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 set_page_links(page, zone, nid, pfn);
1172 init_page_count(page);
1173 page_mapcount_reset(page);
1174 page_cpupid_reset_last(page);
1176 INIT_LIST_HEAD(&page->lru);
1177 #ifdef WANT_PAGE_VIRTUAL
1178 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1179 if (!is_highmem_idx(zone))
1180 set_page_address(page, __va(pfn << PAGE_SHIFT));
1184 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1187 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1190 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1191 static void __meminit init_reserved_page(unsigned long pfn)
1196 if (!early_page_uninitialised(pfn))
1199 nid = early_pfn_to_nid(pfn);
1200 pgdat = NODE_DATA(nid);
1202 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1203 struct zone *zone = &pgdat->node_zones[zid];
1205 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1208 __init_single_pfn(pfn, zid, nid);
1211 static inline void init_reserved_page(unsigned long pfn)
1214 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1217 * Initialised pages do not have PageReserved set. This function is
1218 * called for each range allocated by the bootmem allocator and
1219 * marks the pages PageReserved. The remaining valid pages are later
1220 * sent to the buddy page allocator.
1222 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224 unsigned long start_pfn = PFN_DOWN(start);
1225 unsigned long end_pfn = PFN_UP(end);
1227 for (; start_pfn < end_pfn; start_pfn++) {
1228 if (pfn_valid(start_pfn)) {
1229 struct page *page = pfn_to_page(start_pfn);
1231 init_reserved_page(start_pfn);
1233 /* Avoid false-positive PageTail() */
1234 INIT_LIST_HEAD(&page->lru);
1236 SetPageReserved(page);
1241 static void __free_pages_ok(struct page *page, unsigned int order)
1243 unsigned long flags;
1245 unsigned long pfn = page_to_pfn(page);
1247 if (!free_pages_prepare(page, order, true))
1250 migratetype = get_pfnblock_migratetype(page, pfn);
1251 local_irq_save(flags);
1252 __count_vm_events(PGFREE, 1 << order);
1253 free_one_page(page_zone(page), page, pfn, order, migratetype);
1254 local_irq_restore(flags);
1257 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259 unsigned int nr_pages = 1 << order;
1260 struct page *p = page;
1264 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 __ClearPageReserved(p);
1267 set_page_count(p, 0);
1269 __ClearPageReserved(p);
1270 set_page_count(p, 0);
1272 page_zone(page)->managed_pages += nr_pages;
1273 set_page_refcounted(page);
1274 __free_pages(page, order);
1277 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1278 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282 int __meminit early_pfn_to_nid(unsigned long pfn)
1284 static DEFINE_SPINLOCK(early_pfn_lock);
1287 spin_lock(&early_pfn_lock);
1288 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 nid = first_online_node;
1291 spin_unlock(&early_pfn_lock);
1297 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1298 static inline bool __meminit __maybe_unused
1299 meminit_pfn_in_nid(unsigned long pfn, int node,
1300 struct mminit_pfnnid_cache *state)
1304 nid = __early_pfn_to_nid(pfn, state);
1305 if (nid >= 0 && nid != node)
1310 /* Only safe to use early in boot when initialisation is single-threaded */
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1318 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1322 static inline bool __meminit __maybe_unused
1323 meminit_pfn_in_nid(unsigned long pfn, int node,
1324 struct mminit_pfnnid_cache *state)
1331 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1334 if (early_page_uninitialised(pfn))
1336 return __free_pages_boot_core(page, order);
1340 * Check that the whole (or subset of) a pageblock given by the interval of
1341 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1342 * with the migration of free compaction scanner. The scanners then need to
1343 * use only pfn_valid_within() check for arches that allow holes within
1346 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 * It's possible on some configurations to have a setup like node0 node1 node0
1349 * i.e. it's possible that all pages within a zones range of pages do not
1350 * belong to a single zone. We assume that a border between node0 and node1
1351 * can occur within a single pageblock, but not a node0 node1 node0
1352 * interleaving within a single pageblock. It is therefore sufficient to check
1353 * the first and last page of a pageblock and avoid checking each individual
1354 * page in a pageblock.
1356 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1357 unsigned long end_pfn, struct zone *zone)
1359 struct page *start_page;
1360 struct page *end_page;
1362 /* end_pfn is one past the range we are checking */
1365 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1368 start_page = pfn_to_online_page(start_pfn);
1372 if (page_zone(start_page) != zone)
1375 end_page = pfn_to_page(end_pfn);
1377 /* This gives a shorter code than deriving page_zone(end_page) */
1378 if (page_zone_id(start_page) != page_zone_id(end_page))
1384 void set_zone_contiguous(struct zone *zone)
1386 unsigned long block_start_pfn = zone->zone_start_pfn;
1387 unsigned long block_end_pfn;
1389 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1390 for (; block_start_pfn < zone_end_pfn(zone);
1391 block_start_pfn = block_end_pfn,
1392 block_end_pfn += pageblock_nr_pages) {
1394 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396 if (!__pageblock_pfn_to_page(block_start_pfn,
1397 block_end_pfn, zone))
1401 /* We confirm that there is no hole */
1402 zone->contiguous = true;
1405 void clear_zone_contiguous(struct zone *zone)
1407 zone->contiguous = false;
1410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411 static void __init deferred_free_range(struct page *page,
1412 unsigned long pfn, int nr_pages)
1419 /* Free a large naturally-aligned chunk if possible */
1420 if (nr_pages == pageblock_nr_pages &&
1421 (pfn & (pageblock_nr_pages - 1)) == 0) {
1422 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1423 __free_pages_boot_core(page, pageblock_order);
1427 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1428 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1429 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1430 __free_pages_boot_core(page, 0);
1434 /* Completion tracking for deferred_init_memmap() threads */
1435 static atomic_t pgdat_init_n_undone __initdata;
1436 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1438 static inline void __init pgdat_init_report_one_done(void)
1440 if (atomic_dec_and_test(&pgdat_init_n_undone))
1441 complete(&pgdat_init_all_done_comp);
1444 /* Initialise remaining memory on a node */
1445 static int __init deferred_init_memmap(void *data)
1447 pg_data_t *pgdat = data;
1448 int nid = pgdat->node_id;
1449 struct mminit_pfnnid_cache nid_init_state = { };
1450 unsigned long start = jiffies;
1451 unsigned long nr_pages = 0;
1452 unsigned long walk_start, walk_end;
1455 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1456 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1458 if (first_init_pfn == ULONG_MAX) {
1459 pgdat_init_report_one_done();
1463 /* Bind memory initialisation thread to a local node if possible */
1464 if (!cpumask_empty(cpumask))
1465 set_cpus_allowed_ptr(current, cpumask);
1467 /* Sanity check boundaries */
1468 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1469 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1470 pgdat->first_deferred_pfn = ULONG_MAX;
1472 /* Only the highest zone is deferred so find it */
1473 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1474 zone = pgdat->node_zones + zid;
1475 if (first_init_pfn < zone_end_pfn(zone))
1479 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1480 unsigned long pfn, end_pfn;
1481 struct page *page = NULL;
1482 struct page *free_base_page = NULL;
1483 unsigned long free_base_pfn = 0;
1486 end_pfn = min(walk_end, zone_end_pfn(zone));
1487 pfn = first_init_pfn;
1488 if (pfn < walk_start)
1490 if (pfn < zone->zone_start_pfn)
1491 pfn = zone->zone_start_pfn;
1493 for (; pfn < end_pfn; pfn++) {
1494 if (!pfn_valid_within(pfn))
1498 * Ensure pfn_valid is checked every
1499 * pageblock_nr_pages for memory holes
1501 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1502 if (!pfn_valid(pfn)) {
1508 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1513 /* Minimise pfn page lookups and scheduler checks */
1514 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1517 nr_pages += nr_to_free;
1518 deferred_free_range(free_base_page,
1519 free_base_pfn, nr_to_free);
1520 free_base_page = NULL;
1521 free_base_pfn = nr_to_free = 0;
1523 page = pfn_to_page(pfn);
1528 VM_BUG_ON(page_zone(page) != zone);
1532 __init_single_page(page, pfn, zid, nid);
1533 if (!free_base_page) {
1534 free_base_page = page;
1535 free_base_pfn = pfn;
1540 /* Where possible, batch up pages for a single free */
1543 /* Free the current block of pages to allocator */
1544 nr_pages += nr_to_free;
1545 deferred_free_range(free_base_page, free_base_pfn,
1547 free_base_page = NULL;
1548 free_base_pfn = nr_to_free = 0;
1550 /* Free the last block of pages to allocator */
1551 nr_pages += nr_to_free;
1552 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1554 first_init_pfn = max(end_pfn, first_init_pfn);
1557 /* Sanity check that the next zone really is unpopulated */
1558 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1560 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1561 jiffies_to_msecs(jiffies - start));
1563 pgdat_init_report_one_done();
1566 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1568 void __init page_alloc_init_late(void)
1572 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1575 /* There will be num_node_state(N_MEMORY) threads */
1576 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1577 for_each_node_state(nid, N_MEMORY) {
1578 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1581 /* Block until all are initialised */
1582 wait_for_completion(&pgdat_init_all_done_comp);
1584 /* Reinit limits that are based on free pages after the kernel is up */
1585 files_maxfiles_init();
1587 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1588 /* Discard memblock private memory */
1592 for_each_populated_zone(zone)
1593 set_zone_contiguous(zone);
1597 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1598 void __init init_cma_reserved_pageblock(struct page *page)
1600 unsigned i = pageblock_nr_pages;
1601 struct page *p = page;
1604 __ClearPageReserved(p);
1605 set_page_count(p, 0);
1608 set_pageblock_migratetype(page, MIGRATE_CMA);
1610 if (pageblock_order >= MAX_ORDER) {
1611 i = pageblock_nr_pages;
1614 set_page_refcounted(p);
1615 __free_pages(p, MAX_ORDER - 1);
1616 p += MAX_ORDER_NR_PAGES;
1617 } while (i -= MAX_ORDER_NR_PAGES);
1619 set_page_refcounted(page);
1620 __free_pages(page, pageblock_order);
1623 adjust_managed_page_count(page, pageblock_nr_pages);
1628 * The order of subdivision here is critical for the IO subsystem.
1629 * Please do not alter this order without good reasons and regression
1630 * testing. Specifically, as large blocks of memory are subdivided,
1631 * the order in which smaller blocks are delivered depends on the order
1632 * they're subdivided in this function. This is the primary factor
1633 * influencing the order in which pages are delivered to the IO
1634 * subsystem according to empirical testing, and this is also justified
1635 * by considering the behavior of a buddy system containing a single
1636 * large block of memory acted on by a series of small allocations.
1637 * This behavior is a critical factor in sglist merging's success.
1641 static inline void expand(struct zone *zone, struct page *page,
1642 int low, int high, struct free_area *area,
1645 unsigned long size = 1 << high;
1647 while (high > low) {
1651 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1654 * Mark as guard pages (or page), that will allow to
1655 * merge back to allocator when buddy will be freed.
1656 * Corresponding page table entries will not be touched,
1657 * pages will stay not present in virtual address space
1659 if (set_page_guard(zone, &page[size], high, migratetype))
1662 list_add(&page[size].lru, &area->free_list[migratetype]);
1664 set_page_order(&page[size], high);
1668 static void check_new_page_bad(struct page *page)
1670 const char *bad_reason = NULL;
1671 unsigned long bad_flags = 0;
1673 if (unlikely(atomic_read(&page->_mapcount) != -1))
1674 bad_reason = "nonzero mapcount";
1675 if (unlikely(page->mapping != NULL))
1676 bad_reason = "non-NULL mapping";
1677 if (unlikely(page_ref_count(page) != 0))
1678 bad_reason = "nonzero _count";
1679 if (unlikely(page->flags & __PG_HWPOISON)) {
1680 bad_reason = "HWPoisoned (hardware-corrupted)";
1681 bad_flags = __PG_HWPOISON;
1682 /* Don't complain about hwpoisoned pages */
1683 page_mapcount_reset(page); /* remove PageBuddy */
1686 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1687 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1688 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1691 if (unlikely(page->mem_cgroup))
1692 bad_reason = "page still charged to cgroup";
1694 bad_page(page, bad_reason, bad_flags);
1698 * This page is about to be returned from the page allocator
1700 static inline int check_new_page(struct page *page)
1702 if (likely(page_expected_state(page,
1703 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1706 check_new_page_bad(page);
1710 static inline bool free_pages_prezeroed(void)
1712 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1713 page_poisoning_enabled();
1716 #ifdef CONFIG_DEBUG_VM
1717 static bool check_pcp_refill(struct page *page)
1722 static bool check_new_pcp(struct page *page)
1724 return check_new_page(page);
1727 static bool check_pcp_refill(struct page *page)
1729 return check_new_page(page);
1731 static bool check_new_pcp(struct page *page)
1735 #endif /* CONFIG_DEBUG_VM */
1737 static bool check_new_pages(struct page *page, unsigned int order)
1740 for (i = 0; i < (1 << order); i++) {
1741 struct page *p = page + i;
1743 if (unlikely(check_new_page(p)))
1750 inline void post_alloc_hook(struct page *page, unsigned int order,
1753 set_page_private(page, 0);
1754 set_page_refcounted(page);
1756 arch_alloc_page(page, order);
1757 kernel_map_pages(page, 1 << order, 1);
1758 kernel_poison_pages(page, 1 << order, 1);
1759 kasan_alloc_pages(page, order);
1760 set_page_owner(page, order, gfp_flags);
1763 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1764 unsigned int alloc_flags)
1768 post_alloc_hook(page, order, gfp_flags);
1770 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1771 for (i = 0; i < (1 << order); i++)
1772 clear_highpage(page + i);
1774 if (order && (gfp_flags & __GFP_COMP))
1775 prep_compound_page(page, order);
1778 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1779 * allocate the page. The expectation is that the caller is taking
1780 * steps that will free more memory. The caller should avoid the page
1781 * being used for !PFMEMALLOC purposes.
1783 if (alloc_flags & ALLOC_NO_WATERMARKS)
1784 set_page_pfmemalloc(page);
1786 clear_page_pfmemalloc(page);
1790 * Go through the free lists for the given migratetype and remove
1791 * the smallest available page from the freelists
1794 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1797 unsigned int current_order;
1798 struct free_area *area;
1801 /* Find a page of the appropriate size in the preferred list */
1802 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1803 area = &(zone->free_area[current_order]);
1804 page = list_first_entry_or_null(&area->free_list[migratetype],
1808 list_del(&page->lru);
1809 rmv_page_order(page);
1811 expand(zone, page, order, current_order, area, migratetype);
1812 set_pcppage_migratetype(page, migratetype);
1821 * This array describes the order lists are fallen back to when
1822 * the free lists for the desirable migrate type are depleted
1824 static int fallbacks[MIGRATE_TYPES][4] = {
1825 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1826 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1827 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1829 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1831 #ifdef CONFIG_MEMORY_ISOLATION
1832 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1837 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1840 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1843 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1844 unsigned int order) { return NULL; }
1848 * Move the free pages in a range to the free lists of the requested type.
1849 * Note that start_page and end_pages are not aligned on a pageblock
1850 * boundary. If alignment is required, use move_freepages_block()
1852 static int move_freepages(struct zone *zone,
1853 struct page *start_page, struct page *end_page,
1854 int migratetype, int *num_movable)
1858 int pages_moved = 0;
1860 #ifndef CONFIG_HOLES_IN_ZONE
1862 * page_zone is not safe to call in this context when
1863 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1864 * anyway as we check zone boundaries in move_freepages_block().
1865 * Remove at a later date when no bug reports exist related to
1866 * grouping pages by mobility
1868 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1874 for (page = start_page; page <= end_page;) {
1875 if (!pfn_valid_within(page_to_pfn(page))) {
1880 /* Make sure we are not inadvertently changing nodes */
1881 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1883 if (!PageBuddy(page)) {
1885 * We assume that pages that could be isolated for
1886 * migration are movable. But we don't actually try
1887 * isolating, as that would be expensive.
1890 (PageLRU(page) || __PageMovable(page)))
1897 order = page_order(page);
1898 list_move(&page->lru,
1899 &zone->free_area[order].free_list[migratetype]);
1901 pages_moved += 1 << order;
1907 int move_freepages_block(struct zone *zone, struct page *page,
1908 int migratetype, int *num_movable)
1910 unsigned long start_pfn, end_pfn;
1911 struct page *start_page, *end_page;
1913 start_pfn = page_to_pfn(page);
1914 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1915 start_page = pfn_to_page(start_pfn);
1916 end_page = start_page + pageblock_nr_pages - 1;
1917 end_pfn = start_pfn + pageblock_nr_pages - 1;
1919 /* Do not cross zone boundaries */
1920 if (!zone_spans_pfn(zone, start_pfn))
1922 if (!zone_spans_pfn(zone, end_pfn))
1925 return move_freepages(zone, start_page, end_page, migratetype,
1929 static void change_pageblock_range(struct page *pageblock_page,
1930 int start_order, int migratetype)
1932 int nr_pageblocks = 1 << (start_order - pageblock_order);
1934 while (nr_pageblocks--) {
1935 set_pageblock_migratetype(pageblock_page, migratetype);
1936 pageblock_page += pageblock_nr_pages;
1941 * When we are falling back to another migratetype during allocation, try to
1942 * steal extra free pages from the same pageblocks to satisfy further
1943 * allocations, instead of polluting multiple pageblocks.
1945 * If we are stealing a relatively large buddy page, it is likely there will
1946 * be more free pages in the pageblock, so try to steal them all. For
1947 * reclaimable and unmovable allocations, we steal regardless of page size,
1948 * as fragmentation caused by those allocations polluting movable pageblocks
1949 * is worse than movable allocations stealing from unmovable and reclaimable
1952 static bool can_steal_fallback(unsigned int order, int start_mt)
1955 * Leaving this order check is intended, although there is
1956 * relaxed order check in next check. The reason is that
1957 * we can actually steal whole pageblock if this condition met,
1958 * but, below check doesn't guarantee it and that is just heuristic
1959 * so could be changed anytime.
1961 if (order >= pageblock_order)
1964 if (order >= pageblock_order / 2 ||
1965 start_mt == MIGRATE_RECLAIMABLE ||
1966 start_mt == MIGRATE_UNMOVABLE ||
1967 page_group_by_mobility_disabled)
1974 * This function implements actual steal behaviour. If order is large enough,
1975 * we can steal whole pageblock. If not, we first move freepages in this
1976 * pageblock to our migratetype and determine how many already-allocated pages
1977 * are there in the pageblock with a compatible migratetype. If at least half
1978 * of pages are free or compatible, we can change migratetype of the pageblock
1979 * itself, so pages freed in the future will be put on the correct free list.
1981 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1982 int start_type, bool whole_block)
1984 unsigned int current_order = page_order(page);
1985 struct free_area *area;
1986 int free_pages, movable_pages, alike_pages;
1989 old_block_type = get_pageblock_migratetype(page);
1992 * This can happen due to races and we want to prevent broken
1993 * highatomic accounting.
1995 if (is_migrate_highatomic(old_block_type))
1998 /* Take ownership for orders >= pageblock_order */
1999 if (current_order >= pageblock_order) {
2000 change_pageblock_range(page, current_order, start_type);
2004 /* We are not allowed to try stealing from the whole block */
2008 free_pages = move_freepages_block(zone, page, start_type,
2011 * Determine how many pages are compatible with our allocation.
2012 * For movable allocation, it's the number of movable pages which
2013 * we just obtained. For other types it's a bit more tricky.
2015 if (start_type == MIGRATE_MOVABLE) {
2016 alike_pages = movable_pages;
2019 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2020 * to MOVABLE pageblock, consider all non-movable pages as
2021 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2022 * vice versa, be conservative since we can't distinguish the
2023 * exact migratetype of non-movable pages.
2025 if (old_block_type == MIGRATE_MOVABLE)
2026 alike_pages = pageblock_nr_pages
2027 - (free_pages + movable_pages);
2032 /* moving whole block can fail due to zone boundary conditions */
2037 * If a sufficient number of pages in the block are either free or of
2038 * comparable migratability as our allocation, claim the whole block.
2040 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2041 page_group_by_mobility_disabled)
2042 set_pageblock_migratetype(page, start_type);
2047 area = &zone->free_area[current_order];
2048 list_move(&page->lru, &area->free_list[start_type]);
2052 * Check whether there is a suitable fallback freepage with requested order.
2053 * If only_stealable is true, this function returns fallback_mt only if
2054 * we can steal other freepages all together. This would help to reduce
2055 * fragmentation due to mixed migratetype pages in one pageblock.
2057 int find_suitable_fallback(struct free_area *area, unsigned int order,
2058 int migratetype, bool only_stealable, bool *can_steal)
2063 if (area->nr_free == 0)
2068 fallback_mt = fallbacks[migratetype][i];
2069 if (fallback_mt == MIGRATE_TYPES)
2072 if (list_empty(&area->free_list[fallback_mt]))
2075 if (can_steal_fallback(order, migratetype))
2078 if (!only_stealable)
2089 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2090 * there are no empty page blocks that contain a page with a suitable order
2092 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2093 unsigned int alloc_order)
2096 unsigned long max_managed, flags;
2099 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2100 * Check is race-prone but harmless.
2102 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2103 if (zone->nr_reserved_highatomic >= max_managed)
2106 spin_lock_irqsave(&zone->lock, flags);
2108 /* Recheck the nr_reserved_highatomic limit under the lock */
2109 if (zone->nr_reserved_highatomic >= max_managed)
2113 mt = get_pageblock_migratetype(page);
2114 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2115 && !is_migrate_cma(mt)) {
2116 zone->nr_reserved_highatomic += pageblock_nr_pages;
2117 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2118 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2122 spin_unlock_irqrestore(&zone->lock, flags);
2126 * Used when an allocation is about to fail under memory pressure. This
2127 * potentially hurts the reliability of high-order allocations when under
2128 * intense memory pressure but failed atomic allocations should be easier
2129 * to recover from than an OOM.
2131 * If @force is true, try to unreserve a pageblock even though highatomic
2132 * pageblock is exhausted.
2134 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2137 struct zonelist *zonelist = ac->zonelist;
2138 unsigned long flags;
2145 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2148 * Preserve at least one pageblock unless memory pressure
2151 if (!force && zone->nr_reserved_highatomic <=
2155 spin_lock_irqsave(&zone->lock, flags);
2156 for (order = 0; order < MAX_ORDER; order++) {
2157 struct free_area *area = &(zone->free_area[order]);
2159 page = list_first_entry_or_null(
2160 &area->free_list[MIGRATE_HIGHATOMIC],
2166 * In page freeing path, migratetype change is racy so
2167 * we can counter several free pages in a pageblock
2168 * in this loop althoug we changed the pageblock type
2169 * from highatomic to ac->migratetype. So we should
2170 * adjust the count once.
2172 if (is_migrate_highatomic_page(page)) {
2174 * It should never happen but changes to
2175 * locking could inadvertently allow a per-cpu
2176 * drain to add pages to MIGRATE_HIGHATOMIC
2177 * while unreserving so be safe and watch for
2180 zone->nr_reserved_highatomic -= min(
2182 zone->nr_reserved_highatomic);
2186 * Convert to ac->migratetype and avoid the normal
2187 * pageblock stealing heuristics. Minimally, the caller
2188 * is doing the work and needs the pages. More
2189 * importantly, if the block was always converted to
2190 * MIGRATE_UNMOVABLE or another type then the number
2191 * of pageblocks that cannot be completely freed
2194 set_pageblock_migratetype(page, ac->migratetype);
2195 ret = move_freepages_block(zone, page, ac->migratetype,
2198 spin_unlock_irqrestore(&zone->lock, flags);
2202 spin_unlock_irqrestore(&zone->lock, flags);
2209 * Try finding a free buddy page on the fallback list and put it on the free
2210 * list of requested migratetype, possibly along with other pages from the same
2211 * block, depending on fragmentation avoidance heuristics. Returns true if
2212 * fallback was found so that __rmqueue_smallest() can grab it.
2214 * The use of signed ints for order and current_order is a deliberate
2215 * deviation from the rest of this file, to make the for loop
2216 * condition simpler.
2219 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2221 struct free_area *area;
2228 * Find the largest available free page in the other list. This roughly
2229 * approximates finding the pageblock with the most free pages, which
2230 * would be too costly to do exactly.
2232 for (current_order = MAX_ORDER - 1; current_order >= order;
2234 area = &(zone->free_area[current_order]);
2235 fallback_mt = find_suitable_fallback(area, current_order,
2236 start_migratetype, false, &can_steal);
2237 if (fallback_mt == -1)
2241 * We cannot steal all free pages from the pageblock and the
2242 * requested migratetype is movable. In that case it's better to
2243 * steal and split the smallest available page instead of the
2244 * largest available page, because even if the next movable
2245 * allocation falls back into a different pageblock than this
2246 * one, it won't cause permanent fragmentation.
2248 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2249 && current_order > order)
2258 for (current_order = order; current_order < MAX_ORDER;
2260 area = &(zone->free_area[current_order]);
2261 fallback_mt = find_suitable_fallback(area, current_order,
2262 start_migratetype, false, &can_steal);
2263 if (fallback_mt != -1)
2268 * This should not happen - we already found a suitable fallback
2269 * when looking for the largest page.
2271 VM_BUG_ON(current_order == MAX_ORDER);
2274 page = list_first_entry(&area->free_list[fallback_mt],
2277 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2279 trace_mm_page_alloc_extfrag(page, order, current_order,
2280 start_migratetype, fallback_mt);
2287 * Do the hard work of removing an element from the buddy allocator.
2288 * Call me with the zone->lock already held.
2290 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2296 page = __rmqueue_smallest(zone, order, migratetype);
2297 if (unlikely(!page)) {
2298 if (migratetype == MIGRATE_MOVABLE)
2299 page = __rmqueue_cma_fallback(zone, order);
2301 if (!page && __rmqueue_fallback(zone, order, migratetype))
2305 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2310 * Obtain a specified number of elements from the buddy allocator, all under
2311 * a single hold of the lock, for efficiency. Add them to the supplied list.
2312 * Returns the number of new pages which were placed at *list.
2314 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2315 unsigned long count, struct list_head *list,
2316 int migratetype, bool cold)
2320 spin_lock(&zone->lock);
2321 for (i = 0; i < count; ++i) {
2322 struct page *page = __rmqueue(zone, order, migratetype);
2323 if (unlikely(page == NULL))
2326 if (unlikely(check_pcp_refill(page)))
2330 * Split buddy pages returned by expand() are received here
2331 * in physical page order. The page is added to the callers and
2332 * list and the list head then moves forward. From the callers
2333 * perspective, the linked list is ordered by page number in
2334 * some conditions. This is useful for IO devices that can
2335 * merge IO requests if the physical pages are ordered
2339 list_add(&page->lru, list);
2341 list_add_tail(&page->lru, list);
2344 if (is_migrate_cma(get_pcppage_migratetype(page)))
2345 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2350 * i pages were removed from the buddy list even if some leak due
2351 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2352 * on i. Do not confuse with 'alloced' which is the number of
2353 * pages added to the pcp list.
2355 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2356 spin_unlock(&zone->lock);
2362 * Called from the vmstat counter updater to drain pagesets of this
2363 * currently executing processor on remote nodes after they have
2366 * Note that this function must be called with the thread pinned to
2367 * a single processor.
2369 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2371 unsigned long flags;
2372 int to_drain, batch;
2374 local_irq_save(flags);
2375 batch = READ_ONCE(pcp->batch);
2376 to_drain = min(pcp->count, batch);
2378 free_pcppages_bulk(zone, to_drain, pcp);
2379 pcp->count -= to_drain;
2381 local_irq_restore(flags);
2386 * Drain pcplists of the indicated processor and zone.
2388 * The processor must either be the current processor and the
2389 * thread pinned to the current processor or a processor that
2392 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2394 unsigned long flags;
2395 struct per_cpu_pageset *pset;
2396 struct per_cpu_pages *pcp;
2398 local_irq_save(flags);
2399 pset = per_cpu_ptr(zone->pageset, cpu);
2403 free_pcppages_bulk(zone, pcp->count, pcp);
2406 local_irq_restore(flags);
2410 * Drain pcplists of all zones on the indicated processor.
2412 * The processor must either be the current processor and the
2413 * thread pinned to the current processor or a processor that
2416 static void drain_pages(unsigned int cpu)
2420 for_each_populated_zone(zone) {
2421 drain_pages_zone(cpu, zone);
2426 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2428 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2429 * the single zone's pages.
2431 void drain_local_pages(struct zone *zone)
2433 int cpu = smp_processor_id();
2436 drain_pages_zone(cpu, zone);
2441 static void drain_local_pages_wq(struct work_struct *work)
2444 * drain_all_pages doesn't use proper cpu hotplug protection so
2445 * we can race with cpu offline when the WQ can move this from
2446 * a cpu pinned worker to an unbound one. We can operate on a different
2447 * cpu which is allright but we also have to make sure to not move to
2451 drain_local_pages(NULL);
2456 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2458 * When zone parameter is non-NULL, spill just the single zone's pages.
2460 * Note that this can be extremely slow as the draining happens in a workqueue.
2462 void drain_all_pages(struct zone *zone)
2467 * Allocate in the BSS so we wont require allocation in
2468 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2470 static cpumask_t cpus_with_pcps;
2473 * Make sure nobody triggers this path before mm_percpu_wq is fully
2476 if (WARN_ON_ONCE(!mm_percpu_wq))
2479 /* Workqueues cannot recurse */
2480 if (current->flags & PF_WQ_WORKER)
2484 * Do not drain if one is already in progress unless it's specific to
2485 * a zone. Such callers are primarily CMA and memory hotplug and need
2486 * the drain to be complete when the call returns.
2488 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2491 mutex_lock(&pcpu_drain_mutex);
2495 * We don't care about racing with CPU hotplug event
2496 * as offline notification will cause the notified
2497 * cpu to drain that CPU pcps and on_each_cpu_mask
2498 * disables preemption as part of its processing
2500 for_each_online_cpu(cpu) {
2501 struct per_cpu_pageset *pcp;
2503 bool has_pcps = false;
2506 pcp = per_cpu_ptr(zone->pageset, cpu);
2510 for_each_populated_zone(z) {
2511 pcp = per_cpu_ptr(z->pageset, cpu);
2512 if (pcp->pcp.count) {
2520 cpumask_set_cpu(cpu, &cpus_with_pcps);
2522 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2525 for_each_cpu(cpu, &cpus_with_pcps) {
2526 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2527 INIT_WORK(work, drain_local_pages_wq);
2528 queue_work_on(cpu, mm_percpu_wq, work);
2530 for_each_cpu(cpu, &cpus_with_pcps)
2531 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2533 mutex_unlock(&pcpu_drain_mutex);
2536 #ifdef CONFIG_HIBERNATION
2539 * Touch the watchdog for every WD_PAGE_COUNT pages.
2541 #define WD_PAGE_COUNT (128*1024)
2543 void mark_free_pages(struct zone *zone)
2545 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2546 unsigned long flags;
2547 unsigned int order, t;
2550 if (zone_is_empty(zone))
2553 spin_lock_irqsave(&zone->lock, flags);
2555 max_zone_pfn = zone_end_pfn(zone);
2556 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2557 if (pfn_valid(pfn)) {
2558 page = pfn_to_page(pfn);
2560 if (!--page_count) {
2561 touch_nmi_watchdog();
2562 page_count = WD_PAGE_COUNT;
2565 if (page_zone(page) != zone)
2568 if (!swsusp_page_is_forbidden(page))
2569 swsusp_unset_page_free(page);
2572 for_each_migratetype_order(order, t) {
2573 list_for_each_entry(page,
2574 &zone->free_area[order].free_list[t], lru) {
2577 pfn = page_to_pfn(page);
2578 for (i = 0; i < (1UL << order); i++) {
2579 if (!--page_count) {
2580 touch_nmi_watchdog();
2581 page_count = WD_PAGE_COUNT;
2583 swsusp_set_page_free(pfn_to_page(pfn + i));
2587 spin_unlock_irqrestore(&zone->lock, flags);
2589 #endif /* CONFIG_PM */
2592 * Free a 0-order page
2593 * cold == true ? free a cold page : free a hot page
2595 void free_hot_cold_page(struct page *page, bool cold)
2597 struct zone *zone = page_zone(page);
2598 struct per_cpu_pages *pcp;
2599 unsigned long flags;
2600 unsigned long pfn = page_to_pfn(page);
2603 if (!free_pcp_prepare(page))
2606 migratetype = get_pfnblock_migratetype(page, pfn);
2607 set_pcppage_migratetype(page, migratetype);
2608 local_irq_save(flags);
2609 __count_vm_event(PGFREE);
2612 * We only track unmovable, reclaimable and movable on pcp lists.
2613 * Free ISOLATE pages back to the allocator because they are being
2614 * offlined but treat HIGHATOMIC as movable pages so we can get those
2615 * areas back if necessary. Otherwise, we may have to free
2616 * excessively into the page allocator
2618 if (migratetype >= MIGRATE_PCPTYPES) {
2619 if (unlikely(is_migrate_isolate(migratetype))) {
2620 free_one_page(zone, page, pfn, 0, migratetype);
2623 migratetype = MIGRATE_MOVABLE;
2626 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2628 list_add(&page->lru, &pcp->lists[migratetype]);
2630 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2632 if (pcp->count >= pcp->high) {
2633 unsigned long batch = READ_ONCE(pcp->batch);
2634 free_pcppages_bulk(zone, batch, pcp);
2635 pcp->count -= batch;
2639 local_irq_restore(flags);
2643 * Free a list of 0-order pages
2645 void free_hot_cold_page_list(struct list_head *list, bool cold)
2647 struct page *page, *next;
2649 list_for_each_entry_safe(page, next, list, lru) {
2650 trace_mm_page_free_batched(page, cold);
2651 free_hot_cold_page(page, cold);
2656 * split_page takes a non-compound higher-order page, and splits it into
2657 * n (1<<order) sub-pages: page[0..n]
2658 * Each sub-page must be freed individually.
2660 * Note: this is probably too low level an operation for use in drivers.
2661 * Please consult with lkml before using this in your driver.
2663 void split_page(struct page *page, unsigned int order)
2667 VM_BUG_ON_PAGE(PageCompound(page), page);
2668 VM_BUG_ON_PAGE(!page_count(page), page);
2670 for (i = 1; i < (1 << order); i++)
2671 set_page_refcounted(page + i);
2672 split_page_owner(page, order);
2674 EXPORT_SYMBOL_GPL(split_page);
2676 int __isolate_free_page(struct page *page, unsigned int order)
2678 unsigned long watermark;
2682 BUG_ON(!PageBuddy(page));
2684 zone = page_zone(page);
2685 mt = get_pageblock_migratetype(page);
2687 if (!is_migrate_isolate(mt)) {
2689 * Obey watermarks as if the page was being allocated. We can
2690 * emulate a high-order watermark check with a raised order-0
2691 * watermark, because we already know our high-order page
2694 watermark = min_wmark_pages(zone) + (1UL << order);
2695 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2698 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2701 /* Remove page from free list */
2702 list_del(&page->lru);
2703 zone->free_area[order].nr_free--;
2704 rmv_page_order(page);
2707 * Set the pageblock if the isolated page is at least half of a
2710 if (order >= pageblock_order - 1) {
2711 struct page *endpage = page + (1 << order) - 1;
2712 for (; page < endpage; page += pageblock_nr_pages) {
2713 int mt = get_pageblock_migratetype(page);
2714 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2715 && !is_migrate_highatomic(mt))
2716 set_pageblock_migratetype(page,
2722 return 1UL << order;
2726 * Update NUMA hit/miss statistics
2728 * Must be called with interrupts disabled.
2730 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2733 enum numa_stat_item local_stat = NUMA_LOCAL;
2735 if (z->node != numa_node_id())
2736 local_stat = NUMA_OTHER;
2738 if (z->node == preferred_zone->node)
2739 __inc_numa_state(z, NUMA_HIT);
2741 __inc_numa_state(z, NUMA_MISS);
2742 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2744 __inc_numa_state(z, local_stat);
2748 /* Remove page from the per-cpu list, caller must protect the list */
2749 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2750 bool cold, struct per_cpu_pages *pcp,
2751 struct list_head *list)
2756 if (list_empty(list)) {
2757 pcp->count += rmqueue_bulk(zone, 0,
2760 if (unlikely(list_empty(list)))
2765 page = list_last_entry(list, struct page, lru);
2767 page = list_first_entry(list, struct page, lru);
2769 list_del(&page->lru);
2771 } while (check_new_pcp(page));
2776 /* Lock and remove page from the per-cpu list */
2777 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2778 struct zone *zone, unsigned int order,
2779 gfp_t gfp_flags, int migratetype)
2781 struct per_cpu_pages *pcp;
2782 struct list_head *list;
2783 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2785 unsigned long flags;
2787 local_irq_save(flags);
2788 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2789 list = &pcp->lists[migratetype];
2790 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2792 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2793 zone_statistics(preferred_zone, zone);
2795 local_irq_restore(flags);
2800 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2803 struct page *rmqueue(struct zone *preferred_zone,
2804 struct zone *zone, unsigned int order,
2805 gfp_t gfp_flags, unsigned int alloc_flags,
2808 unsigned long flags;
2811 if (likely(order == 0)) {
2812 page = rmqueue_pcplist(preferred_zone, zone, order,
2813 gfp_flags, migratetype);
2818 * We most definitely don't want callers attempting to
2819 * allocate greater than order-1 page units with __GFP_NOFAIL.
2821 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2822 spin_lock_irqsave(&zone->lock, flags);
2826 if (alloc_flags & ALLOC_HARDER) {
2827 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2829 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2832 page = __rmqueue(zone, order, migratetype);
2833 } while (page && check_new_pages(page, order));
2834 spin_unlock(&zone->lock);
2837 __mod_zone_freepage_state(zone, -(1 << order),
2838 get_pcppage_migratetype(page));
2840 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2841 zone_statistics(preferred_zone, zone);
2842 local_irq_restore(flags);
2845 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2849 local_irq_restore(flags);
2853 #ifdef CONFIG_FAIL_PAGE_ALLOC
2856 struct fault_attr attr;
2858 bool ignore_gfp_highmem;
2859 bool ignore_gfp_reclaim;
2861 } fail_page_alloc = {
2862 .attr = FAULT_ATTR_INITIALIZER,
2863 .ignore_gfp_reclaim = true,
2864 .ignore_gfp_highmem = true,
2868 static int __init setup_fail_page_alloc(char *str)
2870 return setup_fault_attr(&fail_page_alloc.attr, str);
2872 __setup("fail_page_alloc=", setup_fail_page_alloc);
2874 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2876 if (order < fail_page_alloc.min_order)
2878 if (gfp_mask & __GFP_NOFAIL)
2880 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2882 if (fail_page_alloc.ignore_gfp_reclaim &&
2883 (gfp_mask & __GFP_DIRECT_RECLAIM))
2886 return should_fail(&fail_page_alloc.attr, 1 << order);
2889 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2891 static int __init fail_page_alloc_debugfs(void)
2893 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2896 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2897 &fail_page_alloc.attr);
2899 return PTR_ERR(dir);
2901 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2902 &fail_page_alloc.ignore_gfp_reclaim))
2904 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2905 &fail_page_alloc.ignore_gfp_highmem))
2907 if (!debugfs_create_u32("min-order", mode, dir,
2908 &fail_page_alloc.min_order))
2913 debugfs_remove_recursive(dir);
2918 late_initcall(fail_page_alloc_debugfs);
2920 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2922 #else /* CONFIG_FAIL_PAGE_ALLOC */
2924 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2929 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2932 * Return true if free base pages are above 'mark'. For high-order checks it
2933 * will return true of the order-0 watermark is reached and there is at least
2934 * one free page of a suitable size. Checking now avoids taking the zone lock
2935 * to check in the allocation paths if no pages are free.
2937 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2938 int classzone_idx, unsigned int alloc_flags,
2943 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2945 /* free_pages may go negative - that's OK */
2946 free_pages -= (1 << order) - 1;
2948 if (alloc_flags & ALLOC_HIGH)
2952 * If the caller does not have rights to ALLOC_HARDER then subtract
2953 * the high-atomic reserves. This will over-estimate the size of the
2954 * atomic reserve but it avoids a search.
2956 if (likely(!alloc_harder)) {
2957 free_pages -= z->nr_reserved_highatomic;
2960 * OOM victims can try even harder than normal ALLOC_HARDER
2961 * users on the grounds that it's definitely going to be in
2962 * the exit path shortly and free memory. Any allocation it
2963 * makes during the free path will be small and short-lived.
2965 if (alloc_flags & ALLOC_OOM)
2973 /* If allocation can't use CMA areas don't use free CMA pages */
2974 if (!(alloc_flags & ALLOC_CMA))
2975 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2979 * Check watermarks for an order-0 allocation request. If these
2980 * are not met, then a high-order request also cannot go ahead
2981 * even if a suitable page happened to be free.
2983 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2986 /* If this is an order-0 request then the watermark is fine */
2990 /* For a high-order request, check at least one suitable page is free */
2991 for (o = order; o < MAX_ORDER; o++) {
2992 struct free_area *area = &z->free_area[o];
3001 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3002 if (!list_empty(&area->free_list[mt]))
3007 if ((alloc_flags & ALLOC_CMA) &&
3008 !list_empty(&area->free_list[MIGRATE_CMA])) {
3016 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3017 int classzone_idx, unsigned int alloc_flags)
3019 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3020 zone_page_state(z, NR_FREE_PAGES));
3023 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3024 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3026 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3030 /* If allocation can't use CMA areas don't use free CMA pages */
3031 if (!(alloc_flags & ALLOC_CMA))
3032 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3036 * Fast check for order-0 only. If this fails then the reserves
3037 * need to be calculated. There is a corner case where the check
3038 * passes but only the high-order atomic reserve are free. If
3039 * the caller is !atomic then it'll uselessly search the free
3040 * list. That corner case is then slower but it is harmless.
3042 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3045 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3049 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3050 unsigned long mark, int classzone_idx)
3052 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3054 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3055 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3057 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3062 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3064 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3067 #else /* CONFIG_NUMA */
3068 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3072 #endif /* CONFIG_NUMA */
3075 * get_page_from_freelist goes through the zonelist trying to allocate
3078 static struct page *
3079 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3080 const struct alloc_context *ac)
3082 struct zoneref *z = ac->preferred_zoneref;
3084 struct pglist_data *last_pgdat_dirty_limit = NULL;
3087 * Scan zonelist, looking for a zone with enough free.
3088 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3090 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3095 if (cpusets_enabled() &&
3096 (alloc_flags & ALLOC_CPUSET) &&
3097 !__cpuset_zone_allowed(zone, gfp_mask))
3100 * When allocating a page cache page for writing, we
3101 * want to get it from a node that is within its dirty
3102 * limit, such that no single node holds more than its
3103 * proportional share of globally allowed dirty pages.
3104 * The dirty limits take into account the node's
3105 * lowmem reserves and high watermark so that kswapd
3106 * should be able to balance it without having to
3107 * write pages from its LRU list.
3109 * XXX: For now, allow allocations to potentially
3110 * exceed the per-node dirty limit in the slowpath
3111 * (spread_dirty_pages unset) before going into reclaim,
3112 * which is important when on a NUMA setup the allowed
3113 * nodes are together not big enough to reach the
3114 * global limit. The proper fix for these situations
3115 * will require awareness of nodes in the
3116 * dirty-throttling and the flusher threads.
3118 if (ac->spread_dirty_pages) {
3119 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3122 if (!node_dirty_ok(zone->zone_pgdat)) {
3123 last_pgdat_dirty_limit = zone->zone_pgdat;
3128 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3129 if (!zone_watermark_fast(zone, order, mark,
3130 ac_classzone_idx(ac), alloc_flags)) {
3133 /* Checked here to keep the fast path fast */
3134 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3135 if (alloc_flags & ALLOC_NO_WATERMARKS)
3138 if (node_reclaim_mode == 0 ||
3139 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3142 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3144 case NODE_RECLAIM_NOSCAN:
3147 case NODE_RECLAIM_FULL:
3148 /* scanned but unreclaimable */
3151 /* did we reclaim enough */
3152 if (zone_watermark_ok(zone, order, mark,
3153 ac_classzone_idx(ac), alloc_flags))
3161 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3162 gfp_mask, alloc_flags, ac->migratetype);
3164 prep_new_page(page, order, gfp_mask, alloc_flags);
3167 * If this is a high-order atomic allocation then check
3168 * if the pageblock should be reserved for the future
3170 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3171 reserve_highatomic_pageblock(page, zone, order);
3181 * Large machines with many possible nodes should not always dump per-node
3182 * meminfo in irq context.
3184 static inline bool should_suppress_show_mem(void)
3189 ret = in_interrupt();
3194 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3196 unsigned int filter = SHOW_MEM_FILTER_NODES;
3197 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3199 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3203 * This documents exceptions given to allocations in certain
3204 * contexts that are allowed to allocate outside current's set
3207 if (!(gfp_mask & __GFP_NOMEMALLOC))
3208 if (tsk_is_oom_victim(current) ||
3209 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3210 filter &= ~SHOW_MEM_FILTER_NODES;
3211 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3212 filter &= ~SHOW_MEM_FILTER_NODES;
3214 show_mem(filter, nodemask);
3217 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3219 struct va_format vaf;
3221 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3222 DEFAULT_RATELIMIT_BURST);
3224 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3227 pr_warn("%s: ", current->comm);
3229 va_start(args, fmt);
3232 pr_cont("%pV", &vaf);
3235 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3237 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3239 pr_cont("(null)\n");
3241 cpuset_print_current_mems_allowed();
3244 warn_alloc_show_mem(gfp_mask, nodemask);
3247 static inline struct page *
3248 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3249 unsigned int alloc_flags,
3250 const struct alloc_context *ac)
3254 page = get_page_from_freelist(gfp_mask, order,
3255 alloc_flags|ALLOC_CPUSET, ac);
3257 * fallback to ignore cpuset restriction if our nodes
3261 page = get_page_from_freelist(gfp_mask, order,
3267 static inline struct page *
3268 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3269 const struct alloc_context *ac, unsigned long *did_some_progress)
3271 struct oom_control oc = {
3272 .zonelist = ac->zonelist,
3273 .nodemask = ac->nodemask,
3275 .gfp_mask = gfp_mask,
3280 *did_some_progress = 0;
3283 * Acquire the oom lock. If that fails, somebody else is
3284 * making progress for us.
3286 if (!mutex_trylock(&oom_lock)) {
3287 *did_some_progress = 1;
3288 schedule_timeout_uninterruptible(1);
3293 * Go through the zonelist yet one more time, keep very high watermark
3294 * here, this is only to catch a parallel oom killing, we must fail if
3295 * we're still under heavy pressure. But make sure that this reclaim
3296 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3297 * allocation which will never fail due to oom_lock already held.
3299 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3300 ~__GFP_DIRECT_RECLAIM, order,
3301 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3305 /* Coredumps can quickly deplete all memory reserves */
3306 if (current->flags & PF_DUMPCORE)
3308 /* The OOM killer will not help higher order allocs */
3309 if (order > PAGE_ALLOC_COSTLY_ORDER)
3312 * We have already exhausted all our reclaim opportunities without any
3313 * success so it is time to admit defeat. We will skip the OOM killer
3314 * because it is very likely that the caller has a more reasonable
3315 * fallback than shooting a random task.
3317 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3319 /* The OOM killer does not needlessly kill tasks for lowmem */
3320 if (ac->high_zoneidx < ZONE_NORMAL)
3322 if (pm_suspended_storage())
3325 * XXX: GFP_NOFS allocations should rather fail than rely on
3326 * other request to make a forward progress.
3327 * We are in an unfortunate situation where out_of_memory cannot
3328 * do much for this context but let's try it to at least get
3329 * access to memory reserved if the current task is killed (see
3330 * out_of_memory). Once filesystems are ready to handle allocation
3331 * failures more gracefully we should just bail out here.
3334 /* The OOM killer may not free memory on a specific node */
3335 if (gfp_mask & __GFP_THISNODE)
3338 /* Exhausted what can be done so it's blamo time */
3339 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3340 *did_some_progress = 1;
3343 * Help non-failing allocations by giving them access to memory
3346 if (gfp_mask & __GFP_NOFAIL)
3347 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3348 ALLOC_NO_WATERMARKS, ac);
3351 mutex_unlock(&oom_lock);
3356 * Maximum number of compaction retries wit a progress before OOM
3357 * killer is consider as the only way to move forward.
3359 #define MAX_COMPACT_RETRIES 16
3361 #ifdef CONFIG_COMPACTION
3362 /* Try memory compaction for high-order allocations before reclaim */
3363 static struct page *
3364 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3365 unsigned int alloc_flags, const struct alloc_context *ac,
3366 enum compact_priority prio, enum compact_result *compact_result)
3369 unsigned int noreclaim_flag;
3374 noreclaim_flag = memalloc_noreclaim_save();
3375 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3377 memalloc_noreclaim_restore(noreclaim_flag);
3379 if (*compact_result <= COMPACT_INACTIVE)
3383 * At least in one zone compaction wasn't deferred or skipped, so let's
3384 * count a compaction stall
3386 count_vm_event(COMPACTSTALL);
3388 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3391 struct zone *zone = page_zone(page);
3393 zone->compact_blockskip_flush = false;
3394 compaction_defer_reset(zone, order, true);
3395 count_vm_event(COMPACTSUCCESS);
3400 * It's bad if compaction run occurs and fails. The most likely reason
3401 * is that pages exist, but not enough to satisfy watermarks.
3403 count_vm_event(COMPACTFAIL);
3411 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3412 enum compact_result compact_result,
3413 enum compact_priority *compact_priority,
3414 int *compaction_retries)
3416 int max_retries = MAX_COMPACT_RETRIES;
3419 int retries = *compaction_retries;
3420 enum compact_priority priority = *compact_priority;
3425 if (compaction_made_progress(compact_result))
3426 (*compaction_retries)++;
3429 * compaction considers all the zone as desperately out of memory
3430 * so it doesn't really make much sense to retry except when the
3431 * failure could be caused by insufficient priority
3433 if (compaction_failed(compact_result))
3434 goto check_priority;
3437 * make sure the compaction wasn't deferred or didn't bail out early
3438 * due to locks contention before we declare that we should give up.
3439 * But do not retry if the given zonelist is not suitable for
3442 if (compaction_withdrawn(compact_result)) {
3443 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3448 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3449 * costly ones because they are de facto nofail and invoke OOM
3450 * killer to move on while costly can fail and users are ready
3451 * to cope with that. 1/4 retries is rather arbitrary but we
3452 * would need much more detailed feedback from compaction to
3453 * make a better decision.
3455 if (order > PAGE_ALLOC_COSTLY_ORDER)
3457 if (*compaction_retries <= max_retries) {
3463 * Make sure there are attempts at the highest priority if we exhausted
3464 * all retries or failed at the lower priorities.
3467 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3468 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3470 if (*compact_priority > min_priority) {
3471 (*compact_priority)--;
3472 *compaction_retries = 0;
3476 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3480 static inline struct page *
3481 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3482 unsigned int alloc_flags, const struct alloc_context *ac,
3483 enum compact_priority prio, enum compact_result *compact_result)
3485 *compact_result = COMPACT_SKIPPED;
3490 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3491 enum compact_result compact_result,
3492 enum compact_priority *compact_priority,
3493 int *compaction_retries)
3498 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3502 * There are setups with compaction disabled which would prefer to loop
3503 * inside the allocator rather than hit the oom killer prematurely.
3504 * Let's give them a good hope and keep retrying while the order-0
3505 * watermarks are OK.
3507 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3509 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3510 ac_classzone_idx(ac), alloc_flags))
3515 #endif /* CONFIG_COMPACTION */
3517 #ifdef CONFIG_LOCKDEP
3518 struct lockdep_map __fs_reclaim_map =
3519 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3521 static bool __need_fs_reclaim(gfp_t gfp_mask)
3523 gfp_mask = current_gfp_context(gfp_mask);
3525 /* no reclaim without waiting on it */
3526 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3529 /* this guy won't enter reclaim */
3530 if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
3533 /* We're only interested __GFP_FS allocations for now */
3534 if (!(gfp_mask & __GFP_FS))
3537 if (gfp_mask & __GFP_NOLOCKDEP)
3543 void fs_reclaim_acquire(gfp_t gfp_mask)
3545 if (__need_fs_reclaim(gfp_mask))
3546 lock_map_acquire(&__fs_reclaim_map);
3548 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3550 void fs_reclaim_release(gfp_t gfp_mask)
3552 if (__need_fs_reclaim(gfp_mask))
3553 lock_map_release(&__fs_reclaim_map);
3555 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3558 /* Perform direct synchronous page reclaim */
3560 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3561 const struct alloc_context *ac)
3563 struct reclaim_state reclaim_state;
3565 unsigned int noreclaim_flag;
3569 /* We now go into synchronous reclaim */
3570 cpuset_memory_pressure_bump();
3571 noreclaim_flag = memalloc_noreclaim_save();
3572 fs_reclaim_acquire(gfp_mask);
3573 reclaim_state.reclaimed_slab = 0;
3574 current->reclaim_state = &reclaim_state;
3576 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3579 current->reclaim_state = NULL;
3580 fs_reclaim_release(gfp_mask);
3581 memalloc_noreclaim_restore(noreclaim_flag);
3588 /* The really slow allocator path where we enter direct reclaim */
3589 static inline struct page *
3590 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3591 unsigned int alloc_flags, const struct alloc_context *ac,
3592 unsigned long *did_some_progress)
3594 struct page *page = NULL;
3595 bool drained = false;
3597 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3598 if (unlikely(!(*did_some_progress)))
3602 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3605 * If an allocation failed after direct reclaim, it could be because
3606 * pages are pinned on the per-cpu lists or in high alloc reserves.
3607 * Shrink them them and try again
3609 if (!page && !drained) {
3610 unreserve_highatomic_pageblock(ac, false);
3611 drain_all_pages(NULL);
3619 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3623 pg_data_t *last_pgdat = NULL;
3625 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3626 ac->high_zoneidx, ac->nodemask) {
3627 if (last_pgdat != zone->zone_pgdat)
3628 wakeup_kswapd(zone, order, ac->high_zoneidx);
3629 last_pgdat = zone->zone_pgdat;
3633 static inline unsigned int
3634 gfp_to_alloc_flags(gfp_t gfp_mask)
3636 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3638 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3639 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3642 * The caller may dip into page reserves a bit more if the caller
3643 * cannot run direct reclaim, or if the caller has realtime scheduling
3644 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3645 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3647 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3649 if (gfp_mask & __GFP_ATOMIC) {
3651 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3652 * if it can't schedule.
3654 if (!(gfp_mask & __GFP_NOMEMALLOC))
3655 alloc_flags |= ALLOC_HARDER;
3657 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3658 * comment for __cpuset_node_allowed().
3660 alloc_flags &= ~ALLOC_CPUSET;
3661 } else if (unlikely(rt_task(current)) && !in_interrupt())
3662 alloc_flags |= ALLOC_HARDER;
3665 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3666 alloc_flags |= ALLOC_CMA;
3671 static bool oom_reserves_allowed(struct task_struct *tsk)
3673 if (!tsk_is_oom_victim(tsk))
3677 * !MMU doesn't have oom reaper so give access to memory reserves
3678 * only to the thread with TIF_MEMDIE set
3680 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3687 * Distinguish requests which really need access to full memory
3688 * reserves from oom victims which can live with a portion of it
3690 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3692 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3694 if (gfp_mask & __GFP_MEMALLOC)
3695 return ALLOC_NO_WATERMARKS;
3696 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3697 return ALLOC_NO_WATERMARKS;
3698 if (!in_interrupt()) {
3699 if (current->flags & PF_MEMALLOC)
3700 return ALLOC_NO_WATERMARKS;
3701 else if (oom_reserves_allowed(current))
3708 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3710 return !!__gfp_pfmemalloc_flags(gfp_mask);
3714 * Checks whether it makes sense to retry the reclaim to make a forward progress
3715 * for the given allocation request.
3717 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3718 * without success, or when we couldn't even meet the watermark if we
3719 * reclaimed all remaining pages on the LRU lists.
3721 * Returns true if a retry is viable or false to enter the oom path.
3724 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3725 struct alloc_context *ac, int alloc_flags,
3726 bool did_some_progress, int *no_progress_loops)
3732 * Costly allocations might have made a progress but this doesn't mean
3733 * their order will become available due to high fragmentation so
3734 * always increment the no progress counter for them
3736 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3737 *no_progress_loops = 0;
3739 (*no_progress_loops)++;
3742 * Make sure we converge to OOM if we cannot make any progress
3743 * several times in the row.
3745 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3746 /* Before OOM, exhaust highatomic_reserve */
3747 return unreserve_highatomic_pageblock(ac, true);
3751 * Keep reclaiming pages while there is a chance this will lead
3752 * somewhere. If none of the target zones can satisfy our allocation
3753 * request even if all reclaimable pages are considered then we are
3754 * screwed and have to go OOM.
3756 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3758 unsigned long available;
3759 unsigned long reclaimable;
3760 unsigned long min_wmark = min_wmark_pages(zone);
3763 available = reclaimable = zone_reclaimable_pages(zone);
3764 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3767 * Would the allocation succeed if we reclaimed all
3768 * reclaimable pages?
3770 wmark = __zone_watermark_ok(zone, order, min_wmark,
3771 ac_classzone_idx(ac), alloc_flags, available);
3772 trace_reclaim_retry_zone(z, order, reclaimable,
3773 available, min_wmark, *no_progress_loops, wmark);
3776 * If we didn't make any progress and have a lot of
3777 * dirty + writeback pages then we should wait for
3778 * an IO to complete to slow down the reclaim and
3779 * prevent from pre mature OOM
3781 if (!did_some_progress) {
3782 unsigned long write_pending;
3784 write_pending = zone_page_state_snapshot(zone,
3785 NR_ZONE_WRITE_PENDING);
3787 if (2 * write_pending > reclaimable) {
3788 congestion_wait(BLK_RW_ASYNC, HZ/10);
3794 * Memory allocation/reclaim might be called from a WQ
3795 * context and the current implementation of the WQ
3796 * concurrency control doesn't recognize that
3797 * a particular WQ is congested if the worker thread is
3798 * looping without ever sleeping. Therefore we have to
3799 * do a short sleep here rather than calling
3802 if (current->flags & PF_WQ_WORKER)
3803 schedule_timeout_uninterruptible(1);
3815 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3818 * It's possible that cpuset's mems_allowed and the nodemask from
3819 * mempolicy don't intersect. This should be normally dealt with by
3820 * policy_nodemask(), but it's possible to race with cpuset update in
3821 * such a way the check therein was true, and then it became false
3822 * before we got our cpuset_mems_cookie here.
3823 * This assumes that for all allocations, ac->nodemask can come only
3824 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3825 * when it does not intersect with the cpuset restrictions) or the
3826 * caller can deal with a violated nodemask.
3828 if (cpusets_enabled() && ac->nodemask &&
3829 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3830 ac->nodemask = NULL;
3835 * When updating a task's mems_allowed or mempolicy nodemask, it is
3836 * possible to race with parallel threads in such a way that our
3837 * allocation can fail while the mask is being updated. If we are about
3838 * to fail, check if the cpuset changed during allocation and if so,
3841 if (read_mems_allowed_retry(cpuset_mems_cookie))
3847 static inline struct page *
3848 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3849 struct alloc_context *ac)
3851 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3852 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3853 struct page *page = NULL;
3854 unsigned int alloc_flags;
3855 unsigned long did_some_progress;
3856 enum compact_priority compact_priority;
3857 enum compact_result compact_result;
3858 int compaction_retries;
3859 int no_progress_loops;
3860 unsigned long alloc_start = jiffies;
3861 unsigned int stall_timeout = 10 * HZ;
3862 unsigned int cpuset_mems_cookie;
3866 * In the slowpath, we sanity check order to avoid ever trying to
3867 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3868 * be using allocators in order of preference for an area that is
3871 if (order >= MAX_ORDER) {
3872 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3877 * We also sanity check to catch abuse of atomic reserves being used by
3878 * callers that are not in atomic context.
3880 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3881 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3882 gfp_mask &= ~__GFP_ATOMIC;
3885 compaction_retries = 0;
3886 no_progress_loops = 0;
3887 compact_priority = DEF_COMPACT_PRIORITY;
3888 cpuset_mems_cookie = read_mems_allowed_begin();
3891 * The fast path uses conservative alloc_flags to succeed only until
3892 * kswapd needs to be woken up, and to avoid the cost of setting up
3893 * alloc_flags precisely. So we do that now.
3895 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3898 * We need to recalculate the starting point for the zonelist iterator
3899 * because we might have used different nodemask in the fast path, or
3900 * there was a cpuset modification and we are retrying - otherwise we
3901 * could end up iterating over non-eligible zones endlessly.
3903 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3904 ac->high_zoneidx, ac->nodemask);
3905 if (!ac->preferred_zoneref->zone)
3908 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3909 wake_all_kswapds(order, ac);
3912 * The adjusted alloc_flags might result in immediate success, so try
3915 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3920 * For costly allocations, try direct compaction first, as it's likely
3921 * that we have enough base pages and don't need to reclaim. For non-
3922 * movable high-order allocations, do that as well, as compaction will
3923 * try prevent permanent fragmentation by migrating from blocks of the
3925 * Don't try this for allocations that are allowed to ignore
3926 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3928 if (can_direct_reclaim &&
3930 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3931 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3932 page = __alloc_pages_direct_compact(gfp_mask, order,
3934 INIT_COMPACT_PRIORITY,
3940 * Checks for costly allocations with __GFP_NORETRY, which
3941 * includes THP page fault allocations
3943 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3945 * If compaction is deferred for high-order allocations,
3946 * it is because sync compaction recently failed. If
3947 * this is the case and the caller requested a THP
3948 * allocation, we do not want to heavily disrupt the
3949 * system, so we fail the allocation instead of entering
3952 if (compact_result == COMPACT_DEFERRED)
3956 * Looks like reclaim/compaction is worth trying, but
3957 * sync compaction could be very expensive, so keep
3958 * using async compaction.
3960 compact_priority = INIT_COMPACT_PRIORITY;
3965 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3966 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3967 wake_all_kswapds(order, ac);
3969 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
3971 alloc_flags = reserve_flags;
3974 * Reset the zonelist iterators if memory policies can be ignored.
3975 * These allocations are high priority and system rather than user
3978 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
3979 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3980 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3981 ac->high_zoneidx, ac->nodemask);
3984 /* Attempt with potentially adjusted zonelist and alloc_flags */
3985 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3989 /* Caller is not willing to reclaim, we can't balance anything */
3990 if (!can_direct_reclaim)
3993 /* Make sure we know about allocations which stall for too long */
3994 if (time_after(jiffies, alloc_start + stall_timeout)) {
3995 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
3996 "page allocation stalls for %ums, order:%u",
3997 jiffies_to_msecs(jiffies-alloc_start), order);
3998 stall_timeout += 10 * HZ;
4001 /* Avoid recursion of direct reclaim */
4002 if (current->flags & PF_MEMALLOC)
4005 /* Try direct reclaim and then allocating */
4006 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4007 &did_some_progress);
4011 /* Try direct compaction and then allocating */
4012 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4013 compact_priority, &compact_result);
4017 /* Do not loop if specifically requested */
4018 if (gfp_mask & __GFP_NORETRY)
4022 * Do not retry costly high order allocations unless they are
4023 * __GFP_RETRY_MAYFAIL
4025 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4028 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4029 did_some_progress > 0, &no_progress_loops))
4033 * It doesn't make any sense to retry for the compaction if the order-0
4034 * reclaim is not able to make any progress because the current
4035 * implementation of the compaction depends on the sufficient amount
4036 * of free memory (see __compaction_suitable)
4038 if (did_some_progress > 0 &&
4039 should_compact_retry(ac, order, alloc_flags,
4040 compact_result, &compact_priority,
4041 &compaction_retries))
4045 /* Deal with possible cpuset update races before we start OOM killing */
4046 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4049 /* Reclaim has failed us, start killing things */
4050 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4054 /* Avoid allocations with no watermarks from looping endlessly */
4055 if (tsk_is_oom_victim(current) &&
4056 (alloc_flags == ALLOC_OOM ||
4057 (gfp_mask & __GFP_NOMEMALLOC)))
4060 /* Retry as long as the OOM killer is making progress */
4061 if (did_some_progress) {
4062 no_progress_loops = 0;
4067 /* Deal with possible cpuset update races before we fail */
4068 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4072 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4075 if (gfp_mask & __GFP_NOFAIL) {
4077 * All existing users of the __GFP_NOFAIL are blockable, so warn
4078 * of any new users that actually require GFP_NOWAIT
4080 if (WARN_ON_ONCE(!can_direct_reclaim))
4084 * PF_MEMALLOC request from this context is rather bizarre
4085 * because we cannot reclaim anything and only can loop waiting
4086 * for somebody to do a work for us
4088 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4091 * non failing costly orders are a hard requirement which we
4092 * are not prepared for much so let's warn about these users
4093 * so that we can identify them and convert them to something
4096 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4099 * Help non-failing allocations by giving them access to memory
4100 * reserves but do not use ALLOC_NO_WATERMARKS because this
4101 * could deplete whole memory reserves which would just make
4102 * the situation worse
4104 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4112 warn_alloc(gfp_mask, ac->nodemask,
4113 "page allocation failure: order:%u", order);
4118 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4119 int preferred_nid, nodemask_t *nodemask,
4120 struct alloc_context *ac, gfp_t *alloc_mask,
4121 unsigned int *alloc_flags)
4123 ac->high_zoneidx = gfp_zone(gfp_mask);
4124 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4125 ac->nodemask = nodemask;
4126 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4128 if (cpusets_enabled()) {
4129 *alloc_mask |= __GFP_HARDWALL;
4131 ac->nodemask = &cpuset_current_mems_allowed;
4133 *alloc_flags |= ALLOC_CPUSET;
4136 fs_reclaim_acquire(gfp_mask);
4137 fs_reclaim_release(gfp_mask);
4139 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4141 if (should_fail_alloc_page(gfp_mask, order))
4144 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4145 *alloc_flags |= ALLOC_CMA;
4150 /* Determine whether to spread dirty pages and what the first usable zone */
4151 static inline void finalise_ac(gfp_t gfp_mask,
4152 unsigned int order, struct alloc_context *ac)
4154 /* Dirty zone balancing only done in the fast path */
4155 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4158 * The preferred zone is used for statistics but crucially it is
4159 * also used as the starting point for the zonelist iterator. It
4160 * may get reset for allocations that ignore memory policies.
4162 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4163 ac->high_zoneidx, ac->nodemask);
4167 * This is the 'heart' of the zoned buddy allocator.
4170 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4171 nodemask_t *nodemask)
4174 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4175 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4176 struct alloc_context ac = { };
4178 gfp_mask &= gfp_allowed_mask;
4179 alloc_mask = gfp_mask;
4180 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4183 finalise_ac(gfp_mask, order, &ac);
4185 /* First allocation attempt */
4186 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4191 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4192 * resp. GFP_NOIO which has to be inherited for all allocation requests
4193 * from a particular context which has been marked by
4194 * memalloc_no{fs,io}_{save,restore}.
4196 alloc_mask = current_gfp_context(gfp_mask);
4197 ac.spread_dirty_pages = false;
4200 * Restore the original nodemask if it was potentially replaced with
4201 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4203 if (unlikely(ac.nodemask != nodemask))
4204 ac.nodemask = nodemask;
4206 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4209 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4210 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4211 __free_pages(page, order);
4215 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4219 EXPORT_SYMBOL(__alloc_pages_nodemask);
4222 * Common helper functions.
4224 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4229 * __get_free_pages() returns a 32-bit address, which cannot represent
4232 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4234 page = alloc_pages(gfp_mask, order);
4237 return (unsigned long) page_address(page);
4239 EXPORT_SYMBOL(__get_free_pages);
4241 unsigned long get_zeroed_page(gfp_t gfp_mask)
4243 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4245 EXPORT_SYMBOL(get_zeroed_page);
4247 void __free_pages(struct page *page, unsigned int order)
4249 if (put_page_testzero(page)) {
4251 free_hot_cold_page(page, false);
4253 __free_pages_ok(page, order);
4257 EXPORT_SYMBOL(__free_pages);
4259 void free_pages(unsigned long addr, unsigned int order)
4262 VM_BUG_ON(!virt_addr_valid((void *)addr));
4263 __free_pages(virt_to_page((void *)addr), order);
4267 EXPORT_SYMBOL(free_pages);
4271 * An arbitrary-length arbitrary-offset area of memory which resides
4272 * within a 0 or higher order page. Multiple fragments within that page
4273 * are individually refcounted, in the page's reference counter.
4275 * The page_frag functions below provide a simple allocation framework for
4276 * page fragments. This is used by the network stack and network device
4277 * drivers to provide a backing region of memory for use as either an
4278 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4280 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4283 struct page *page = NULL;
4284 gfp_t gfp = gfp_mask;
4286 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4287 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4289 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4290 PAGE_FRAG_CACHE_MAX_ORDER);
4291 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4293 if (unlikely(!page))
4294 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4296 nc->va = page ? page_address(page) : NULL;
4301 void __page_frag_cache_drain(struct page *page, unsigned int count)
4303 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4305 if (page_ref_sub_and_test(page, count)) {
4306 unsigned int order = compound_order(page);
4309 free_hot_cold_page(page, false);
4311 __free_pages_ok(page, order);
4314 EXPORT_SYMBOL(__page_frag_cache_drain);
4316 void *page_frag_alloc(struct page_frag_cache *nc,
4317 unsigned int fragsz, gfp_t gfp_mask)
4319 unsigned int size = PAGE_SIZE;
4323 if (unlikely(!nc->va)) {
4325 page = __page_frag_cache_refill(nc, gfp_mask);
4329 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4330 /* if size can vary use size else just use PAGE_SIZE */
4333 /* Even if we own the page, we do not use atomic_set().
4334 * This would break get_page_unless_zero() users.
4336 page_ref_add(page, size - 1);
4338 /* reset page count bias and offset to start of new frag */
4339 nc->pfmemalloc = page_is_pfmemalloc(page);
4340 nc->pagecnt_bias = size;
4344 offset = nc->offset - fragsz;
4345 if (unlikely(offset < 0)) {
4346 page = virt_to_page(nc->va);
4348 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4351 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4352 /* if size can vary use size else just use PAGE_SIZE */
4355 /* OK, page count is 0, we can safely set it */
4356 set_page_count(page, size);
4358 /* reset page count bias and offset to start of new frag */
4359 nc->pagecnt_bias = size;
4360 offset = size - fragsz;
4364 nc->offset = offset;
4366 return nc->va + offset;
4368 EXPORT_SYMBOL(page_frag_alloc);
4371 * Frees a page fragment allocated out of either a compound or order 0 page.
4373 void page_frag_free(void *addr)
4375 struct page *page = virt_to_head_page(addr);
4377 if (unlikely(put_page_testzero(page)))
4378 __free_pages_ok(page, compound_order(page));
4380 EXPORT_SYMBOL(page_frag_free);
4382 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4386 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4387 unsigned long used = addr + PAGE_ALIGN(size);
4389 split_page(virt_to_page((void *)addr), order);
4390 while (used < alloc_end) {
4395 return (void *)addr;
4399 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4400 * @size: the number of bytes to allocate
4401 * @gfp_mask: GFP flags for the allocation
4403 * This function is similar to alloc_pages(), except that it allocates the
4404 * minimum number of pages to satisfy the request. alloc_pages() can only
4405 * allocate memory in power-of-two pages.
4407 * This function is also limited by MAX_ORDER.
4409 * Memory allocated by this function must be released by free_pages_exact().
4411 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4413 unsigned int order = get_order(size);
4416 addr = __get_free_pages(gfp_mask, order);
4417 return make_alloc_exact(addr, order, size);
4419 EXPORT_SYMBOL(alloc_pages_exact);
4422 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4424 * @nid: the preferred node ID where memory should be allocated
4425 * @size: the number of bytes to allocate
4426 * @gfp_mask: GFP flags for the allocation
4428 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4431 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4433 unsigned int order = get_order(size);
4434 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4437 return make_alloc_exact((unsigned long)page_address(p), order, size);
4441 * free_pages_exact - release memory allocated via alloc_pages_exact()
4442 * @virt: the value returned by alloc_pages_exact.
4443 * @size: size of allocation, same value as passed to alloc_pages_exact().
4445 * Release the memory allocated by a previous call to alloc_pages_exact.
4447 void free_pages_exact(void *virt, size_t size)
4449 unsigned long addr = (unsigned long)virt;
4450 unsigned long end = addr + PAGE_ALIGN(size);
4452 while (addr < end) {
4457 EXPORT_SYMBOL(free_pages_exact);
4460 * nr_free_zone_pages - count number of pages beyond high watermark
4461 * @offset: The zone index of the highest zone
4463 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4464 * high watermark within all zones at or below a given zone index. For each
4465 * zone, the number of pages is calculated as:
4467 * nr_free_zone_pages = managed_pages - high_pages
4469 static unsigned long nr_free_zone_pages(int offset)
4474 /* Just pick one node, since fallback list is circular */
4475 unsigned long sum = 0;
4477 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4479 for_each_zone_zonelist(zone, z, zonelist, offset) {
4480 unsigned long size = zone->managed_pages;
4481 unsigned long high = high_wmark_pages(zone);
4490 * nr_free_buffer_pages - count number of pages beyond high watermark
4492 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4493 * watermark within ZONE_DMA and ZONE_NORMAL.
4495 unsigned long nr_free_buffer_pages(void)
4497 return nr_free_zone_pages(gfp_zone(GFP_USER));
4499 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4502 * nr_free_pagecache_pages - count number of pages beyond high watermark
4504 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4505 * high watermark within all zones.
4507 unsigned long nr_free_pagecache_pages(void)
4509 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4512 static inline void show_node(struct zone *zone)
4514 if (IS_ENABLED(CONFIG_NUMA))
4515 printk("Node %d ", zone_to_nid(zone));
4518 long si_mem_available(void)
4521 unsigned long pagecache;
4522 unsigned long wmark_low = 0;
4523 unsigned long pages[NR_LRU_LISTS];
4527 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4528 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4531 wmark_low += zone->watermark[WMARK_LOW];
4534 * Estimate the amount of memory available for userspace allocations,
4535 * without causing swapping.
4537 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4540 * Not all the page cache can be freed, otherwise the system will
4541 * start swapping. Assume at least half of the page cache, or the
4542 * low watermark worth of cache, needs to stay.
4544 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4545 pagecache -= min(pagecache / 2, wmark_low);
4546 available += pagecache;
4549 * Part of the reclaimable slab consists of items that are in use,
4550 * and cannot be freed. Cap this estimate at the low watermark.
4552 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4553 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4560 EXPORT_SYMBOL_GPL(si_mem_available);
4562 void si_meminfo(struct sysinfo *val)
4564 val->totalram = totalram_pages;
4565 val->sharedram = global_node_page_state(NR_SHMEM);
4566 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4567 val->bufferram = nr_blockdev_pages();
4568 val->totalhigh = totalhigh_pages;
4569 val->freehigh = nr_free_highpages();
4570 val->mem_unit = PAGE_SIZE;
4573 EXPORT_SYMBOL(si_meminfo);
4576 void si_meminfo_node(struct sysinfo *val, int nid)
4578 int zone_type; /* needs to be signed */
4579 unsigned long managed_pages = 0;
4580 unsigned long managed_highpages = 0;
4581 unsigned long free_highpages = 0;
4582 pg_data_t *pgdat = NODE_DATA(nid);
4584 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4585 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4586 val->totalram = managed_pages;
4587 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4588 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4589 #ifdef CONFIG_HIGHMEM
4590 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4591 struct zone *zone = &pgdat->node_zones[zone_type];
4593 if (is_highmem(zone)) {
4594 managed_highpages += zone->managed_pages;
4595 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4598 val->totalhigh = managed_highpages;
4599 val->freehigh = free_highpages;
4601 val->totalhigh = managed_highpages;
4602 val->freehigh = free_highpages;
4604 val->mem_unit = PAGE_SIZE;
4609 * Determine whether the node should be displayed or not, depending on whether
4610 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4612 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4614 if (!(flags & SHOW_MEM_FILTER_NODES))
4618 * no node mask - aka implicit memory numa policy. Do not bother with
4619 * the synchronization - read_mems_allowed_begin - because we do not
4620 * have to be precise here.
4623 nodemask = &cpuset_current_mems_allowed;
4625 return !node_isset(nid, *nodemask);
4628 #define K(x) ((x) << (PAGE_SHIFT-10))
4630 static void show_migration_types(unsigned char type)
4632 static const char types[MIGRATE_TYPES] = {
4633 [MIGRATE_UNMOVABLE] = 'U',
4634 [MIGRATE_MOVABLE] = 'M',
4635 [MIGRATE_RECLAIMABLE] = 'E',
4636 [MIGRATE_HIGHATOMIC] = 'H',
4638 [MIGRATE_CMA] = 'C',
4640 #ifdef CONFIG_MEMORY_ISOLATION
4641 [MIGRATE_ISOLATE] = 'I',
4644 char tmp[MIGRATE_TYPES + 1];
4648 for (i = 0; i < MIGRATE_TYPES; i++) {
4649 if (type & (1 << i))
4654 printk(KERN_CONT "(%s) ", tmp);
4658 * Show free area list (used inside shift_scroll-lock stuff)
4659 * We also calculate the percentage fragmentation. We do this by counting the
4660 * memory on each free list with the exception of the first item on the list.
4663 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4666 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4668 unsigned long free_pcp = 0;
4673 for_each_populated_zone(zone) {
4674 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4677 for_each_online_cpu(cpu)
4678 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4681 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4682 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4683 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4684 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4685 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4686 " free:%lu free_pcp:%lu free_cma:%lu\n",
4687 global_node_page_state(NR_ACTIVE_ANON),
4688 global_node_page_state(NR_INACTIVE_ANON),
4689 global_node_page_state(NR_ISOLATED_ANON),
4690 global_node_page_state(NR_ACTIVE_FILE),
4691 global_node_page_state(NR_INACTIVE_FILE),
4692 global_node_page_state(NR_ISOLATED_FILE),
4693 global_node_page_state(NR_UNEVICTABLE),
4694 global_node_page_state(NR_FILE_DIRTY),
4695 global_node_page_state(NR_WRITEBACK),
4696 global_node_page_state(NR_UNSTABLE_NFS),
4697 global_node_page_state(NR_SLAB_RECLAIMABLE),
4698 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4699 global_node_page_state(NR_FILE_MAPPED),
4700 global_node_page_state(NR_SHMEM),
4701 global_zone_page_state(NR_PAGETABLE),
4702 global_zone_page_state(NR_BOUNCE),
4703 global_zone_page_state(NR_FREE_PAGES),
4705 global_zone_page_state(NR_FREE_CMA_PAGES));
4707 for_each_online_pgdat(pgdat) {
4708 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4712 " active_anon:%lukB"
4713 " inactive_anon:%lukB"
4714 " active_file:%lukB"
4715 " inactive_file:%lukB"
4716 " unevictable:%lukB"
4717 " isolated(anon):%lukB"
4718 " isolated(file):%lukB"
4723 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4725 " shmem_pmdmapped: %lukB"
4728 " writeback_tmp:%lukB"
4730 " all_unreclaimable? %s"
4733 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4734 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4735 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4736 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4737 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4738 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4739 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4740 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4741 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4742 K(node_page_state(pgdat, NR_WRITEBACK)),
4743 K(node_page_state(pgdat, NR_SHMEM)),
4744 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4745 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4746 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4748 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4750 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4751 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4752 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4756 for_each_populated_zone(zone) {
4759 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4763 for_each_online_cpu(cpu)
4764 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4773 " active_anon:%lukB"
4774 " inactive_anon:%lukB"
4775 " active_file:%lukB"
4776 " inactive_file:%lukB"
4777 " unevictable:%lukB"
4778 " writepending:%lukB"
4782 " kernel_stack:%lukB"
4790 K(zone_page_state(zone, NR_FREE_PAGES)),
4791 K(min_wmark_pages(zone)),
4792 K(low_wmark_pages(zone)),
4793 K(high_wmark_pages(zone)),
4794 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4795 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4796 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4797 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4798 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4799 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4800 K(zone->present_pages),
4801 K(zone->managed_pages),
4802 K(zone_page_state(zone, NR_MLOCK)),
4803 zone_page_state(zone, NR_KERNEL_STACK_KB),
4804 K(zone_page_state(zone, NR_PAGETABLE)),
4805 K(zone_page_state(zone, NR_BOUNCE)),
4807 K(this_cpu_read(zone->pageset->pcp.count)),
4808 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4809 printk("lowmem_reserve[]:");
4810 for (i = 0; i < MAX_NR_ZONES; i++)
4811 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4812 printk(KERN_CONT "\n");
4815 for_each_populated_zone(zone) {
4817 unsigned long nr[MAX_ORDER], flags, total = 0;
4818 unsigned char types[MAX_ORDER];
4820 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4823 printk(KERN_CONT "%s: ", zone->name);
4825 spin_lock_irqsave(&zone->lock, flags);
4826 for (order = 0; order < MAX_ORDER; order++) {
4827 struct free_area *area = &zone->free_area[order];
4830 nr[order] = area->nr_free;
4831 total += nr[order] << order;
4834 for (type = 0; type < MIGRATE_TYPES; type++) {
4835 if (!list_empty(&area->free_list[type]))
4836 types[order] |= 1 << type;
4839 spin_unlock_irqrestore(&zone->lock, flags);
4840 for (order = 0; order < MAX_ORDER; order++) {
4841 printk(KERN_CONT "%lu*%lukB ",
4842 nr[order], K(1UL) << order);
4844 show_migration_types(types[order]);
4846 printk(KERN_CONT "= %lukB\n", K(total));
4849 hugetlb_show_meminfo();
4851 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4853 show_swap_cache_info();
4856 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4858 zoneref->zone = zone;
4859 zoneref->zone_idx = zone_idx(zone);
4863 * Builds allocation fallback zone lists.
4865 * Add all populated zones of a node to the zonelist.
4867 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4870 enum zone_type zone_type = MAX_NR_ZONES;
4875 zone = pgdat->node_zones + zone_type;
4876 if (managed_zone(zone)) {
4877 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4878 check_highest_zone(zone_type);
4880 } while (zone_type);
4887 static int __parse_numa_zonelist_order(char *s)
4890 * We used to support different zonlists modes but they turned
4891 * out to be just not useful. Let's keep the warning in place
4892 * if somebody still use the cmd line parameter so that we do
4893 * not fail it silently
4895 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4896 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4902 static __init int setup_numa_zonelist_order(char *s)
4907 return __parse_numa_zonelist_order(s);
4909 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4911 char numa_zonelist_order[] = "Node";
4914 * sysctl handler for numa_zonelist_order
4916 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4917 void __user *buffer, size_t *length,
4924 return proc_dostring(table, write, buffer, length, ppos);
4925 str = memdup_user_nul(buffer, 16);
4927 return PTR_ERR(str);
4929 ret = __parse_numa_zonelist_order(str);
4935 #define MAX_NODE_LOAD (nr_online_nodes)
4936 static int node_load[MAX_NUMNODES];
4939 * find_next_best_node - find the next node that should appear in a given node's fallback list
4940 * @node: node whose fallback list we're appending
4941 * @used_node_mask: nodemask_t of already used nodes
4943 * We use a number of factors to determine which is the next node that should
4944 * appear on a given node's fallback list. The node should not have appeared
4945 * already in @node's fallback list, and it should be the next closest node
4946 * according to the distance array (which contains arbitrary distance values
4947 * from each node to each node in the system), and should also prefer nodes
4948 * with no CPUs, since presumably they'll have very little allocation pressure
4949 * on them otherwise.
4950 * It returns -1 if no node is found.
4952 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4955 int min_val = INT_MAX;
4956 int best_node = NUMA_NO_NODE;
4957 const struct cpumask *tmp = cpumask_of_node(0);
4959 /* Use the local node if we haven't already */
4960 if (!node_isset(node, *used_node_mask)) {
4961 node_set(node, *used_node_mask);
4965 for_each_node_state(n, N_MEMORY) {
4967 /* Don't want a node to appear more than once */
4968 if (node_isset(n, *used_node_mask))
4971 /* Use the distance array to find the distance */
4972 val = node_distance(node, n);
4974 /* Penalize nodes under us ("prefer the next node") */
4977 /* Give preference to headless and unused nodes */
4978 tmp = cpumask_of_node(n);
4979 if (!cpumask_empty(tmp))
4980 val += PENALTY_FOR_NODE_WITH_CPUS;
4982 /* Slight preference for less loaded node */
4983 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4984 val += node_load[n];
4986 if (val < min_val) {
4993 node_set(best_node, *used_node_mask);
5000 * Build zonelists ordered by node and zones within node.
5001 * This results in maximum locality--normal zone overflows into local
5002 * DMA zone, if any--but risks exhausting DMA zone.
5004 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5007 struct zoneref *zonerefs;
5010 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5012 for (i = 0; i < nr_nodes; i++) {
5015 pg_data_t *node = NODE_DATA(node_order[i]);
5017 nr_zones = build_zonerefs_node(node, zonerefs);
5018 zonerefs += nr_zones;
5020 zonerefs->zone = NULL;
5021 zonerefs->zone_idx = 0;
5025 * Build gfp_thisnode zonelists
5027 static void build_thisnode_zonelists(pg_data_t *pgdat)
5029 struct zoneref *zonerefs;
5032 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5033 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5034 zonerefs += nr_zones;
5035 zonerefs->zone = NULL;
5036 zonerefs->zone_idx = 0;
5040 * Build zonelists ordered by zone and nodes within zones.
5041 * This results in conserving DMA zone[s] until all Normal memory is
5042 * exhausted, but results in overflowing to remote node while memory
5043 * may still exist in local DMA zone.
5046 static void build_zonelists(pg_data_t *pgdat)
5048 static int node_order[MAX_NUMNODES];
5049 int node, load, nr_nodes = 0;
5050 nodemask_t used_mask;
5051 int local_node, prev_node;
5053 /* NUMA-aware ordering of nodes */
5054 local_node = pgdat->node_id;
5055 load = nr_online_nodes;
5056 prev_node = local_node;
5057 nodes_clear(used_mask);
5059 memset(node_order, 0, sizeof(node_order));
5060 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5062 * We don't want to pressure a particular node.
5063 * So adding penalty to the first node in same
5064 * distance group to make it round-robin.
5066 if (node_distance(local_node, node) !=
5067 node_distance(local_node, prev_node))
5068 node_load[node] = load;
5070 node_order[nr_nodes++] = node;
5075 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5076 build_thisnode_zonelists(pgdat);
5079 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5081 * Return node id of node used for "local" allocations.
5082 * I.e., first node id of first zone in arg node's generic zonelist.
5083 * Used for initializing percpu 'numa_mem', which is used primarily
5084 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5086 int local_memory_node(int node)
5090 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5091 gfp_zone(GFP_KERNEL),
5093 return z->zone->node;
5097 static void setup_min_unmapped_ratio(void);
5098 static void setup_min_slab_ratio(void);
5099 #else /* CONFIG_NUMA */
5101 static void build_zonelists(pg_data_t *pgdat)
5103 int node, local_node;
5104 struct zoneref *zonerefs;
5107 local_node = pgdat->node_id;
5109 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5110 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5111 zonerefs += nr_zones;
5114 * Now we build the zonelist so that it contains the zones
5115 * of all the other nodes.
5116 * We don't want to pressure a particular node, so when
5117 * building the zones for node N, we make sure that the
5118 * zones coming right after the local ones are those from
5119 * node N+1 (modulo N)
5121 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5122 if (!node_online(node))
5124 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5125 zonerefs += nr_zones;
5127 for (node = 0; node < local_node; node++) {
5128 if (!node_online(node))
5130 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5131 zonerefs += nr_zones;
5134 zonerefs->zone = NULL;
5135 zonerefs->zone_idx = 0;
5138 #endif /* CONFIG_NUMA */
5141 * Boot pageset table. One per cpu which is going to be used for all
5142 * zones and all nodes. The parameters will be set in such a way
5143 * that an item put on a list will immediately be handed over to
5144 * the buddy list. This is safe since pageset manipulation is done
5145 * with interrupts disabled.
5147 * The boot_pagesets must be kept even after bootup is complete for
5148 * unused processors and/or zones. They do play a role for bootstrapping
5149 * hotplugged processors.
5151 * zoneinfo_show() and maybe other functions do
5152 * not check if the processor is online before following the pageset pointer.
5153 * Other parts of the kernel may not check if the zone is available.
5155 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5156 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5157 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5159 static void __build_all_zonelists(void *data)
5162 int __maybe_unused cpu;
5163 pg_data_t *self = data;
5164 static DEFINE_SPINLOCK(lock);
5169 memset(node_load, 0, sizeof(node_load));
5173 * This node is hotadded and no memory is yet present. So just
5174 * building zonelists is fine - no need to touch other nodes.
5176 if (self && !node_online(self->node_id)) {
5177 build_zonelists(self);
5179 for_each_online_node(nid) {
5180 pg_data_t *pgdat = NODE_DATA(nid);
5182 build_zonelists(pgdat);
5185 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5187 * We now know the "local memory node" for each node--
5188 * i.e., the node of the first zone in the generic zonelist.
5189 * Set up numa_mem percpu variable for on-line cpus. During
5190 * boot, only the boot cpu should be on-line; we'll init the
5191 * secondary cpus' numa_mem as they come on-line. During
5192 * node/memory hotplug, we'll fixup all on-line cpus.
5194 for_each_online_cpu(cpu)
5195 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5202 static noinline void __init
5203 build_all_zonelists_init(void)
5207 __build_all_zonelists(NULL);
5210 * Initialize the boot_pagesets that are going to be used
5211 * for bootstrapping processors. The real pagesets for
5212 * each zone will be allocated later when the per cpu
5213 * allocator is available.
5215 * boot_pagesets are used also for bootstrapping offline
5216 * cpus if the system is already booted because the pagesets
5217 * are needed to initialize allocators on a specific cpu too.
5218 * F.e. the percpu allocator needs the page allocator which
5219 * needs the percpu allocator in order to allocate its pagesets
5220 * (a chicken-egg dilemma).
5222 for_each_possible_cpu(cpu)
5223 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5225 mminit_verify_zonelist();
5226 cpuset_init_current_mems_allowed();
5230 * unless system_state == SYSTEM_BOOTING.
5232 * __ref due to call of __init annotated helper build_all_zonelists_init
5233 * [protected by SYSTEM_BOOTING].
5235 void __ref build_all_zonelists(pg_data_t *pgdat)
5237 if (system_state == SYSTEM_BOOTING) {
5238 build_all_zonelists_init();
5240 __build_all_zonelists(pgdat);
5241 /* cpuset refresh routine should be here */
5243 vm_total_pages = nr_free_pagecache_pages();
5245 * Disable grouping by mobility if the number of pages in the
5246 * system is too low to allow the mechanism to work. It would be
5247 * more accurate, but expensive to check per-zone. This check is
5248 * made on memory-hotadd so a system can start with mobility
5249 * disabled and enable it later
5251 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5252 page_group_by_mobility_disabled = 1;
5254 page_group_by_mobility_disabled = 0;
5256 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5258 page_group_by_mobility_disabled ? "off" : "on",
5261 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5266 * Initially all pages are reserved - free ones are freed
5267 * up by free_all_bootmem() once the early boot process is
5268 * done. Non-atomic initialization, single-pass.
5270 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5271 unsigned long start_pfn, enum memmap_context context)
5273 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5274 unsigned long end_pfn = start_pfn + size;
5275 pg_data_t *pgdat = NODE_DATA(nid);
5277 unsigned long nr_initialised = 0;
5278 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5279 struct memblock_region *r = NULL, *tmp;
5282 if (highest_memmap_pfn < end_pfn - 1)
5283 highest_memmap_pfn = end_pfn - 1;
5286 * Honor reservation requested by the driver for this ZONE_DEVICE
5289 if (altmap && start_pfn == altmap->base_pfn)
5290 start_pfn += altmap->reserve;
5292 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5294 * There can be holes in boot-time mem_map[]s handed to this
5295 * function. They do not exist on hotplugged memory.
5297 if (context != MEMMAP_EARLY)
5300 if (!early_pfn_valid(pfn)) {
5301 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5303 * Skip to the pfn preceding the next valid one (or
5304 * end_pfn), such that we hit a valid pfn (or end_pfn)
5305 * on our next iteration of the loop.
5307 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5311 if (!early_pfn_in_nid(pfn, nid))
5313 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5316 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5318 * Check given memblock attribute by firmware which can affect
5319 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5320 * mirrored, it's an overlapped memmap init. skip it.
5322 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5323 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5324 for_each_memblock(memory, tmp)
5325 if (pfn < memblock_region_memory_end_pfn(tmp))
5329 if (pfn >= memblock_region_memory_base_pfn(r) &&
5330 memblock_is_mirror(r)) {
5331 /* already initialized as NORMAL */
5332 pfn = memblock_region_memory_end_pfn(r);
5340 * Mark the block movable so that blocks are reserved for
5341 * movable at startup. This will force kernel allocations
5342 * to reserve their blocks rather than leaking throughout
5343 * the address space during boot when many long-lived
5344 * kernel allocations are made.
5346 * bitmap is created for zone's valid pfn range. but memmap
5347 * can be created for invalid pages (for alignment)
5348 * check here not to call set_pageblock_migratetype() against
5351 if (!(pfn & (pageblock_nr_pages - 1))) {
5352 struct page *page = pfn_to_page(pfn);
5354 __init_single_page(page, pfn, zone, nid);
5355 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5358 __init_single_pfn(pfn, zone, nid);
5363 static void __meminit zone_init_free_lists(struct zone *zone)
5365 unsigned int order, t;
5366 for_each_migratetype_order(order, t) {
5367 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5368 zone->free_area[order].nr_free = 0;
5372 #ifndef __HAVE_ARCH_MEMMAP_INIT
5373 #define memmap_init(size, nid, zone, start_pfn) \
5374 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5377 static int zone_batchsize(struct zone *zone)
5383 * The per-cpu-pages pools are set to around 1000th of the
5384 * size of the zone. But no more than 1/2 of a meg.
5386 * OK, so we don't know how big the cache is. So guess.
5388 batch = zone->managed_pages / 1024;
5389 if (batch * PAGE_SIZE > 512 * 1024)
5390 batch = (512 * 1024) / PAGE_SIZE;
5391 batch /= 4; /* We effectively *= 4 below */
5396 * Clamp the batch to a 2^n - 1 value. Having a power
5397 * of 2 value was found to be more likely to have
5398 * suboptimal cache aliasing properties in some cases.
5400 * For example if 2 tasks are alternately allocating
5401 * batches of pages, one task can end up with a lot
5402 * of pages of one half of the possible page colors
5403 * and the other with pages of the other colors.
5405 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5410 /* The deferral and batching of frees should be suppressed under NOMMU
5413 * The problem is that NOMMU needs to be able to allocate large chunks
5414 * of contiguous memory as there's no hardware page translation to
5415 * assemble apparent contiguous memory from discontiguous pages.
5417 * Queueing large contiguous runs of pages for batching, however,
5418 * causes the pages to actually be freed in smaller chunks. As there
5419 * can be a significant delay between the individual batches being
5420 * recycled, this leads to the once large chunks of space being
5421 * fragmented and becoming unavailable for high-order allocations.
5428 * pcp->high and pcp->batch values are related and dependent on one another:
5429 * ->batch must never be higher then ->high.
5430 * The following function updates them in a safe manner without read side
5433 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5434 * those fields changing asynchronously (acording the the above rule).
5436 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5437 * outside of boot time (or some other assurance that no concurrent updaters
5440 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5441 unsigned long batch)
5443 /* start with a fail safe value for batch */
5447 /* Update high, then batch, in order */
5454 /* a companion to pageset_set_high() */
5455 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5457 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5460 static void pageset_init(struct per_cpu_pageset *p)
5462 struct per_cpu_pages *pcp;
5465 memset(p, 0, sizeof(*p));
5469 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5470 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5473 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5476 pageset_set_batch(p, batch);
5480 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5481 * to the value high for the pageset p.
5483 static void pageset_set_high(struct per_cpu_pageset *p,
5486 unsigned long batch = max(1UL, high / 4);
5487 if ((high / 4) > (PAGE_SHIFT * 8))
5488 batch = PAGE_SHIFT * 8;
5490 pageset_update(&p->pcp, high, batch);
5493 static void pageset_set_high_and_batch(struct zone *zone,
5494 struct per_cpu_pageset *pcp)
5496 if (percpu_pagelist_fraction)
5497 pageset_set_high(pcp,
5498 (zone->managed_pages /
5499 percpu_pagelist_fraction));
5501 pageset_set_batch(pcp, zone_batchsize(zone));
5504 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5506 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5509 pageset_set_high_and_batch(zone, pcp);
5512 void __meminit setup_zone_pageset(struct zone *zone)
5515 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5516 for_each_possible_cpu(cpu)
5517 zone_pageset_init(zone, cpu);
5521 * Allocate per cpu pagesets and initialize them.
5522 * Before this call only boot pagesets were available.
5524 void __init setup_per_cpu_pageset(void)
5526 struct pglist_data *pgdat;
5529 for_each_populated_zone(zone)
5530 setup_zone_pageset(zone);
5532 for_each_online_pgdat(pgdat)
5533 pgdat->per_cpu_nodestats =
5534 alloc_percpu(struct per_cpu_nodestat);
5537 static __meminit void zone_pcp_init(struct zone *zone)
5540 * per cpu subsystem is not up at this point. The following code
5541 * relies on the ability of the linker to provide the
5542 * offset of a (static) per cpu variable into the per cpu area.
5544 zone->pageset = &boot_pageset;
5546 if (populated_zone(zone))
5547 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5548 zone->name, zone->present_pages,
5549 zone_batchsize(zone));
5552 void __meminit init_currently_empty_zone(struct zone *zone,
5553 unsigned long zone_start_pfn,
5556 struct pglist_data *pgdat = zone->zone_pgdat;
5558 pgdat->nr_zones = zone_idx(zone) + 1;
5560 zone->zone_start_pfn = zone_start_pfn;
5562 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5563 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5565 (unsigned long)zone_idx(zone),
5566 zone_start_pfn, (zone_start_pfn + size));
5568 zone_init_free_lists(zone);
5569 zone->initialized = 1;
5572 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5573 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5576 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5578 int __meminit __early_pfn_to_nid(unsigned long pfn,
5579 struct mminit_pfnnid_cache *state)
5581 unsigned long start_pfn, end_pfn;
5584 if (state->last_start <= pfn && pfn < state->last_end)
5585 return state->last_nid;
5587 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5589 state->last_start = start_pfn;
5590 state->last_end = end_pfn;
5591 state->last_nid = nid;
5596 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5599 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5600 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5601 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5603 * If an architecture guarantees that all ranges registered contain no holes
5604 * and may be freed, this this function may be used instead of calling
5605 * memblock_free_early_nid() manually.
5607 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5609 unsigned long start_pfn, end_pfn;
5612 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5613 start_pfn = min(start_pfn, max_low_pfn);
5614 end_pfn = min(end_pfn, max_low_pfn);
5616 if (start_pfn < end_pfn)
5617 memblock_free_early_nid(PFN_PHYS(start_pfn),
5618 (end_pfn - start_pfn) << PAGE_SHIFT,
5624 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5625 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5627 * If an architecture guarantees that all ranges registered contain no holes and may
5628 * be freed, this function may be used instead of calling memory_present() manually.
5630 void __init sparse_memory_present_with_active_regions(int nid)
5632 unsigned long start_pfn, end_pfn;
5635 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5636 memory_present(this_nid, start_pfn, end_pfn);
5640 * get_pfn_range_for_nid - Return the start and end page frames for a node
5641 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5642 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5643 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5645 * It returns the start and end page frame of a node based on information
5646 * provided by memblock_set_node(). If called for a node
5647 * with no available memory, a warning is printed and the start and end
5650 void __meminit get_pfn_range_for_nid(unsigned int nid,
5651 unsigned long *start_pfn, unsigned long *end_pfn)
5653 unsigned long this_start_pfn, this_end_pfn;
5659 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5660 *start_pfn = min(*start_pfn, this_start_pfn);
5661 *end_pfn = max(*end_pfn, this_end_pfn);
5664 if (*start_pfn == -1UL)
5669 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5670 * assumption is made that zones within a node are ordered in monotonic
5671 * increasing memory addresses so that the "highest" populated zone is used
5673 static void __init find_usable_zone_for_movable(void)
5676 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5677 if (zone_index == ZONE_MOVABLE)
5680 if (arch_zone_highest_possible_pfn[zone_index] >
5681 arch_zone_lowest_possible_pfn[zone_index])
5685 VM_BUG_ON(zone_index == -1);
5686 movable_zone = zone_index;
5690 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5691 * because it is sized independent of architecture. Unlike the other zones,
5692 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5693 * in each node depending on the size of each node and how evenly kernelcore
5694 * is distributed. This helper function adjusts the zone ranges
5695 * provided by the architecture for a given node by using the end of the
5696 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5697 * zones within a node are in order of monotonic increases memory addresses
5699 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5700 unsigned long zone_type,
5701 unsigned long node_start_pfn,
5702 unsigned long node_end_pfn,
5703 unsigned long *zone_start_pfn,
5704 unsigned long *zone_end_pfn)
5706 /* Only adjust if ZONE_MOVABLE is on this node */
5707 if (zone_movable_pfn[nid]) {
5708 /* Size ZONE_MOVABLE */
5709 if (zone_type == ZONE_MOVABLE) {
5710 *zone_start_pfn = zone_movable_pfn[nid];
5711 *zone_end_pfn = min(node_end_pfn,
5712 arch_zone_highest_possible_pfn[movable_zone]);
5714 /* Adjust for ZONE_MOVABLE starting within this range */
5715 } else if (!mirrored_kernelcore &&
5716 *zone_start_pfn < zone_movable_pfn[nid] &&
5717 *zone_end_pfn > zone_movable_pfn[nid]) {
5718 *zone_end_pfn = zone_movable_pfn[nid];
5720 /* Check if this whole range is within ZONE_MOVABLE */
5721 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5722 *zone_start_pfn = *zone_end_pfn;
5727 * Return the number of pages a zone spans in a node, including holes
5728 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5730 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5731 unsigned long zone_type,
5732 unsigned long node_start_pfn,
5733 unsigned long node_end_pfn,
5734 unsigned long *zone_start_pfn,
5735 unsigned long *zone_end_pfn,
5736 unsigned long *ignored)
5738 /* When hotadd a new node from cpu_up(), the node should be empty */
5739 if (!node_start_pfn && !node_end_pfn)
5742 /* Get the start and end of the zone */
5743 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5744 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5745 adjust_zone_range_for_zone_movable(nid, zone_type,
5746 node_start_pfn, node_end_pfn,
5747 zone_start_pfn, zone_end_pfn);
5749 /* Check that this node has pages within the zone's required range */
5750 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5753 /* Move the zone boundaries inside the node if necessary */
5754 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5755 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5757 /* Return the spanned pages */
5758 return *zone_end_pfn - *zone_start_pfn;
5762 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5763 * then all holes in the requested range will be accounted for.
5765 unsigned long __meminit __absent_pages_in_range(int nid,
5766 unsigned long range_start_pfn,
5767 unsigned long range_end_pfn)
5769 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5770 unsigned long start_pfn, end_pfn;
5773 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5774 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5775 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5776 nr_absent -= end_pfn - start_pfn;
5782 * absent_pages_in_range - Return number of page frames in holes within a range
5783 * @start_pfn: The start PFN to start searching for holes
5784 * @end_pfn: The end PFN to stop searching for holes
5786 * It returns the number of pages frames in memory holes within a range.
5788 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5789 unsigned long end_pfn)
5791 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5794 /* Return the number of page frames in holes in a zone on a node */
5795 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5796 unsigned long zone_type,
5797 unsigned long node_start_pfn,
5798 unsigned long node_end_pfn,
5799 unsigned long *ignored)
5801 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5802 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5803 unsigned long zone_start_pfn, zone_end_pfn;
5804 unsigned long nr_absent;
5806 /* When hotadd a new node from cpu_up(), the node should be empty */
5807 if (!node_start_pfn && !node_end_pfn)
5810 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5811 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5813 adjust_zone_range_for_zone_movable(nid, zone_type,
5814 node_start_pfn, node_end_pfn,
5815 &zone_start_pfn, &zone_end_pfn);
5816 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5819 * ZONE_MOVABLE handling.
5820 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5823 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5824 unsigned long start_pfn, end_pfn;
5825 struct memblock_region *r;
5827 for_each_memblock(memory, r) {
5828 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5829 zone_start_pfn, zone_end_pfn);
5830 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5831 zone_start_pfn, zone_end_pfn);
5833 if (zone_type == ZONE_MOVABLE &&
5834 memblock_is_mirror(r))
5835 nr_absent += end_pfn - start_pfn;
5837 if (zone_type == ZONE_NORMAL &&
5838 !memblock_is_mirror(r))
5839 nr_absent += end_pfn - start_pfn;
5846 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5847 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5848 unsigned long zone_type,
5849 unsigned long node_start_pfn,
5850 unsigned long node_end_pfn,
5851 unsigned long *zone_start_pfn,
5852 unsigned long *zone_end_pfn,
5853 unsigned long *zones_size)
5857 *zone_start_pfn = node_start_pfn;
5858 for (zone = 0; zone < zone_type; zone++)
5859 *zone_start_pfn += zones_size[zone];
5861 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5863 return zones_size[zone_type];
5866 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5867 unsigned long zone_type,
5868 unsigned long node_start_pfn,
5869 unsigned long node_end_pfn,
5870 unsigned long *zholes_size)
5875 return zholes_size[zone_type];
5878 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5880 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5881 unsigned long node_start_pfn,
5882 unsigned long node_end_pfn,
5883 unsigned long *zones_size,
5884 unsigned long *zholes_size)
5886 unsigned long realtotalpages = 0, totalpages = 0;
5889 for (i = 0; i < MAX_NR_ZONES; i++) {
5890 struct zone *zone = pgdat->node_zones + i;
5891 unsigned long zone_start_pfn, zone_end_pfn;
5892 unsigned long size, real_size;
5894 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5900 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5901 node_start_pfn, node_end_pfn,
5904 zone->zone_start_pfn = zone_start_pfn;
5906 zone->zone_start_pfn = 0;
5907 zone->spanned_pages = size;
5908 zone->present_pages = real_size;
5911 realtotalpages += real_size;
5914 pgdat->node_spanned_pages = totalpages;
5915 pgdat->node_present_pages = realtotalpages;
5916 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5920 #ifndef CONFIG_SPARSEMEM
5922 * Calculate the size of the zone->blockflags rounded to an unsigned long
5923 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5924 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5925 * round what is now in bits to nearest long in bits, then return it in
5928 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5930 unsigned long usemapsize;
5932 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5933 usemapsize = roundup(zonesize, pageblock_nr_pages);
5934 usemapsize = usemapsize >> pageblock_order;
5935 usemapsize *= NR_PAGEBLOCK_BITS;
5936 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5938 return usemapsize / 8;
5941 static void __init setup_usemap(struct pglist_data *pgdat,
5943 unsigned long zone_start_pfn,
5944 unsigned long zonesize)
5946 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5947 zone->pageblock_flags = NULL;
5949 zone->pageblock_flags =
5950 memblock_virt_alloc_node_nopanic(usemapsize,
5954 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5955 unsigned long zone_start_pfn, unsigned long zonesize) {}
5956 #endif /* CONFIG_SPARSEMEM */
5958 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5960 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5961 void __paginginit set_pageblock_order(void)
5965 /* Check that pageblock_nr_pages has not already been setup */
5966 if (pageblock_order)
5969 if (HPAGE_SHIFT > PAGE_SHIFT)
5970 order = HUGETLB_PAGE_ORDER;
5972 order = MAX_ORDER - 1;
5975 * Assume the largest contiguous order of interest is a huge page.
5976 * This value may be variable depending on boot parameters on IA64 and
5979 pageblock_order = order;
5981 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5984 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5985 * is unused as pageblock_order is set at compile-time. See
5986 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5989 void __paginginit set_pageblock_order(void)
5993 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5995 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5996 unsigned long present_pages)
5998 unsigned long pages = spanned_pages;
6001 * Provide a more accurate estimation if there are holes within
6002 * the zone and SPARSEMEM is in use. If there are holes within the
6003 * zone, each populated memory region may cost us one or two extra
6004 * memmap pages due to alignment because memmap pages for each
6005 * populated regions may not be naturally aligned on page boundary.
6006 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6008 if (spanned_pages > present_pages + (present_pages >> 4) &&
6009 IS_ENABLED(CONFIG_SPARSEMEM))
6010 pages = present_pages;
6012 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6016 * Set up the zone data structures:
6017 * - mark all pages reserved
6018 * - mark all memory queues empty
6019 * - clear the memory bitmaps
6021 * NOTE: pgdat should get zeroed by caller.
6023 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6026 int nid = pgdat->node_id;
6028 pgdat_resize_init(pgdat);
6029 #ifdef CONFIG_NUMA_BALANCING
6030 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6031 pgdat->numabalancing_migrate_nr_pages = 0;
6032 pgdat->numabalancing_migrate_next_window = jiffies;
6034 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6035 spin_lock_init(&pgdat->split_queue_lock);
6036 INIT_LIST_HEAD(&pgdat->split_queue);
6037 pgdat->split_queue_len = 0;
6039 init_waitqueue_head(&pgdat->kswapd_wait);
6040 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6041 #ifdef CONFIG_COMPACTION
6042 init_waitqueue_head(&pgdat->kcompactd_wait);
6044 pgdat_page_ext_init(pgdat);
6045 spin_lock_init(&pgdat->lru_lock);
6046 lruvec_init(node_lruvec(pgdat));
6048 pgdat->per_cpu_nodestats = &boot_nodestats;
6050 for (j = 0; j < MAX_NR_ZONES; j++) {
6051 struct zone *zone = pgdat->node_zones + j;
6052 unsigned long size, realsize, freesize, memmap_pages;
6053 unsigned long zone_start_pfn = zone->zone_start_pfn;
6055 size = zone->spanned_pages;
6056 realsize = freesize = zone->present_pages;
6059 * Adjust freesize so that it accounts for how much memory
6060 * is used by this zone for memmap. This affects the watermark
6061 * and per-cpu initialisations
6063 memmap_pages = calc_memmap_size(size, realsize);
6064 if (!is_highmem_idx(j)) {
6065 if (freesize >= memmap_pages) {
6066 freesize -= memmap_pages;
6069 " %s zone: %lu pages used for memmap\n",
6070 zone_names[j], memmap_pages);
6072 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6073 zone_names[j], memmap_pages, freesize);
6076 /* Account for reserved pages */
6077 if (j == 0 && freesize > dma_reserve) {
6078 freesize -= dma_reserve;
6079 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6080 zone_names[0], dma_reserve);
6083 if (!is_highmem_idx(j))
6084 nr_kernel_pages += freesize;
6085 /* Charge for highmem memmap if there are enough kernel pages */
6086 else if (nr_kernel_pages > memmap_pages * 2)
6087 nr_kernel_pages -= memmap_pages;
6088 nr_all_pages += freesize;
6091 * Set an approximate value for lowmem here, it will be adjusted
6092 * when the bootmem allocator frees pages into the buddy system.
6093 * And all highmem pages will be managed by the buddy system.
6095 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6099 zone->name = zone_names[j];
6100 zone->zone_pgdat = pgdat;
6101 spin_lock_init(&zone->lock);
6102 zone_seqlock_init(zone);
6103 zone_pcp_init(zone);
6108 set_pageblock_order();
6109 setup_usemap(pgdat, zone, zone_start_pfn, size);
6110 init_currently_empty_zone(zone, zone_start_pfn, size);
6111 memmap_init(size, nid, j, zone_start_pfn);
6115 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6117 unsigned long __maybe_unused start = 0;
6118 unsigned long __maybe_unused offset = 0;
6120 /* Skip empty nodes */
6121 if (!pgdat->node_spanned_pages)
6124 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6125 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6126 offset = pgdat->node_start_pfn - start;
6127 /* ia64 gets its own node_mem_map, before this, without bootmem */
6128 if (!pgdat->node_mem_map) {
6129 unsigned long size, end;
6133 * The zone's endpoints aren't required to be MAX_ORDER
6134 * aligned but the node_mem_map endpoints must be in order
6135 * for the buddy allocator to function correctly.
6137 end = pgdat_end_pfn(pgdat);
6138 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6139 size = (end - start) * sizeof(struct page);
6140 map = alloc_remap(pgdat->node_id, size);
6142 map = memblock_virt_alloc_node_nopanic(size,
6144 pgdat->node_mem_map = map + offset;
6146 #ifndef CONFIG_NEED_MULTIPLE_NODES
6148 * With no DISCONTIG, the global mem_map is just set as node 0's
6150 if (pgdat == NODE_DATA(0)) {
6151 mem_map = NODE_DATA(0)->node_mem_map;
6152 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6153 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6155 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6158 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6161 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6162 unsigned long node_start_pfn, unsigned long *zholes_size)
6164 pg_data_t *pgdat = NODE_DATA(nid);
6165 unsigned long start_pfn = 0;
6166 unsigned long end_pfn = 0;
6168 /* pg_data_t should be reset to zero when it's allocated */
6169 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6171 pgdat->node_id = nid;
6172 pgdat->node_start_pfn = node_start_pfn;
6173 pgdat->per_cpu_nodestats = NULL;
6174 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6175 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6176 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6177 (u64)start_pfn << PAGE_SHIFT,
6178 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6180 start_pfn = node_start_pfn;
6182 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6183 zones_size, zholes_size);
6185 alloc_node_mem_map(pgdat);
6186 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6187 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6188 nid, (unsigned long)pgdat,
6189 (unsigned long)pgdat->node_mem_map);
6192 reset_deferred_meminit(pgdat);
6193 free_area_init_core(pgdat);
6196 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6198 #if MAX_NUMNODES > 1
6200 * Figure out the number of possible node ids.
6202 void __init setup_nr_node_ids(void)
6204 unsigned int highest;
6206 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6207 nr_node_ids = highest + 1;
6212 * node_map_pfn_alignment - determine the maximum internode alignment
6214 * This function should be called after node map is populated and sorted.
6215 * It calculates the maximum power of two alignment which can distinguish
6218 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6219 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6220 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6221 * shifted, 1GiB is enough and this function will indicate so.
6223 * This is used to test whether pfn -> nid mapping of the chosen memory
6224 * model has fine enough granularity to avoid incorrect mapping for the
6225 * populated node map.
6227 * Returns the determined alignment in pfn's. 0 if there is no alignment
6228 * requirement (single node).
6230 unsigned long __init node_map_pfn_alignment(void)
6232 unsigned long accl_mask = 0, last_end = 0;
6233 unsigned long start, end, mask;
6237 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6238 if (!start || last_nid < 0 || last_nid == nid) {
6245 * Start with a mask granular enough to pin-point to the
6246 * start pfn and tick off bits one-by-one until it becomes
6247 * too coarse to separate the current node from the last.
6249 mask = ~((1 << __ffs(start)) - 1);
6250 while (mask && last_end <= (start & (mask << 1)))
6253 /* accumulate all internode masks */
6257 /* convert mask to number of pages */
6258 return ~accl_mask + 1;
6261 /* Find the lowest pfn for a node */
6262 static unsigned long __init find_min_pfn_for_node(int nid)
6264 unsigned long min_pfn = ULONG_MAX;
6265 unsigned long start_pfn;
6268 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6269 min_pfn = min(min_pfn, start_pfn);
6271 if (min_pfn == ULONG_MAX) {
6272 pr_warn("Could not find start_pfn for node %d\n", nid);
6280 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6282 * It returns the minimum PFN based on information provided via
6283 * memblock_set_node().
6285 unsigned long __init find_min_pfn_with_active_regions(void)
6287 return find_min_pfn_for_node(MAX_NUMNODES);
6291 * early_calculate_totalpages()
6292 * Sum pages in active regions for movable zone.
6293 * Populate N_MEMORY for calculating usable_nodes.
6295 static unsigned long __init early_calculate_totalpages(void)
6297 unsigned long totalpages = 0;
6298 unsigned long start_pfn, end_pfn;
6301 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6302 unsigned long pages = end_pfn - start_pfn;
6304 totalpages += pages;
6306 node_set_state(nid, N_MEMORY);
6312 * Find the PFN the Movable zone begins in each node. Kernel memory
6313 * is spread evenly between nodes as long as the nodes have enough
6314 * memory. When they don't, some nodes will have more kernelcore than
6317 static void __init find_zone_movable_pfns_for_nodes(void)
6320 unsigned long usable_startpfn;
6321 unsigned long kernelcore_node, kernelcore_remaining;
6322 /* save the state before borrow the nodemask */
6323 nodemask_t saved_node_state = node_states[N_MEMORY];
6324 unsigned long totalpages = early_calculate_totalpages();
6325 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6326 struct memblock_region *r;
6328 /* Need to find movable_zone earlier when movable_node is specified. */
6329 find_usable_zone_for_movable();
6332 * If movable_node is specified, ignore kernelcore and movablecore
6335 if (movable_node_is_enabled()) {
6336 for_each_memblock(memory, r) {
6337 if (!memblock_is_hotpluggable(r))
6342 usable_startpfn = PFN_DOWN(r->base);
6343 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6344 min(usable_startpfn, zone_movable_pfn[nid]) :
6352 * If kernelcore=mirror is specified, ignore movablecore option
6354 if (mirrored_kernelcore) {
6355 bool mem_below_4gb_not_mirrored = false;
6357 for_each_memblock(memory, r) {
6358 if (memblock_is_mirror(r))
6363 usable_startpfn = memblock_region_memory_base_pfn(r);
6365 if (usable_startpfn < 0x100000) {
6366 mem_below_4gb_not_mirrored = true;
6370 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6371 min(usable_startpfn, zone_movable_pfn[nid]) :
6375 if (mem_below_4gb_not_mirrored)
6376 pr_warn("This configuration results in unmirrored kernel memory.");
6382 * If movablecore=nn[KMG] was specified, calculate what size of
6383 * kernelcore that corresponds so that memory usable for
6384 * any allocation type is evenly spread. If both kernelcore
6385 * and movablecore are specified, then the value of kernelcore
6386 * will be used for required_kernelcore if it's greater than
6387 * what movablecore would have allowed.
6389 if (required_movablecore) {
6390 unsigned long corepages;
6393 * Round-up so that ZONE_MOVABLE is at least as large as what
6394 * was requested by the user
6396 required_movablecore =
6397 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6398 required_movablecore = min(totalpages, required_movablecore);
6399 corepages = totalpages - required_movablecore;
6401 required_kernelcore = max(required_kernelcore, corepages);
6405 * If kernelcore was not specified or kernelcore size is larger
6406 * than totalpages, there is no ZONE_MOVABLE.
6408 if (!required_kernelcore || required_kernelcore >= totalpages)
6411 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6412 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6415 /* Spread kernelcore memory as evenly as possible throughout nodes */
6416 kernelcore_node = required_kernelcore / usable_nodes;
6417 for_each_node_state(nid, N_MEMORY) {
6418 unsigned long start_pfn, end_pfn;
6421 * Recalculate kernelcore_node if the division per node
6422 * now exceeds what is necessary to satisfy the requested
6423 * amount of memory for the kernel
6425 if (required_kernelcore < kernelcore_node)
6426 kernelcore_node = required_kernelcore / usable_nodes;
6429 * As the map is walked, we track how much memory is usable
6430 * by the kernel using kernelcore_remaining. When it is
6431 * 0, the rest of the node is usable by ZONE_MOVABLE
6433 kernelcore_remaining = kernelcore_node;
6435 /* Go through each range of PFNs within this node */
6436 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6437 unsigned long size_pages;
6439 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6440 if (start_pfn >= end_pfn)
6443 /* Account for what is only usable for kernelcore */
6444 if (start_pfn < usable_startpfn) {
6445 unsigned long kernel_pages;
6446 kernel_pages = min(end_pfn, usable_startpfn)
6449 kernelcore_remaining -= min(kernel_pages,
6450 kernelcore_remaining);
6451 required_kernelcore -= min(kernel_pages,
6452 required_kernelcore);
6454 /* Continue if range is now fully accounted */
6455 if (end_pfn <= usable_startpfn) {
6458 * Push zone_movable_pfn to the end so
6459 * that if we have to rebalance
6460 * kernelcore across nodes, we will
6461 * not double account here
6463 zone_movable_pfn[nid] = end_pfn;
6466 start_pfn = usable_startpfn;
6470 * The usable PFN range for ZONE_MOVABLE is from
6471 * start_pfn->end_pfn. Calculate size_pages as the
6472 * number of pages used as kernelcore
6474 size_pages = end_pfn - start_pfn;
6475 if (size_pages > kernelcore_remaining)
6476 size_pages = kernelcore_remaining;
6477 zone_movable_pfn[nid] = start_pfn + size_pages;
6480 * Some kernelcore has been met, update counts and
6481 * break if the kernelcore for this node has been
6484 required_kernelcore -= min(required_kernelcore,
6486 kernelcore_remaining -= size_pages;
6487 if (!kernelcore_remaining)
6493 * If there is still required_kernelcore, we do another pass with one
6494 * less node in the count. This will push zone_movable_pfn[nid] further
6495 * along on the nodes that still have memory until kernelcore is
6499 if (usable_nodes && required_kernelcore > usable_nodes)
6503 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6504 for (nid = 0; nid < MAX_NUMNODES; nid++)
6505 zone_movable_pfn[nid] =
6506 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6509 /* restore the node_state */
6510 node_states[N_MEMORY] = saved_node_state;
6513 /* Any regular or high memory on that node ? */
6514 static void check_for_memory(pg_data_t *pgdat, int nid)
6516 enum zone_type zone_type;
6518 if (N_MEMORY == N_NORMAL_MEMORY)
6521 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6522 struct zone *zone = &pgdat->node_zones[zone_type];
6523 if (populated_zone(zone)) {
6524 node_set_state(nid, N_HIGH_MEMORY);
6525 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6526 zone_type <= ZONE_NORMAL)
6527 node_set_state(nid, N_NORMAL_MEMORY);
6534 * free_area_init_nodes - Initialise all pg_data_t and zone data
6535 * @max_zone_pfn: an array of max PFNs for each zone
6537 * This will call free_area_init_node() for each active node in the system.
6538 * Using the page ranges provided by memblock_set_node(), the size of each
6539 * zone in each node and their holes is calculated. If the maximum PFN
6540 * between two adjacent zones match, it is assumed that the zone is empty.
6541 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6542 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6543 * starts where the previous one ended. For example, ZONE_DMA32 starts
6544 * at arch_max_dma_pfn.
6546 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6548 unsigned long start_pfn, end_pfn;
6551 /* Record where the zone boundaries are */
6552 memset(arch_zone_lowest_possible_pfn, 0,
6553 sizeof(arch_zone_lowest_possible_pfn));
6554 memset(arch_zone_highest_possible_pfn, 0,
6555 sizeof(arch_zone_highest_possible_pfn));
6557 start_pfn = find_min_pfn_with_active_regions();
6559 for (i = 0; i < MAX_NR_ZONES; i++) {
6560 if (i == ZONE_MOVABLE)
6563 end_pfn = max(max_zone_pfn[i], start_pfn);
6564 arch_zone_lowest_possible_pfn[i] = start_pfn;
6565 arch_zone_highest_possible_pfn[i] = end_pfn;
6567 start_pfn = end_pfn;
6570 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6571 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6572 find_zone_movable_pfns_for_nodes();
6574 /* Print out the zone ranges */
6575 pr_info("Zone ranges:\n");
6576 for (i = 0; i < MAX_NR_ZONES; i++) {
6577 if (i == ZONE_MOVABLE)
6579 pr_info(" %-8s ", zone_names[i]);
6580 if (arch_zone_lowest_possible_pfn[i] ==
6581 arch_zone_highest_possible_pfn[i])
6584 pr_cont("[mem %#018Lx-%#018Lx]\n",
6585 (u64)arch_zone_lowest_possible_pfn[i]
6587 ((u64)arch_zone_highest_possible_pfn[i]
6588 << PAGE_SHIFT) - 1);
6591 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6592 pr_info("Movable zone start for each node\n");
6593 for (i = 0; i < MAX_NUMNODES; i++) {
6594 if (zone_movable_pfn[i])
6595 pr_info(" Node %d: %#018Lx\n", i,
6596 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6599 /* Print out the early node map */
6600 pr_info("Early memory node ranges\n");
6601 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6602 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6603 (u64)start_pfn << PAGE_SHIFT,
6604 ((u64)end_pfn << PAGE_SHIFT) - 1);
6606 /* Initialise every node */
6607 mminit_verify_pageflags_layout();
6608 setup_nr_node_ids();
6609 for_each_online_node(nid) {
6610 pg_data_t *pgdat = NODE_DATA(nid);
6611 free_area_init_node(nid, NULL,
6612 find_min_pfn_for_node(nid), NULL);
6614 /* Any memory on that node */
6615 if (pgdat->node_present_pages)
6616 node_set_state(nid, N_MEMORY);
6617 check_for_memory(pgdat, nid);
6621 static int __init cmdline_parse_core(char *p, unsigned long *core)
6623 unsigned long long coremem;
6627 coremem = memparse(p, &p);
6628 *core = coremem >> PAGE_SHIFT;
6630 /* Paranoid check that UL is enough for the coremem value */
6631 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6637 * kernelcore=size sets the amount of memory for use for allocations that
6638 * cannot be reclaimed or migrated.
6640 static int __init cmdline_parse_kernelcore(char *p)
6642 /* parse kernelcore=mirror */
6643 if (parse_option_str(p, "mirror")) {
6644 mirrored_kernelcore = true;
6648 return cmdline_parse_core(p, &required_kernelcore);
6652 * movablecore=size sets the amount of memory for use for allocations that
6653 * can be reclaimed or migrated.
6655 static int __init cmdline_parse_movablecore(char *p)
6657 return cmdline_parse_core(p, &required_movablecore);
6660 early_param("kernelcore", cmdline_parse_kernelcore);
6661 early_param("movablecore", cmdline_parse_movablecore);
6663 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6665 void adjust_managed_page_count(struct page *page, long count)
6667 spin_lock(&managed_page_count_lock);
6668 page_zone(page)->managed_pages += count;
6669 totalram_pages += count;
6670 #ifdef CONFIG_HIGHMEM
6671 if (PageHighMem(page))
6672 totalhigh_pages += count;
6674 spin_unlock(&managed_page_count_lock);
6676 EXPORT_SYMBOL(adjust_managed_page_count);
6678 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6681 unsigned long pages = 0;
6683 start = (void *)PAGE_ALIGN((unsigned long)start);
6684 end = (void *)((unsigned long)end & PAGE_MASK);
6685 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6686 if ((unsigned int)poison <= 0xFF)
6687 memset(pos, poison, PAGE_SIZE);
6688 free_reserved_page(virt_to_page(pos));
6692 pr_info("Freeing %s memory: %ldK\n",
6693 s, pages << (PAGE_SHIFT - 10));
6697 EXPORT_SYMBOL(free_reserved_area);
6699 #ifdef CONFIG_HIGHMEM
6700 void free_highmem_page(struct page *page)
6702 __free_reserved_page(page);
6704 page_zone(page)->managed_pages++;
6710 void __init mem_init_print_info(const char *str)
6712 unsigned long physpages, codesize, datasize, rosize, bss_size;
6713 unsigned long init_code_size, init_data_size;
6715 physpages = get_num_physpages();
6716 codesize = _etext - _stext;
6717 datasize = _edata - _sdata;
6718 rosize = __end_rodata - __start_rodata;
6719 bss_size = __bss_stop - __bss_start;
6720 init_data_size = __init_end - __init_begin;
6721 init_code_size = _einittext - _sinittext;
6724 * Detect special cases and adjust section sizes accordingly:
6725 * 1) .init.* may be embedded into .data sections
6726 * 2) .init.text.* may be out of [__init_begin, __init_end],
6727 * please refer to arch/tile/kernel/vmlinux.lds.S.
6728 * 3) .rodata.* may be embedded into .text or .data sections.
6730 #define adj_init_size(start, end, size, pos, adj) \
6732 if (start <= pos && pos < end && size > adj) \
6736 adj_init_size(__init_begin, __init_end, init_data_size,
6737 _sinittext, init_code_size);
6738 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6739 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6740 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6741 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6743 #undef adj_init_size
6745 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6746 #ifdef CONFIG_HIGHMEM
6750 nr_free_pages() << (PAGE_SHIFT - 10),
6751 physpages << (PAGE_SHIFT - 10),
6752 codesize >> 10, datasize >> 10, rosize >> 10,
6753 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6754 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6755 totalcma_pages << (PAGE_SHIFT - 10),
6756 #ifdef CONFIG_HIGHMEM
6757 totalhigh_pages << (PAGE_SHIFT - 10),
6759 str ? ", " : "", str ? str : "");
6763 * set_dma_reserve - set the specified number of pages reserved in the first zone
6764 * @new_dma_reserve: The number of pages to mark reserved
6766 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6767 * In the DMA zone, a significant percentage may be consumed by kernel image
6768 * and other unfreeable allocations which can skew the watermarks badly. This
6769 * function may optionally be used to account for unfreeable pages in the
6770 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6771 * smaller per-cpu batchsize.
6773 void __init set_dma_reserve(unsigned long new_dma_reserve)
6775 dma_reserve = new_dma_reserve;
6778 void __init free_area_init(unsigned long *zones_size)
6780 free_area_init_node(0, zones_size,
6781 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6784 static int page_alloc_cpu_dead(unsigned int cpu)
6787 lru_add_drain_cpu(cpu);
6791 * Spill the event counters of the dead processor
6792 * into the current processors event counters.
6793 * This artificially elevates the count of the current
6796 vm_events_fold_cpu(cpu);
6799 * Zero the differential counters of the dead processor
6800 * so that the vm statistics are consistent.
6802 * This is only okay since the processor is dead and cannot
6803 * race with what we are doing.
6805 cpu_vm_stats_fold(cpu);
6809 void __init page_alloc_init(void)
6813 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6814 "mm/page_alloc:dead", NULL,
6815 page_alloc_cpu_dead);
6820 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6821 * or min_free_kbytes changes.
6823 static void calculate_totalreserve_pages(void)
6825 struct pglist_data *pgdat;
6826 unsigned long reserve_pages = 0;
6827 enum zone_type i, j;
6829 for_each_online_pgdat(pgdat) {
6831 pgdat->totalreserve_pages = 0;
6833 for (i = 0; i < MAX_NR_ZONES; i++) {
6834 struct zone *zone = pgdat->node_zones + i;
6837 /* Find valid and maximum lowmem_reserve in the zone */
6838 for (j = i; j < MAX_NR_ZONES; j++) {
6839 if (zone->lowmem_reserve[j] > max)
6840 max = zone->lowmem_reserve[j];
6843 /* we treat the high watermark as reserved pages. */
6844 max += high_wmark_pages(zone);
6846 if (max > zone->managed_pages)
6847 max = zone->managed_pages;
6849 pgdat->totalreserve_pages += max;
6851 reserve_pages += max;
6854 totalreserve_pages = reserve_pages;
6858 * setup_per_zone_lowmem_reserve - called whenever
6859 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6860 * has a correct pages reserved value, so an adequate number of
6861 * pages are left in the zone after a successful __alloc_pages().
6863 static void setup_per_zone_lowmem_reserve(void)
6865 struct pglist_data *pgdat;
6866 enum zone_type j, idx;
6868 for_each_online_pgdat(pgdat) {
6869 for (j = 0; j < MAX_NR_ZONES; j++) {
6870 struct zone *zone = pgdat->node_zones + j;
6871 unsigned long managed_pages = zone->managed_pages;
6873 zone->lowmem_reserve[j] = 0;
6877 struct zone *lower_zone;
6881 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6882 sysctl_lowmem_reserve_ratio[idx] = 1;
6884 lower_zone = pgdat->node_zones + idx;
6885 lower_zone->lowmem_reserve[j] = managed_pages /
6886 sysctl_lowmem_reserve_ratio[idx];
6887 managed_pages += lower_zone->managed_pages;
6892 /* update totalreserve_pages */
6893 calculate_totalreserve_pages();
6896 static void __setup_per_zone_wmarks(void)
6898 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6899 unsigned long lowmem_pages = 0;
6901 unsigned long flags;
6903 /* Calculate total number of !ZONE_HIGHMEM pages */
6904 for_each_zone(zone) {
6905 if (!is_highmem(zone))
6906 lowmem_pages += zone->managed_pages;
6909 for_each_zone(zone) {
6912 spin_lock_irqsave(&zone->lock, flags);
6913 tmp = (u64)pages_min * zone->managed_pages;
6914 do_div(tmp, lowmem_pages);
6915 if (is_highmem(zone)) {
6917 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6918 * need highmem pages, so cap pages_min to a small
6921 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6922 * deltas control asynch page reclaim, and so should
6923 * not be capped for highmem.
6925 unsigned long min_pages;
6927 min_pages = zone->managed_pages / 1024;
6928 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6929 zone->watermark[WMARK_MIN] = min_pages;
6932 * If it's a lowmem zone, reserve a number of pages
6933 * proportionate to the zone's size.
6935 zone->watermark[WMARK_MIN] = tmp;
6939 * Set the kswapd watermarks distance according to the
6940 * scale factor in proportion to available memory, but
6941 * ensure a minimum size on small systems.
6943 tmp = max_t(u64, tmp >> 2,
6944 mult_frac(zone->managed_pages,
6945 watermark_scale_factor, 10000));
6947 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6948 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6950 spin_unlock_irqrestore(&zone->lock, flags);
6953 /* update totalreserve_pages */
6954 calculate_totalreserve_pages();
6958 * setup_per_zone_wmarks - called when min_free_kbytes changes
6959 * or when memory is hot-{added|removed}
6961 * Ensures that the watermark[min,low,high] values for each zone are set
6962 * correctly with respect to min_free_kbytes.
6964 void setup_per_zone_wmarks(void)
6966 static DEFINE_SPINLOCK(lock);
6969 __setup_per_zone_wmarks();
6974 * Initialise min_free_kbytes.
6976 * For small machines we want it small (128k min). For large machines
6977 * we want it large (64MB max). But it is not linear, because network
6978 * bandwidth does not increase linearly with machine size. We use
6980 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6981 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6997 int __meminit init_per_zone_wmark_min(void)
6999 unsigned long lowmem_kbytes;
7000 int new_min_free_kbytes;
7002 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7003 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7005 if (new_min_free_kbytes > user_min_free_kbytes) {
7006 min_free_kbytes = new_min_free_kbytes;
7007 if (min_free_kbytes < 128)
7008 min_free_kbytes = 128;
7009 if (min_free_kbytes > 65536)
7010 min_free_kbytes = 65536;
7012 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7013 new_min_free_kbytes, user_min_free_kbytes);
7015 setup_per_zone_wmarks();
7016 refresh_zone_stat_thresholds();
7017 setup_per_zone_lowmem_reserve();
7020 setup_min_unmapped_ratio();
7021 setup_min_slab_ratio();
7026 core_initcall(init_per_zone_wmark_min)
7029 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7030 * that we can call two helper functions whenever min_free_kbytes
7033 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7034 void __user *buffer, size_t *length, loff_t *ppos)
7038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7043 user_min_free_kbytes = min_free_kbytes;
7044 setup_per_zone_wmarks();
7049 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7050 void __user *buffer, size_t *length, loff_t *ppos)
7054 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7059 setup_per_zone_wmarks();
7065 static void setup_min_unmapped_ratio(void)
7070 for_each_online_pgdat(pgdat)
7071 pgdat->min_unmapped_pages = 0;
7074 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7075 sysctl_min_unmapped_ratio) / 100;
7079 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7080 void __user *buffer, size_t *length, loff_t *ppos)
7084 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7088 setup_min_unmapped_ratio();
7093 static void setup_min_slab_ratio(void)
7098 for_each_online_pgdat(pgdat)
7099 pgdat->min_slab_pages = 0;
7102 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7103 sysctl_min_slab_ratio) / 100;
7106 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7107 void __user *buffer, size_t *length, loff_t *ppos)
7111 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7115 setup_min_slab_ratio();
7122 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7123 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7124 * whenever sysctl_lowmem_reserve_ratio changes.
7126 * The reserve ratio obviously has absolutely no relation with the
7127 * minimum watermarks. The lowmem reserve ratio can only make sense
7128 * if in function of the boot time zone sizes.
7130 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7131 void __user *buffer, size_t *length, loff_t *ppos)
7133 proc_dointvec_minmax(table, write, buffer, length, ppos);
7134 setup_per_zone_lowmem_reserve();
7139 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7140 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7141 * pagelist can have before it gets flushed back to buddy allocator.
7143 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7144 void __user *buffer, size_t *length, loff_t *ppos)
7147 int old_percpu_pagelist_fraction;
7150 mutex_lock(&pcp_batch_high_lock);
7151 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7153 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7154 if (!write || ret < 0)
7157 /* Sanity checking to avoid pcp imbalance */
7158 if (percpu_pagelist_fraction &&
7159 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7160 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7166 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7169 for_each_populated_zone(zone) {
7172 for_each_possible_cpu(cpu)
7173 pageset_set_high_and_batch(zone,
7174 per_cpu_ptr(zone->pageset, cpu));
7177 mutex_unlock(&pcp_batch_high_lock);
7182 int hashdist = HASHDIST_DEFAULT;
7184 static int __init set_hashdist(char *str)
7188 hashdist = simple_strtoul(str, &str, 0);
7191 __setup("hashdist=", set_hashdist);
7194 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7196 * Returns the number of pages that arch has reserved but
7197 * is not known to alloc_large_system_hash().
7199 static unsigned long __init arch_reserved_kernel_pages(void)
7206 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7207 * machines. As memory size is increased the scale is also increased but at
7208 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7209 * quadruples the scale is increased by one, which means the size of hash table
7210 * only doubles, instead of quadrupling as well.
7211 * Because 32-bit systems cannot have large physical memory, where this scaling
7212 * makes sense, it is disabled on such platforms.
7214 #if __BITS_PER_LONG > 32
7215 #define ADAPT_SCALE_BASE (64ul << 30)
7216 #define ADAPT_SCALE_SHIFT 2
7217 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7221 * allocate a large system hash table from bootmem
7222 * - it is assumed that the hash table must contain an exact power-of-2
7223 * quantity of entries
7224 * - limit is the number of hash buckets, not the total allocation size
7226 void *__init alloc_large_system_hash(const char *tablename,
7227 unsigned long bucketsize,
7228 unsigned long numentries,
7231 unsigned int *_hash_shift,
7232 unsigned int *_hash_mask,
7233 unsigned long low_limit,
7234 unsigned long high_limit)
7236 unsigned long long max = high_limit;
7237 unsigned long log2qty, size;
7241 /* allow the kernel cmdline to have a say */
7243 /* round applicable memory size up to nearest megabyte */
7244 numentries = nr_kernel_pages;
7245 numentries -= arch_reserved_kernel_pages();
7247 /* It isn't necessary when PAGE_SIZE >= 1MB */
7248 if (PAGE_SHIFT < 20)
7249 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7251 #if __BITS_PER_LONG > 32
7253 unsigned long adapt;
7255 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7256 adapt <<= ADAPT_SCALE_SHIFT)
7261 /* limit to 1 bucket per 2^scale bytes of low memory */
7262 if (scale > PAGE_SHIFT)
7263 numentries >>= (scale - PAGE_SHIFT);
7265 numentries <<= (PAGE_SHIFT - scale);
7267 /* Make sure we've got at least a 0-order allocation.. */
7268 if (unlikely(flags & HASH_SMALL)) {
7269 /* Makes no sense without HASH_EARLY */
7270 WARN_ON(!(flags & HASH_EARLY));
7271 if (!(numentries >> *_hash_shift)) {
7272 numentries = 1UL << *_hash_shift;
7273 BUG_ON(!numentries);
7275 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7276 numentries = PAGE_SIZE / bucketsize;
7278 numentries = roundup_pow_of_two(numentries);
7280 /* limit allocation size to 1/16 total memory by default */
7282 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7283 do_div(max, bucketsize);
7285 max = min(max, 0x80000000ULL);
7287 if (numentries < low_limit)
7288 numentries = low_limit;
7289 if (numentries > max)
7292 log2qty = ilog2(numentries);
7295 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7296 * currently not used when HASH_EARLY is specified.
7298 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7300 size = bucketsize << log2qty;
7301 if (flags & HASH_EARLY)
7302 table = memblock_virt_alloc_nopanic(size, 0);
7304 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7307 * If bucketsize is not a power-of-two, we may free
7308 * some pages at the end of hash table which
7309 * alloc_pages_exact() automatically does
7311 if (get_order(size) < MAX_ORDER) {
7312 table = alloc_pages_exact(size, gfp_flags);
7313 kmemleak_alloc(table, size, 1, gfp_flags);
7316 } while (!table && size > PAGE_SIZE && --log2qty);
7319 panic("Failed to allocate %s hash table\n", tablename);
7321 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7322 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7325 *_hash_shift = log2qty;
7327 *_hash_mask = (1 << log2qty) - 1;
7333 * This function checks whether pageblock includes unmovable pages or not.
7334 * If @count is not zero, it is okay to include less @count unmovable pages
7336 * PageLRU check without isolation or lru_lock could race so that
7337 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7338 * check without lock_page also may miss some movable non-lru pages at
7339 * race condition. So you can't expect this function should be exact.
7341 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7343 bool skip_hwpoisoned_pages)
7345 unsigned long pfn, iter, found;
7348 * For avoiding noise data, lru_add_drain_all() should be called
7349 * If ZONE_MOVABLE, the zone never contains unmovable pages
7351 if (zone_idx(zone) == ZONE_MOVABLE)
7355 * CMA allocations (alloc_contig_range) really need to mark isolate
7356 * CMA pageblocks even when they are not movable in fact so consider
7357 * them movable here.
7359 if (is_migrate_cma(migratetype) &&
7360 is_migrate_cma(get_pageblock_migratetype(page)))
7363 pfn = page_to_pfn(page);
7364 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7365 unsigned long check = pfn + iter;
7367 if (!pfn_valid_within(check))
7370 page = pfn_to_page(check);
7372 if (PageReserved(page))
7376 * Hugepages are not in LRU lists, but they're movable.
7377 * We need not scan over tail pages bacause we don't
7378 * handle each tail page individually in migration.
7380 if (PageHuge(page)) {
7381 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7386 * We can't use page_count without pin a page
7387 * because another CPU can free compound page.
7388 * This check already skips compound tails of THP
7389 * because their page->_refcount is zero at all time.
7391 if (!page_ref_count(page)) {
7392 if (PageBuddy(page))
7393 iter += (1 << page_order(page)) - 1;
7398 * The HWPoisoned page may be not in buddy system, and
7399 * page_count() is not 0.
7401 if (skip_hwpoisoned_pages && PageHWPoison(page))
7404 if (__PageMovable(page))
7410 * If there are RECLAIMABLE pages, we need to check
7411 * it. But now, memory offline itself doesn't call
7412 * shrink_node_slabs() and it still to be fixed.
7415 * If the page is not RAM, page_count()should be 0.
7416 * we don't need more check. This is an _used_ not-movable page.
7418 * The problematic thing here is PG_reserved pages. PG_reserved
7419 * is set to both of a memory hole page and a _used_ kernel
7428 bool is_pageblock_removable_nolock(struct page *page)
7434 * We have to be careful here because we are iterating over memory
7435 * sections which are not zone aware so we might end up outside of
7436 * the zone but still within the section.
7437 * We have to take care about the node as well. If the node is offline
7438 * its NODE_DATA will be NULL - see page_zone.
7440 if (!node_online(page_to_nid(page)))
7443 zone = page_zone(page);
7444 pfn = page_to_pfn(page);
7445 if (!zone_spans_pfn(zone, pfn))
7448 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7451 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7453 static unsigned long pfn_max_align_down(unsigned long pfn)
7455 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7456 pageblock_nr_pages) - 1);
7459 static unsigned long pfn_max_align_up(unsigned long pfn)
7461 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7462 pageblock_nr_pages));
7465 /* [start, end) must belong to a single zone. */
7466 static int __alloc_contig_migrate_range(struct compact_control *cc,
7467 unsigned long start, unsigned long end)
7469 /* This function is based on compact_zone() from compaction.c. */
7470 unsigned long nr_reclaimed;
7471 unsigned long pfn = start;
7472 unsigned int tries = 0;
7477 while (pfn < end || !list_empty(&cc->migratepages)) {
7478 if (fatal_signal_pending(current)) {
7483 if (list_empty(&cc->migratepages)) {
7484 cc->nr_migratepages = 0;
7485 pfn = isolate_migratepages_range(cc, pfn, end);
7491 } else if (++tries == 5) {
7492 ret = ret < 0 ? ret : -EBUSY;
7496 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7498 cc->nr_migratepages -= nr_reclaimed;
7500 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7501 NULL, 0, cc->mode, MR_CMA);
7504 putback_movable_pages(&cc->migratepages);
7511 * alloc_contig_range() -- tries to allocate given range of pages
7512 * @start: start PFN to allocate
7513 * @end: one-past-the-last PFN to allocate
7514 * @migratetype: migratetype of the underlaying pageblocks (either
7515 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7516 * in range must have the same migratetype and it must
7517 * be either of the two.
7518 * @gfp_mask: GFP mask to use during compaction
7520 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7521 * aligned, however it's the caller's responsibility to guarantee that
7522 * we are the only thread that changes migrate type of pageblocks the
7525 * The PFN range must belong to a single zone.
7527 * Returns zero on success or negative error code. On success all
7528 * pages which PFN is in [start, end) are allocated for the caller and
7529 * need to be freed with free_contig_range().
7531 int alloc_contig_range(unsigned long start, unsigned long end,
7532 unsigned migratetype, gfp_t gfp_mask)
7534 unsigned long outer_start, outer_end;
7538 struct compact_control cc = {
7539 .nr_migratepages = 0,
7541 .zone = page_zone(pfn_to_page(start)),
7542 .mode = MIGRATE_SYNC,
7543 .ignore_skip_hint = true,
7544 .gfp_mask = current_gfp_context(gfp_mask),
7546 INIT_LIST_HEAD(&cc.migratepages);
7549 * What we do here is we mark all pageblocks in range as
7550 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7551 * have different sizes, and due to the way page allocator
7552 * work, we align the range to biggest of the two pages so
7553 * that page allocator won't try to merge buddies from
7554 * different pageblocks and change MIGRATE_ISOLATE to some
7555 * other migration type.
7557 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7558 * migrate the pages from an unaligned range (ie. pages that
7559 * we are interested in). This will put all the pages in
7560 * range back to page allocator as MIGRATE_ISOLATE.
7562 * When this is done, we take the pages in range from page
7563 * allocator removing them from the buddy system. This way
7564 * page allocator will never consider using them.
7566 * This lets us mark the pageblocks back as
7567 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7568 * aligned range but not in the unaligned, original range are
7569 * put back to page allocator so that buddy can use them.
7572 ret = start_isolate_page_range(pfn_max_align_down(start),
7573 pfn_max_align_up(end), migratetype,
7579 * In case of -EBUSY, we'd like to know which page causes problem.
7580 * So, just fall through. We will check it in test_pages_isolated().
7582 ret = __alloc_contig_migrate_range(&cc, start, end);
7583 if (ret && ret != -EBUSY)
7587 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7588 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7589 * more, all pages in [start, end) are free in page allocator.
7590 * What we are going to do is to allocate all pages from
7591 * [start, end) (that is remove them from page allocator).
7593 * The only problem is that pages at the beginning and at the
7594 * end of interesting range may be not aligned with pages that
7595 * page allocator holds, ie. they can be part of higher order
7596 * pages. Because of this, we reserve the bigger range and
7597 * once this is done free the pages we are not interested in.
7599 * We don't have to hold zone->lock here because the pages are
7600 * isolated thus they won't get removed from buddy.
7603 lru_add_drain_all();
7604 drain_all_pages(cc.zone);
7607 outer_start = start;
7608 while (!PageBuddy(pfn_to_page(outer_start))) {
7609 if (++order >= MAX_ORDER) {
7610 outer_start = start;
7613 outer_start &= ~0UL << order;
7616 if (outer_start != start) {
7617 order = page_order(pfn_to_page(outer_start));
7620 * outer_start page could be small order buddy page and
7621 * it doesn't include start page. Adjust outer_start
7622 * in this case to report failed page properly
7623 * on tracepoint in test_pages_isolated()
7625 if (outer_start + (1UL << order) <= start)
7626 outer_start = start;
7629 /* Make sure the range is really isolated. */
7630 if (test_pages_isolated(outer_start, end, false)) {
7631 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7632 __func__, outer_start, end);
7637 /* Grab isolated pages from freelists. */
7638 outer_end = isolate_freepages_range(&cc, outer_start, end);
7644 /* Free head and tail (if any) */
7645 if (start != outer_start)
7646 free_contig_range(outer_start, start - outer_start);
7647 if (end != outer_end)
7648 free_contig_range(end, outer_end - end);
7651 undo_isolate_page_range(pfn_max_align_down(start),
7652 pfn_max_align_up(end), migratetype);
7656 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7658 unsigned int count = 0;
7660 for (; nr_pages--; pfn++) {
7661 struct page *page = pfn_to_page(pfn);
7663 count += page_count(page) != 1;
7666 WARN(count != 0, "%d pages are still in use!\n", count);
7670 #ifdef CONFIG_MEMORY_HOTPLUG
7672 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7673 * page high values need to be recalulated.
7675 void __meminit zone_pcp_update(struct zone *zone)
7678 mutex_lock(&pcp_batch_high_lock);
7679 for_each_possible_cpu(cpu)
7680 pageset_set_high_and_batch(zone,
7681 per_cpu_ptr(zone->pageset, cpu));
7682 mutex_unlock(&pcp_batch_high_lock);
7686 void zone_pcp_reset(struct zone *zone)
7688 unsigned long flags;
7690 struct per_cpu_pageset *pset;
7692 /* avoid races with drain_pages() */
7693 local_irq_save(flags);
7694 if (zone->pageset != &boot_pageset) {
7695 for_each_online_cpu(cpu) {
7696 pset = per_cpu_ptr(zone->pageset, cpu);
7697 drain_zonestat(zone, pset);
7699 free_percpu(zone->pageset);
7700 zone->pageset = &boot_pageset;
7702 local_irq_restore(flags);
7705 #ifdef CONFIG_MEMORY_HOTREMOVE
7707 * All pages in the range must be in a single zone and isolated
7708 * before calling this.
7711 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7715 unsigned int order, i;
7717 unsigned long flags;
7718 /* find the first valid pfn */
7719 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7724 offline_mem_sections(pfn, end_pfn);
7725 zone = page_zone(pfn_to_page(pfn));
7726 spin_lock_irqsave(&zone->lock, flags);
7728 while (pfn < end_pfn) {
7729 if (!pfn_valid(pfn)) {
7733 page = pfn_to_page(pfn);
7735 * The HWPoisoned page may be not in buddy system, and
7736 * page_count() is not 0.
7738 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7740 SetPageReserved(page);
7744 BUG_ON(page_count(page));
7745 BUG_ON(!PageBuddy(page));
7746 order = page_order(page);
7747 #ifdef CONFIG_DEBUG_VM
7748 pr_info("remove from free list %lx %d %lx\n",
7749 pfn, 1 << order, end_pfn);
7751 list_del(&page->lru);
7752 rmv_page_order(page);
7753 zone->free_area[order].nr_free--;
7754 for (i = 0; i < (1 << order); i++)
7755 SetPageReserved((page+i));
7756 pfn += (1 << order);
7758 spin_unlock_irqrestore(&zone->lock, flags);
7762 bool is_free_buddy_page(struct page *page)
7764 struct zone *zone = page_zone(page);
7765 unsigned long pfn = page_to_pfn(page);
7766 unsigned long flags;
7769 spin_lock_irqsave(&zone->lock, flags);
7770 for (order = 0; order < MAX_ORDER; order++) {
7771 struct page *page_head = page - (pfn & ((1 << order) - 1));
7773 if (PageBuddy(page_head) && page_order(page_head) >= order)
7776 spin_unlock_irqrestore(&zone->lock, flags);
7778 return order < MAX_ORDER;