup_per_zone_wmarks();
7145 static void setup_min_unmapped_ratio(void)
7150 for_each_online_pgdat(pgdat)
7151 pgdat->min_unmapped_pages = 0;
7154 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7155 sysctl_min_unmapped_ratio) / 100;
7159 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7160 void __user *buffer, size_t *length, loff_t *ppos)
7164 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7168 setup_min_unmapped_ratio();
7173 static void setup_min_slab_ratio(void)
7178 for_each_online_pgdat(pgdat)
7179 pgdat->min_slab_pages = 0;
7182 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7183 sysctl_min_slab_ratio) / 100;
7186 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7187 void __user *buffer, size_t *length, loff_t *ppos)
7191 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7195 setup_min_slab_ratio();
7202 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7203 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7204 * whenever sysctl_lowmem_reserve_ratio changes.
7206 * The reserve ratio obviously has absolutely no relation with the
7207 * minimum watermarks. The lowmem reserve ratio can only make sense
7208 * if in function of the boot time zone sizes.
7210 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7211 void __user *buffer, size_t *length, loff_t *ppos)
7213 proc_dointvec_minmax(table, write, buffer, length, ppos);
7214 setup_per_zone_lowmem_reserve();
7219 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7220 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7221 * pagelist can have before it gets flushed back to buddy allocator.
7223 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7224 void __user *buffer, size_t *length, loff_t *ppos)
7227 int old_percpu_pagelist_fraction;
7230 mutex_lock(&pcp_batch_high_lock);
7231 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7233 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7234 if (!write || ret < 0)
7237 /* Sanity checking to avoid pcp imbalance */
7238 if (percpu_pagelist_fraction &&
7239 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7240 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7246 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7249 for_each_populated_zone(zone) {
7252 for_each_possible_cpu(cpu)
7253 pageset_set_high_and_batch(zone,
7254 per_cpu_ptr(zone->pageset, cpu));
7257 mutex_unlock(&pcp_batch_high_lock);
7262 int hashdist = HASHDIST_DEFAULT;
7264 static int __init set_hashdist(char *str)
7268 hashdist = simple_strtoul(str, &str, 0);
7271 __setup("hashdist=", set_hashdist);
7274 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7276 * Returns the number of pages that arch has reserved but
7277 * is not known to alloc_large_system_hash().
7279 static unsigned long __init arch_reserved_kernel_pages(void)
7286 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7287 * machines. As memory size is increased the scale is also increased but at
7288 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7289 * quadruples the scale is increased by one, which means the size of hash table
7290 * only doubles, instead of quadrupling as well.
7291 * Because 32-bit systems cannot have large physical memory, where this scaling
7292 * makes sense, it is disabled on such platforms.
7294 #if __BITS_PER_LONG > 32
7295 #define ADAPT_SCALE_BASE (64ul << 30)
7296 #define ADAPT_SCALE_SHIFT 2
7297 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7301 * allocate a large system hash table from bootmem
7302 * - it is assumed that the hash table must contain an exact power-of-2
7303 * quantity of entries
7304 * - limit is the number of hash buckets, not the total allocation size
7306 void *__init alloc_large_system_hash(const char *tablename,
7307 unsigned long bucketsize,
7308 unsigned long numentries,
7311 unsigned int *_hash_shift,
7312 unsigned int *_hash_mask,
7313 unsigned long low_limit,
7314 unsigned long high_limit)
7316 unsigned long long max = high_limit;
7317 unsigned long log2qty, size;
7321 /* allow the kernel cmdline to have a say */
7323 /* round applicable memory size up to nearest megabyte */
7324 numentries = nr_kernel_pages;
7325 numentries -= arch_reserved_kernel_pages();
7327 /* It isn't necessary when PAGE_SIZE >= 1MB */
7328 if (PAGE_SHIFT < 20)
7329 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7331 #if __BITS_PER_LONG > 32
7333 unsigned long adapt;
7335 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7336 adapt <<= ADAPT_SCALE_SHIFT)
7341 /* limit to 1 bucket per 2^scale bytes of low memory */
7342 if (scale > PAGE_SHIFT)
7343 numentries >>= (scale - PAGE_SHIFT);
7345 numentries <<= (PAGE_SHIFT - scale);
7347 /* Make sure we've got at least a 0-order allocation.. */
7348 if (unlikely(flags & HASH_SMALL)) {
7349 /* Makes no sense without HASH_EARLY */
7350 WARN_ON(!(flags & HASH_EARLY));
7351 if (!(numentries >> *_hash_shift)) {
7352 numentries = 1UL << *_hash_shift;
7353 BUG_ON(!numentries);
7355 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7356 numentries = PAGE_SIZE / bucketsize;
7358 numentries = roundup_pow_of_two(numentries);
7360 /* limit allocation size to 1/16 total memory by default */
7362 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7363 do_div(max, bucketsize);
7365 max = min(max, 0x80000000ULL);
7367 if (numentries < low_limit)
7368 numentries = low_limit;
7369 if (numentries > max)
7372 log2qty = ilog2(numentries);
7374 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7376 size = bucketsize << log2qty;
7377 if (flags & HASH_EARLY) {
7378 if (flags & HASH_ZERO)
7379 table = memblock_virt_alloc_nopanic(size, 0);
7381 table = memblock_virt_alloc_raw(size, 0);
7382 } else if (hashdist) {
7383 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7386 * If bucketsize is not a power-of-two, we may free
7387 * some pages at the end of hash table which
7388 * alloc_pages_exact() automatically does
7390 if (get_order(size) < MAX_ORDER) {
7391 table = alloc_pages_exact(size, gfp_flags);
7392 kmemleak_alloc(table, size, 1, gfp_flags);
7395 } while (!table && size > PAGE_SIZE && --log2qty);
7398 panic("Failed to allocate %s hash table\n", tablename);
7400 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7401 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7404 *_hash_shift = log2qty;
7406 *_hash_mask = (1 << log2qty) - 1;
7412 * This function checks whether pageblock includes unmovable pages or not.
7413 * If @count is not zero, it is okay to include less @count unmovable pages
7415 * PageLRU check without isolation or lru_lock could race so that
7416 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7417 * check without lock_page also may miss some movable non-lru pages at
7418 * race condition. So you can't expect this function should be exact.
7420 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7422 bool skip_hwpoisoned_pages)
7424 unsigned long pfn, iter, found;
7427 * For avoiding noise data, lru_add_drain_all() should be called
7428 * If ZONE_MOVABLE, the zone never contains unmovable pages
7430 if (zone_idx(zone) == ZONE_MOVABLE)
7434 * CMA allocations (alloc_contig_range) really need to mark isolate
7435 * CMA pageblocks even when they are not movable in fact so consider
7436 * them movable here.
7438 if (is_migrate_cma(migratetype) &&
7439 is_migrate_cma(get_pageblock_migratetype(page)))
7442 pfn = page_to_pfn(page);
7443 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7444 unsigned long check = pfn + iter;
7446 if (!pfn_valid_within(check))
7449 page = pfn_to_page(check);
7451 if (PageReserved(page))
7455 * Hugepages are not in LRU lists, but they're movable.
7456 * We need not scan over tail pages bacause we don't
7457 * handle each tail page individually in migration.
7459 if (PageHuge(page)) {
7460 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7465 * We can't use page_count without pin a page
7466 * because another CPU can free compound page.
7467 * This check already skips compound tails of THP
7468 * because their page->_refcount is zero at all time.
7470 if (!page_ref_count(page)) {
7471 if (PageBuddy(page))
7472 iter += (1 << page_order(page)) - 1;
7477 * The HWPoisoned page may be not in buddy system, and
7478 * page_count() is not 0.
7480 if (skip_hwpoisoned_pages && PageHWPoison(page))
7483 if (__PageMovable(page))
7489 * If there are RECLAIMABLE pages, we need to check
7490 * it. But now, memory offline itself doesn't call
7491 * shrink_node_slabs() and it still to be fixed.
7494 * If the page is not RAM, page_count()should be 0.
7495 * we don't need more check. This is an _used_ not-movable page.
7497 * The problematic thing here is PG_reserved pages. PG_reserved
7498 * is set to both of a memory hole page and a _used_ kernel
7507 bool is_pageblock_removable_nolock(struct page *page)
7513 * We have to be careful here because we are iterating over memory
7514 * sections which are not zone aware so we might end up outside of
7515 * the zone but still within the section.
7516 * We have to take care about the node as well. If the node is offline
7517 * its NODE_DATA will be NULL - see page_zone.
7519 if (!node_online(page_to_nid(page)))
7522 zone = page_zone(page);
7523 pfn = page_to_pfn(page);
7524 if (!zone_spans_pfn(zone, pfn))
7527 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7530 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7532 static unsigned long pfn_max_align_down(unsigned long pfn)
7534 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7535 pageblock_nr_pages) - 1);
7538 static unsigned long pfn_max_align_up(unsigned long pfn)
7540 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7541 pageblock_nr_pages));
7544 /* [start, end) must belong to a single zone. */
7545 static int __alloc_contig_migrate_range(struct compact_control *cc,
7546 unsigned long start, unsigned long end)
7548 /* This function is based on compact_zone() from compaction.c. */
7549 unsigned long nr_reclaimed;
7550 unsigned long pfn = start;
7551 unsigned int tries = 0;
7556 while (pfn < end || !list_empty(&cc->migratepages)) {
7557 if (fatal_signal_pending(current)) {
7562 if (list_empty(&cc->migratepages)) {
7563 cc->nr_migratepages = 0;
7564 pfn = isolate_migratepages_range(cc, pfn, end);
7570 } else if (++tries == 5) {
7571 ret = ret < 0 ? ret : -EBUSY;
7575 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7577 cc->nr_migratepages -= nr_reclaimed;
7579 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7580 NULL, 0, cc->mode, MR_CMA);
7583 putback_movable_pages(&cc->migratepages);
7590 * alloc_contig_range() -- tries to allocate given range of pages
7591 * @start: start PFN to allocate
7592 * @end: one-past-the-last PFN to allocate
7593 * @migratetype: migratetype of the underlaying pageblocks (either
7594 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7595 * in range must have the same migratetype and it must
7596 * be either of the two.
7597 * @gfp_mask: GFP mask to use during compaction
7599 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7600 * aligned, however it's the caller's responsibility to guarantee that
7601 * we are the only thread that changes migrate type of pageblocks the
7604 * The PFN range must belong to a single zone.
7606 * Returns zero on success or negative error code. On success all
7607 * pages which PFN is in [start, end) are allocated for the caller and
7608 * need to be freed with free_contig_range().
7610 int alloc_contig_range(unsigned long start, unsigned long end,
7611 unsigned migratetype, gfp_t gfp_mask)
7613 unsigned long outer_start, outer_end;
7617 struct compact_control cc = {
7618 .nr_migratepages = 0,
7620 .zone = page_zone(pfn_to_page(start)),
7621 .mode = MIGRATE_SYNC,
7622 .ignore_skip_hint = true,
7623 .gfp_mask = current_gfp_context(gfp_mask),
7625 INIT_LIST_HEAD(&cc.migratepages);
7628 * What we do here is we mark all pageblocks in range as
7629 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7630 * have different sizes, and due to the way page allocator
7631 * work, we align the range to biggest of the two pages so
7632 * that page allocator won't try to merge buddies from
7633 * different pageblocks and change MIGRATE_ISOLATE to some
7634 * other migration type.
7636 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7637 * migrate the pages from an unaligned range (ie. pages that
7638 * we are interested in). This will put all the pages in
7639 * range back to page allocator as MIGRATE_ISOLATE.
7641 * When this is done, we take the pages in range from page
7642 * allocator removing them from the buddy system. This way
7643 * page allocator will never consider using them.
7645 * This lets us mark the pageblocks back as
7646 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7647 * aligned range but not in the unaligned, original range are
7648 * put back to page allocator so that buddy can use them.
7651 ret = start_isolate_page_range(pfn_max_align_down(start),
7652 pfn_max_align_up(end), migratetype,
7658 * In case of -EBUSY, we'd like to know which page causes problem.
7659 * So, just fall through. We will check it in test_pages_isolated().
7661 ret = __alloc_contig_migrate_range(&cc, start, end);
7662 if (ret && ret != -EBUSY)
7666 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7667 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7668 * more, all pages in [start, end) are free in page allocator.
7669 * What we are going to do is to allocate all pages from
7670 * [start, end) (that is remove them from page allocator).
7672 * The only problem is that pages at the beginning and at the
7673 * end of interesting range may be not aligned with pages that
7674 * page allocator holds, ie. they can be part of higher order
7675 * pages. Because of this, we reserve the bigger range and
7676 * once this is done free the pages we are not interested in.
7678 * We don't have to hold zone->lock here because the pages are
7679 * isolated thus they won't get removed from buddy.
7682 lru_add_drain_all();
7683 drain_all_pages(cc.zone);
7686 outer_start = start;
7687 while (!PageBuddy(pfn_to_page(outer_start))) {
7688 if (++order >= MAX_ORDER) {
7689 outer_start = start;
7692 outer_start &= ~0UL << order;
7695 if (outer_start != start) {
7696 order = page_order(pfn_to_page(outer_start));
7699 * outer_start page could be small order buddy page and
7700 * it doesn't include start page. Adjust outer_start
7701 * in this case to report failed page properly
7702 * on tracepoint in test_pages_isolated()
7704 if (outer_start + (1UL << order) <= start)
7705 outer_start = start;
7708 /* Make sure the range is really isolated. */
7709 if (test_pages_isolated(outer_start, end, false)) {
7710 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7711 __func__, outer_start, end);
7716 /* Grab isolated pages from freelists. */
7717 outer_end = isolate_freepages_range(&cc, outer_start, end);
7723 /* Free head and tail (if any) */
7724 if (start != outer_start)
7725 free_contig_range(outer_start, start - outer_start);
7726 if (end != outer_end)
7727 free_contig_range(end, outer_end - end);
7730 undo_isolate_page_range(pfn_max_align_down(start),
7731 pfn_max_align_up(end), migratetype);
7735 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7737 unsigned int count = 0;
7739 for (; nr_pages--; pfn++) {
7740 struct page *page = pfn_to_page(pfn);
7742 count += page_count(page) != 1;
7745 WARN(count != 0, "%d pages are still in use!\n", count);
7749 #ifdef CONFIG_MEMORY_HOTPLUG
7751 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7752 * page high values need to be recalulated.
7754 void __meminit zone_pcp_update(struct zone *zone)
7757 mutex_lock(&pcp_batch_high_lock);
7758 for_each_possible_cpu(cpu)
7759 pageset_set_high_and_batch(zone,
7760 per_cpu_ptr(zone->pageset, cpu));
7761 mutex_unlock(&pcp_batch_high_lock);
7765 void zone_pcp_reset(struct zone *zone)
7767 unsigned long flags;
7769 struct per_cpu_pageset *pset;
7771 /* avoid races with drain_pages() */
7772 local_irq_save(flags);
7773 if (zone->pageset != &boot_pageset) {
7774 for_each_online_cpu(cpu) {
7775 pset = per_cpu_ptr(zone->pageset, cpu);
7776 drain_zonestat(zone, pset);
7778 free_percpu(zone->pageset);
7779 zone->pageset = &boot_pageset;
7781 local_irq_restore(flags);
7784 #ifdef CONFIG_MEMORY_HOTREMOVE
7786 * All pages in the range must be in a single zone and isolated
7787 * before calling this.
7790 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7794 unsigned int order, i;
7796 unsigned long flags;
7797 /* find the first valid pfn */
7798 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7803 offline_mem_sections(pfn, end_pfn);
7804 zone = page_zone(pfn_to_page(pfn));
7805 spin_lock_irqsave(&zone->lock, flags);
7807 while (pfn < end_pfn) {
7808 if (!pfn_valid(pfn)) {
7812 page = pfn_to_page(pfn);
7814 * The HWPoisoned page may be not in buddy system, and
7815 * page_count() is not 0.
7817 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7819 SetPageReserved(page);
7823 BUG_ON(page_count(page));
7824 BUG_ON(!PageBuddy(page));
7825 order = page_order(page);
7826 #ifdef CONFIG_DEBUG_VM
7827 pr_info("remove from free list %lx %d %lx\n",
7828 pfn, 1 << order, end_pfn);
7830 list_del(&page->lru);
7831 rmv_page_order(page);
7832 zone->free_area[order].nr_free--;
7833 for (i = 0; i < (1 << order); i++)
7834 SetPageReserved((page+i));
7835 pfn += (1 << order);
7837 spin_unlock_irqrestore(&zone->lock, flags);
7841 bool is_free_buddy_page(struct page *page)
7843 struct zone *zone = page_zone(page);
7844 unsigned long pfn = page_to_pfn(page);
7845 unsigned long flags;
7848 spin_lock_irqsave(&zone->lock, flags);
7849 for (order = 0; order < MAX_ORDER; order++) {
7850 struct page *page_head = page - (pfn & ((1 << order) - 1));
7852 if (PageBuddy(page_head) && page_order(page_head) >= order)
7855 spin_unlock_irqrestore(&zone->lock, flags);
7857 return order < MAX_ORDER;