2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/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);
1015 kmemcheck_free_shadow(page, order);
1018 * Check tail pages before head page information is cleared to
1019 * avoid checking PageCompound for order-0 pages.
1021 if (unlikely(order)) {
1022 bool compound = PageCompound(page);
1025 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1028 ClearPageDoubleMap(page);
1029 for (i = 1; i < (1 << order); i++) {
1031 bad += free_tail_pages_check(page, page + i);
1032 if (unlikely(free_pages_check(page + i))) {
1036 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1039 if (PageMappingFlags(page))
1040 page->mapping = NULL;
1041 if (memcg_kmem_enabled() && PageKmemcg(page))
1042 memcg_kmem_uncharge(page, order);
1044 bad += free_pages_check(page);
1048 page_cpupid_reset_last(page);
1049 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1050 reset_page_owner(page, order);
1052 if (!PageHighMem(page)) {
1053 debug_check_no_locks_freed(page_address(page),
1054 PAGE_SIZE << order);
1055 debug_check_no_obj_freed(page_address(page),
1056 PAGE_SIZE << order);
1058 arch_free_page(page, order);
1059 kernel_poison_pages(page, 1 << order, 0);
1060 kernel_map_pages(page, 1 << order, 0);
1061 kasan_free_pages(page, order);
1066 #ifdef CONFIG_DEBUG_VM
1067 static inline bool free_pcp_prepare(struct page *page)
1069 return free_pages_prepare(page, 0, true);
1072 static inline bool bulkfree_pcp_prepare(struct page *page)
1077 static bool free_pcp_prepare(struct page *page)
1079 return free_pages_prepare(page, 0, false);
1082 static bool bulkfree_pcp_prepare(struct page *page)
1084 return free_pages_check(page);
1086 #endif /* CONFIG_DEBUG_VM */
1089 * Frees a number of pages from the PCP lists
1090 * Assumes all pages on list are in same zone, and of same order.
1091 * count is the number of pages to free.
1093 * If the zone was previously in an "all pages pinned" state then look to
1094 * see if this freeing clears that state.
1096 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1097 * pinned" detection logic.
1099 static void free_pcppages_bulk(struct zone *zone, int count,
1100 struct per_cpu_pages *pcp)
1102 int migratetype = 0;
1104 bool isolated_pageblocks;
1106 spin_lock(&zone->lock);
1107 isolated_pageblocks = has_isolate_pageblock(zone);
1111 struct list_head *list;
1114 * Remove pages from lists in a round-robin fashion. A
1115 * batch_free count is maintained that is incremented when an
1116 * empty list is encountered. This is so more pages are freed
1117 * off fuller lists instead of spinning excessively around empty
1122 if (++migratetype == MIGRATE_PCPTYPES)
1124 list = &pcp->lists[migratetype];
1125 } while (list_empty(list));
1127 /* This is the only non-empty list. Free them all. */
1128 if (batch_free == MIGRATE_PCPTYPES)
1132 int mt; /* migratetype of the to-be-freed page */
1134 page = list_last_entry(list, struct page, lru);
1135 /* must delete as __free_one_page list manipulates */
1136 list_del(&page->lru);
1138 mt = get_pcppage_migratetype(page);
1139 /* MIGRATE_ISOLATE page should not go to pcplists */
1140 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1141 /* Pageblock could have been isolated meanwhile */
1142 if (unlikely(isolated_pageblocks))
1143 mt = get_pageblock_migratetype(page);
1145 if (bulkfree_pcp_prepare(page))
1148 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1149 trace_mm_page_pcpu_drain(page, 0, mt);
1150 } while (--count && --batch_free && !list_empty(list));
1152 spin_unlock(&zone->lock);
1155 static void free_one_page(struct zone *zone,
1156 struct page *page, unsigned long pfn,
1160 spin_lock(&zone->lock);
1161 if (unlikely(has_isolate_pageblock(zone) ||
1162 is_migrate_isolate(migratetype))) {
1163 migratetype = get_pfnblock_migratetype(page, pfn);
1165 __free_one_page(page, pfn, zone, order, migratetype);
1166 spin_unlock(&zone->lock);
1169 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1170 unsigned long zone, int nid)
1172 set_page_links(page, zone, nid, pfn);
1173 init_page_count(page);
1174 page_mapcount_reset(page);
1175 page_cpupid_reset_last(page);
1177 INIT_LIST_HEAD(&page->lru);
1178 #ifdef WANT_PAGE_VIRTUAL
1179 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1180 if (!is_highmem_idx(zone))
1181 set_page_address(page, __va(pfn << PAGE_SHIFT));
1185 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1188 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1191 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1192 static void init_reserved_page(unsigned long pfn)
1197 if (!early_page_uninitialised(pfn))
1200 nid = early_pfn_to_nid(pfn);
1201 pgdat = NODE_DATA(nid);
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 struct zone *zone = &pgdat->node_zones[zid];
1206 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1209 __init_single_pfn(pfn, zid, nid);
1212 static inline void init_reserved_page(unsigned long pfn)
1215 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1218 * Initialised pages do not have PageReserved set. This function is
1219 * called for each range allocated by the bootmem allocator and
1220 * marks the pages PageReserved. The remaining valid pages are later
1221 * sent to the buddy page allocator.
1223 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1225 unsigned long start_pfn = PFN_DOWN(start);
1226 unsigned long end_pfn = PFN_UP(end);
1228 for (; start_pfn < end_pfn; start_pfn++) {
1229 if (pfn_valid(start_pfn)) {
1230 struct page *page = pfn_to_page(start_pfn);
1232 init_reserved_page(start_pfn);
1234 /* Avoid false-positive PageTail() */
1235 INIT_LIST_HEAD(&page->lru);
1237 SetPageReserved(page);
1242 static void __free_pages_ok(struct page *page, unsigned int order)
1244 unsigned long flags;
1246 unsigned long pfn = page_to_pfn(page);
1248 if (!free_pages_prepare(page, order, true))
1251 migratetype = get_pfnblock_migratetype(page, pfn);
1252 local_irq_save(flags);
1253 __count_vm_events(PGFREE, 1 << order);
1254 free_one_page(page_zone(page), page, pfn, order, migratetype);
1255 local_irq_restore(flags);
1258 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1260 unsigned int nr_pages = 1 << order;
1261 struct page *p = page;
1265 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1267 __ClearPageReserved(p);
1268 set_page_count(p, 0);
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1273 page_zone(page)->managed_pages += nr_pages;
1274 set_page_refcounted(page);
1275 __free_pages(page, order);
1278 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1279 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1281 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1283 int __meminit early_pfn_to_nid(unsigned long pfn)
1285 static DEFINE_SPINLOCK(early_pfn_lock);
1288 spin_lock(&early_pfn_lock);
1289 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1291 nid = first_online_node;
1292 spin_unlock(&early_pfn_lock);
1298 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1299 static inline bool __meminit __maybe_unused
1300 meminit_pfn_in_nid(unsigned long pfn, int node,
1301 struct mminit_pfnnid_cache *state)
1305 nid = __early_pfn_to_nid(pfn, state);
1306 if (nid >= 0 && nid != node)
1311 /* Only safe to use early in boot when initialisation is single-threaded */
1312 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1319 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1323 static inline bool __meminit __maybe_unused
1324 meminit_pfn_in_nid(unsigned long pfn, int node,
1325 struct mminit_pfnnid_cache *state)
1332 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1335 if (early_page_uninitialised(pfn))
1337 return __free_pages_boot_core(page, order);
1341 * Check that the whole (or subset of) a pageblock given by the interval of
1342 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1343 * with the migration of free compaction scanner. The scanners then need to
1344 * use only pfn_valid_within() check for arches that allow holes within
1347 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1349 * It's possible on some configurations to have a setup like node0 node1 node0
1350 * i.e. it's possible that all pages within a zones range of pages do not
1351 * belong to a single zone. We assume that a border between node0 and node1
1352 * can occur within a single pageblock, but not a node0 node1 node0
1353 * interleaving within a single pageblock. It is therefore sufficient to check
1354 * the first and last page of a pageblock and avoid checking each individual
1355 * page in a pageblock.
1357 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1358 unsigned long end_pfn, struct zone *zone)
1360 struct page *start_page;
1361 struct page *end_page;
1363 /* end_pfn is one past the range we are checking */
1366 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1369 start_page = pfn_to_online_page(start_pfn);
1373 if (page_zone(start_page) != zone)
1376 end_page = pfn_to_page(end_pfn);
1378 /* This gives a shorter code than deriving page_zone(end_page) */
1379 if (page_zone_id(start_page) != page_zone_id(end_page))
1385 void set_zone_contiguous(struct zone *zone)
1387 unsigned long block_start_pfn = zone->zone_start_pfn;
1388 unsigned long block_end_pfn;
1390 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1391 for (; block_start_pfn < zone_end_pfn(zone);
1392 block_start_pfn = block_end_pfn,
1393 block_end_pfn += pageblock_nr_pages) {
1395 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1397 if (!__pageblock_pfn_to_page(block_start_pfn,
1398 block_end_pfn, zone))
1402 /* We confirm that there is no hole */
1403 zone->contiguous = true;
1406 void clear_zone_contiguous(struct zone *zone)
1408 zone->contiguous = false;
1411 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1412 static void __init deferred_free_range(struct page *page,
1413 unsigned long pfn, int nr_pages)
1420 /* Free a large naturally-aligned chunk if possible */
1421 if (nr_pages == pageblock_nr_pages &&
1422 (pfn & (pageblock_nr_pages - 1)) == 0) {
1423 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1424 __free_pages_boot_core(page, pageblock_order);
1428 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1429 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1430 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1431 __free_pages_boot_core(page, 0);
1435 /* Completion tracking for deferred_init_memmap() threads */
1436 static atomic_t pgdat_init_n_undone __initdata;
1437 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1439 static inline void __init pgdat_init_report_one_done(void)
1441 if (atomic_dec_and_test(&pgdat_init_n_undone))
1442 complete(&pgdat_init_all_done_comp);
1445 /* Initialise remaining memory on a node */
1446 static int __init deferred_init_memmap(void *data)
1448 pg_data_t *pgdat = data;
1449 int nid = pgdat->node_id;
1450 struct mminit_pfnnid_cache nid_init_state = { };
1451 unsigned long start = jiffies;
1452 unsigned long nr_pages = 0;
1453 unsigned long walk_start, walk_end;
1456 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1457 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1459 if (first_init_pfn == ULONG_MAX) {
1460 pgdat_init_report_one_done();
1464 /* Bind memory initialisation thread to a local node if possible */
1465 if (!cpumask_empty(cpumask))
1466 set_cpus_allowed_ptr(current, cpumask);
1468 /* Sanity check boundaries */
1469 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1470 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1471 pgdat->first_deferred_pfn = ULONG_MAX;
1473 /* Only the highest zone is deferred so find it */
1474 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1475 zone = pgdat->node_zones + zid;
1476 if (first_init_pfn < zone_end_pfn(zone))
1480 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1481 unsigned long pfn, end_pfn;
1482 struct page *page = NULL;
1483 struct page *free_base_page = NULL;
1484 unsigned long free_base_pfn = 0;
1487 end_pfn = min(walk_end, zone_end_pfn(zone));
1488 pfn = first_init_pfn;
1489 if (pfn < walk_start)
1491 if (pfn < zone->zone_start_pfn)
1492 pfn = zone->zone_start_pfn;
1494 for (; pfn < end_pfn; pfn++) {
1495 if (!pfn_valid_within(pfn))
1499 * Ensure pfn_valid is checked every
1500 * pageblock_nr_pages for memory holes
1502 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1503 if (!pfn_valid(pfn)) {
1509 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1514 /* Minimise pfn page lookups and scheduler checks */
1515 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1518 nr_pages += nr_to_free;
1519 deferred_free_range(free_base_page,
1520 free_base_pfn, nr_to_free);
1521 free_base_page = NULL;
1522 free_base_pfn = nr_to_free = 0;
1524 page = pfn_to_page(pfn);
1529 VM_BUG_ON(page_zone(page) != zone);
1533 __init_single_page(page, pfn, zid, nid);
1534 if (!free_base_page) {
1535 free_base_page = page;
1536 free_base_pfn = pfn;
1541 /* Where possible, batch up pages for a single free */
1544 /* Free the current block of pages to allocator */
1545 nr_pages += nr_to_free;
1546 deferred_free_range(free_base_page, free_base_pfn,
1548 free_base_page = NULL;
1549 free_base_pfn = nr_to_free = 0;
1551 /* Free the last block of pages to allocator */
1552 nr_pages += nr_to_free;
1553 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1555 first_init_pfn = max(end_pfn, first_init_pfn);
1558 /* Sanity check that the next zone really is unpopulated */
1559 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1561 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1562 jiffies_to_msecs(jiffies - start));
1564 pgdat_init_report_one_done();
1567 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1569 void __init page_alloc_init_late(void)
1573 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1576 /* There will be num_node_state(N_MEMORY) threads */
1577 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1578 for_each_node_state(nid, N_MEMORY) {
1579 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1582 /* Block until all are initialised */
1583 wait_for_completion(&pgdat_init_all_done_comp);
1585 /* Reinit limits that are based on free pages after the kernel is up */
1586 files_maxfiles_init();
1588 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1589 /* Discard memblock private memory */
1593 for_each_populated_zone(zone)
1594 set_zone_contiguous(zone);
1598 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1599 void __init init_cma_reserved_pageblock(struct page *page)
1601 unsigned i = pageblock_nr_pages;
1602 struct page *p = page;
1605 __ClearPageReserved(p);
1606 set_page_count(p, 0);
1609 set_pageblock_migratetype(page, MIGRATE_CMA);
1611 if (pageblock_order >= MAX_ORDER) {
1612 i = pageblock_nr_pages;
1615 set_page_refcounted(p);
1616 __free_pages(p, MAX_ORDER - 1);
1617 p += MAX_ORDER_NR_PAGES;
1618 } while (i -= MAX_ORDER_NR_PAGES);
1620 set_page_refcounted(page);
1621 __free_pages(page, pageblock_order);
1624 adjust_managed_page_count(page, pageblock_nr_pages);
1629 * The order of subdivision here is critical for the IO subsystem.
1630 * Please do not alter this order without good reasons and regression
1631 * testing. Specifically, as large blocks of memory are subdivided,
1632 * the order in which smaller blocks are delivered depends on the order
1633 * they're subdivided in this function. This is the primary factor
1634 * influencing the order in which pages are delivered to the IO
1635 * subsystem according to empirical testing, and this is also justified
1636 * by considering the behavior of a buddy system containing a single
1637 * large block of memory acted on by a series of small allocations.
1638 * This behavior is a critical factor in sglist merging's success.
1642 static inline void expand(struct zone *zone, struct page *page,
1643 int low, int high, struct free_area *area,
1646 unsigned long size = 1 << high;
1648 while (high > low) {
1652 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1655 * Mark as guard pages (or page), that will allow to
1656 * merge back to allocator when buddy will be freed.
1657 * Corresponding page table entries will not be touched,
1658 * pages will stay not present in virtual address space
1660 if (set_page_guard(zone, &page[size], high, migratetype))
1663 list_add(&page[size].lru, &area->free_list[migratetype]);
1665 set_page_order(&page[size], high);
1669 static void check_new_page_bad(struct page *page)
1671 const char *bad_reason = NULL;
1672 unsigned long bad_flags = 0;
1674 if (unlikely(atomic_read(&page->_mapcount) != -1))
1675 bad_reason = "nonzero mapcount";
1676 if (unlikely(page->mapping != NULL))
1677 bad_reason = "non-NULL mapping";
1678 if (unlikely(page_ref_count(page) != 0))
1679 bad_reason = "nonzero _count";
1680 if (unlikely(page->flags & __PG_HWPOISON)) {
1681 bad_reason = "HWPoisoned (hardware-corrupted)";
1682 bad_flags = __PG_HWPOISON;
1683 /* Don't complain about hwpoisoned pages */
1684 page_mapcount_reset(page); /* remove PageBuddy */
1687 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1688 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1689 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1692 if (unlikely(page->mem_cgroup))
1693 bad_reason = "page still charged to cgroup";
1695 bad_page(page, bad_reason, bad_flags);
1699 * This page is about to be returned from the page allocator
1701 static inline int check_new_page(struct page *page)
1703 if (likely(page_expected_state(page,
1704 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1707 check_new_page_bad(page);
1711 static inline bool free_pages_prezeroed(void)
1713 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1714 page_poisoning_enabled();
1717 #ifdef CONFIG_DEBUG_VM
1718 static bool check_pcp_refill(struct page *page)
1723 static bool check_new_pcp(struct page *page)
1725 return check_new_page(page);
1728 static bool check_pcp_refill(struct page *page)
1730 return check_new_page(page);
1732 static bool check_new_pcp(struct page *page)
1736 #endif /* CONFIG_DEBUG_VM */
1738 static bool check_new_pages(struct page *page, unsigned int order)
1741 for (i = 0; i < (1 << order); i++) {
1742 struct page *p = page + i;
1744 if (unlikely(check_new_page(p)))
1751 inline void post_alloc_hook(struct page *page, unsigned int order,
1754 set_page_private(page, 0);
1755 set_page_refcounted(page);
1757 arch_alloc_page(page, order);
1758 kernel_map_pages(page, 1 << order, 1);
1759 kernel_poison_pages(page, 1 << order, 1);
1760 kasan_alloc_pages(page, order);
1761 set_page_owner(page, order, gfp_flags);
1764 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1765 unsigned int alloc_flags)
1769 post_alloc_hook(page, order, gfp_flags);
1771 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1772 for (i = 0; i < (1 << order); i++)
1773 clear_highpage(page + i);
1775 if (order && (gfp_flags & __GFP_COMP))
1776 prep_compound_page(page, order);
1779 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1780 * allocate the page. The expectation is that the caller is taking
1781 * steps that will free more memory. The caller should avoid the page
1782 * being used for !PFMEMALLOC purposes.
1784 if (alloc_flags & ALLOC_NO_WATERMARKS)
1785 set_page_pfmemalloc(page);
1787 clear_page_pfmemalloc(page);
1791 * Go through the free lists for the given migratetype and remove
1792 * the smallest available page from the freelists
1795 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1798 unsigned int current_order;
1799 struct free_area *area;
1802 /* Find a page of the appropriate size in the preferred list */
1803 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1804 area = &(zone->free_area[current_order]);
1805 page = list_first_entry_or_null(&area->free_list[migratetype],
1809 list_del(&page->lru);
1810 rmv_page_order(page);
1812 expand(zone, page, order, current_order, area, migratetype);
1813 set_pcppage_migratetype(page, migratetype);
1822 * This array describes the order lists are fallen back to when
1823 * the free lists for the desirable migrate type are depleted
1825 static int fallbacks[MIGRATE_TYPES][4] = {
1826 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1827 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1828 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1830 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1832 #ifdef CONFIG_MEMORY_ISOLATION
1833 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1838 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1841 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1844 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1845 unsigned int order) { return NULL; }
1849 * Move the free pages in a range to the free lists of the requested type.
1850 * Note that start_page and end_pages are not aligned on a pageblock
1851 * boundary. If alignment is required, use move_freepages_block()
1853 static int move_freepages(struct zone *zone,
1854 struct page *start_page, struct page *end_page,
1855 int migratetype, int *num_movable)
1859 int pages_moved = 0;
1861 #ifndef CONFIG_HOLES_IN_ZONE
1863 * page_zone is not safe to call in this context when
1864 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1865 * anyway as we check zone boundaries in move_freepages_block().
1866 * Remove at a later date when no bug reports exist related to
1867 * grouping pages by mobility
1869 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1875 for (page = start_page; page <= end_page;) {
1876 if (!pfn_valid_within(page_to_pfn(page))) {
1881 /* Make sure we are not inadvertently changing nodes */
1882 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1884 if (!PageBuddy(page)) {
1886 * We assume that pages that could be isolated for
1887 * migration are movable. But we don't actually try
1888 * isolating, as that would be expensive.
1891 (PageLRU(page) || __PageMovable(page)))
1898 order = page_order(page);
1899 list_move(&page->lru,
1900 &zone->free_area[order].free_list[migratetype]);
1902 pages_moved += 1 << order;
1908 int move_freepages_block(struct zone *zone, struct page *page,
1909 int migratetype, int *num_movable)
1911 unsigned long start_pfn, end_pfn;
1912 struct page *start_page, *end_page;
1914 start_pfn = page_to_pfn(page);
1915 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1916 start_page = pfn_to_page(start_pfn);
1917 end_page = start_page + pageblock_nr_pages - 1;
1918 end_pfn = start_pfn + pageblock_nr_pages - 1;
1920 /* Do not cross zone boundaries */
1921 if (!zone_spans_pfn(zone, start_pfn))
1923 if (!zone_spans_pfn(zone, end_pfn))
1926 return move_freepages(zone, start_page, end_page, migratetype,
1930 static void change_pageblock_range(struct page *pageblock_page,
1931 int start_order, int migratetype)
1933 int nr_pageblocks = 1 << (start_order - pageblock_order);
1935 while (nr_pageblocks--) {
1936 set_pageblock_migratetype(pageblock_page, migratetype);
1937 pageblock_page += pageblock_nr_pages;
1942 * When we are falling back to another migratetype during allocation, try to
1943 * steal extra free pages from the same pageblocks to satisfy further
1944 * allocations, instead of polluting multiple pageblocks.
1946 * If we are stealing a relatively large buddy page, it is likely there will
1947 * be more free pages in the pageblock, so try to steal them all. For
1948 * reclaimable and unmovable allocations, we steal regardless of page size,
1949 * as fragmentation caused by those allocations polluting movable pageblocks
1950 * is worse than movable allocations stealing from unmovable and reclaimable
1953 static bool can_steal_fallback(unsigned int order, int start_mt)
1956 * Leaving this order check is intended, although there is
1957 * relaxed order check in next check. The reason is that
1958 * we can actually steal whole pageblock if this condition met,
1959 * but, below check doesn't guarantee it and that is just heuristic
1960 * so could be changed anytime.
1962 if (order >= pageblock_order)
1965 if (order >= pageblock_order / 2 ||
1966 start_mt == MIGRATE_RECLAIMABLE ||
1967 start_mt == MIGRATE_UNMOVABLE ||
1968 page_group_by_mobility_disabled)
1975 * This function implements actual steal behaviour. If order is large enough,
1976 * we can steal whole pageblock. If not, we first move freepages in this
1977 * pageblock to our migratetype and determine how many already-allocated pages
1978 * are there in the pageblock with a compatible migratetype. If at least half
1979 * of pages are free or compatible, we can change migratetype of the pageblock
1980 * itself, so pages freed in the future will be put on the correct free list.
1982 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1983 int start_type, bool whole_block)
1985 unsigned int current_order = page_order(page);
1986 struct free_area *area;
1987 int free_pages, movable_pages, alike_pages;
1990 old_block_type = get_pageblock_migratetype(page);
1993 * This can happen due to races and we want to prevent broken
1994 * highatomic accounting.
1996 if (is_migrate_highatomic(old_block_type))
1999 /* Take ownership for orders >= pageblock_order */
2000 if (current_order >= pageblock_order) {
2001 change_pageblock_range(page, current_order, start_type);
2005 /* We are not allowed to try stealing from the whole block */
2009 free_pages = move_freepages_block(zone, page, start_type,
2012 * Determine how many pages are compatible with our allocation.
2013 * For movable allocation, it's the number of movable pages which
2014 * we just obtained. For other types it's a bit more tricky.
2016 if (start_type == MIGRATE_MOVABLE) {
2017 alike_pages = movable_pages;
2020 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2021 * to MOVABLE pageblock, consider all non-movable pages as
2022 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2023 * vice versa, be conservative since we can't distinguish the
2024 * exact migratetype of non-movable pages.
2026 if (old_block_type == MIGRATE_MOVABLE)
2027 alike_pages = pageblock_nr_pages
2028 - (free_pages + movable_pages);
2033 /* moving whole block can fail due to zone boundary conditions */
2038 * If a sufficient number of pages in the block are either free or of
2039 * comparable migratability as our allocation, claim the whole block.
2041 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2042 page_group_by_mobility_disabled)
2043 set_pageblock_migratetype(page, start_type);
2048 area = &zone->free_area[current_order];
2049 list_move(&page->lru, &area->free_list[start_type]);
2053 * Check whether there is a suitable fallback freepage with requested order.
2054 * If only_stealable is true, this function returns fallback_mt only if
2055 * we can steal other freepages all together. This would help to reduce
2056 * fragmentation due to mixed migratetype pages in one pageblock.
2058 int find_suitable_fallback(struct free_area *area, unsigned int order,
2059 int migratetype, bool only_stealable, bool *can_steal)
2064 if (area->nr_free == 0)
2069 fallback_mt = fallbacks[migratetype][i];
2070 if (fallback_mt == MIGRATE_TYPES)
2073 if (list_empty(&area->free_list[fallback_mt]))
2076 if (can_steal_fallback(order, migratetype))
2079 if (!only_stealable)
2090 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2091 * there are no empty page blocks that contain a page with a suitable order
2093 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2094 unsigned int alloc_order)
2097 unsigned long max_managed, flags;
2100 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2101 * Check is race-prone but harmless.
2103 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2104 if (zone->nr_reserved_highatomic >= max_managed)
2107 spin_lock_irqsave(&zone->lock, flags);
2109 /* Recheck the nr_reserved_highatomic limit under the lock */
2110 if (zone->nr_reserved_highatomic >= max_managed)
2114 mt = get_pageblock_migratetype(page);
2115 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2116 && !is_migrate_cma(mt)) {
2117 zone->nr_reserved_highatomic += pageblock_nr_pages;
2118 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2119 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2123 spin_unlock_irqrestore(&zone->lock, flags);
2127 * Used when an allocation is about to fail under memory pressure. This
2128 * potentially hurts the reliability of high-order allocations when under
2129 * intense memory pressure but failed atomic allocations should be easier
2130 * to recover from than an OOM.
2132 * If @force is true, try to unreserve a pageblock even though highatomic
2133 * pageblock is exhausted.
2135 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2138 struct zonelist *zonelist = ac->zonelist;
2139 unsigned long flags;
2146 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2149 * Preserve at least one pageblock unless memory pressure
2152 if (!force && zone->nr_reserved_highatomic <=
2156 spin_lock_irqsave(&zone->lock, flags);
2157 for (order = 0; order < MAX_ORDER; order++) {
2158 struct free_area *area = &(zone->free_area[order]);
2160 page = list_first_entry_or_null(
2161 &area->free_list[MIGRATE_HIGHATOMIC],
2167 * In page freeing path, migratetype change is racy so
2168 * we can counter several free pages in a pageblock
2169 * in this loop althoug we changed the pageblock type
2170 * from highatomic to ac->migratetype. So we should
2171 * adjust the count once.
2173 if (is_migrate_highatomic_page(page)) {
2175 * It should never happen but changes to
2176 * locking could inadvertently allow a per-cpu
2177 * drain to add pages to MIGRATE_HIGHATOMIC
2178 * while unreserving so be safe and watch for
2181 zone->nr_reserved_highatomic -= min(
2183 zone->nr_reserved_highatomic);
2187 * Convert to ac->migratetype and avoid the normal
2188 * pageblock stealing heuristics. Minimally, the caller
2189 * is doing the work and needs the pages. More
2190 * importantly, if the block was always converted to
2191 * MIGRATE_UNMOVABLE or another type then the number
2192 * of pageblocks that cannot be completely freed
2195 set_pageblock_migratetype(page, ac->migratetype);
2196 ret = move_freepages_block(zone, page, ac->migratetype,
2199 spin_unlock_irqrestore(&zone->lock, flags);
2203 spin_unlock_irqrestore(&zone->lock, flags);
2210 * Try finding a free buddy page on the fallback list and put it on the free
2211 * list of requested migratetype, possibly along with other pages from the same
2212 * block, depending on fragmentation avoidance heuristics. Returns true if
2213 * fallback was found so that __rmqueue_smallest() can grab it.
2215 * The use of signed ints for order and current_order is a deliberate
2216 * deviation from the rest of this file, to make the for loop
2217 * condition simpler.
2220 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2222 struct free_area *area;
2229 * Find the largest available free page in the other list. This roughly
2230 * approximates finding the pageblock with the most free pages, which
2231 * would be too costly to do exactly.
2233 for (current_order = MAX_ORDER - 1; current_order >= order;
2235 area = &(zone->free_area[current_order]);
2236 fallback_mt = find_suitable_fallback(area, current_order,
2237 start_migratetype, false, &can_steal);
2238 if (fallback_mt == -1)
2242 * We cannot steal all free pages from the pageblock and the
2243 * requested migratetype is movable. In that case it's better to
2244 * steal and split the smallest available page instead of the
2245 * largest available page, because even if the next movable
2246 * allocation falls back into a different pageblock than this
2247 * one, it won't cause permanent fragmentation.
2249 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2250 && current_order > order)
2259 for (current_order = order; current_order < MAX_ORDER;
2261 area = &(zone->free_area[current_order]);
2262 fallback_mt = find_suitable_fallback(area, current_order,
2263 start_migratetype, false, &can_steal);
2264 if (fallback_mt != -1)
2269 * This should not happen - we already found a suitable fallback
2270 * when looking for the largest page.
2272 VM_BUG_ON(current_order == MAX_ORDER);
2275 page = list_first_entry(&area->free_list[fallback_mt],
2278 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2280 trace_mm_page_alloc_extfrag(page, order, current_order,
2281 start_migratetype, fallback_mt);
2288 * Do the hard work of removing an element from the buddy allocator.
2289 * Call me with the zone->lock already held.
2291 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2297 page = __rmqueue_smallest(zone, order, migratetype);
2298 if (unlikely(!page)) {
2299 if (migratetype == MIGRATE_MOVABLE)
2300 page = __rmqueue_cma_fallback(zone, order);
2302 if (!page && __rmqueue_fallback(zone, order, migratetype))
2306 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2311 * Obtain a specified number of elements from the buddy allocator, all under
2312 * a single hold of the lock, for efficiency. Add them to the supplied list.
2313 * Returns the number of new pages which were placed at *list.
2315 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2316 unsigned long count, struct list_head *list,
2317 int migratetype, bool cold)
2321 spin_lock(&zone->lock);
2322 for (i = 0; i < count; ++i) {
2323 struct page *page = __rmqueue(zone, order, migratetype);
2324 if (unlikely(page == NULL))
2327 if (unlikely(check_pcp_refill(page)))
2331 * Split buddy pages returned by expand() are received here
2332 * in physical page order. The page is added to the callers and
2333 * list and the list head then moves forward. From the callers
2334 * perspective, the linked list is ordered by page number in
2335 * some conditions. This is useful for IO devices that can
2336 * merge IO requests if the physical pages are ordered
2340 list_add(&page->lru, list);
2342 list_add_tail(&page->lru, list);
2345 if (is_migrate_cma(get_pcppage_migratetype(page)))
2346 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2351 * i pages were removed from the buddy list even if some leak due
2352 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2353 * on i. Do not confuse with 'alloced' which is the number of
2354 * pages added to the pcp list.
2356 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2357 spin_unlock(&zone->lock);
2363 * Called from the vmstat counter updater to drain pagesets of this
2364 * currently executing processor on remote nodes after they have
2367 * Note that this function must be called with the thread pinned to
2368 * a single processor.
2370 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2372 unsigned long flags;
2373 int to_drain, batch;
2375 local_irq_save(flags);
2376 batch = READ_ONCE(pcp->batch);
2377 to_drain = min(pcp->count, batch);
2379 free_pcppages_bulk(zone, to_drain, pcp);
2380 pcp->count -= to_drain;
2382 local_irq_restore(flags);
2387 * Drain pcplists of the indicated processor and zone.
2389 * The processor must either be the current processor and the
2390 * thread pinned to the current processor or a processor that
2393 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2395 unsigned long flags;
2396 struct per_cpu_pageset *pset;
2397 struct per_cpu_pages *pcp;
2399 local_irq_save(flags);
2400 pset = per_cpu_ptr(zone->pageset, cpu);
2404 free_pcppages_bulk(zone, pcp->count, pcp);
2407 local_irq_restore(flags);
2411 * Drain pcplists of all zones on the indicated processor.
2413 * The processor must either be the current processor and the
2414 * thread pinned to the current processor or a processor that
2417 static void drain_pages(unsigned int cpu)
2421 for_each_populated_zone(zone) {
2422 drain_pages_zone(cpu, zone);
2427 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2429 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2430 * the single zone's pages.
2432 void drain_local_pages(struct zone *zone)
2434 int cpu = smp_processor_id();
2437 drain_pages_zone(cpu, zone);
2442 static void drain_local_pages_wq(struct work_struct *work)
2445 * drain_all_pages doesn't use proper cpu hotplug protection so
2446 * we can race with cpu offline when the WQ can move this from
2447 * a cpu pinned worker to an unbound one. We can operate on a different
2448 * cpu which is allright but we also have to make sure to not move to
2452 drain_local_pages(NULL);
2457 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2459 * When zone parameter is non-NULL, spill just the single zone's pages.
2461 * Note that this can be extremely slow as the draining happens in a workqueue.
2463 void drain_all_pages(struct zone *zone)
2468 * Allocate in the BSS so we wont require allocation in
2469 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2471 static cpumask_t cpus_with_pcps;
2474 * Make sure nobody triggers this path before mm_percpu_wq is fully
2477 if (WARN_ON_ONCE(!mm_percpu_wq))
2480 /* Workqueues cannot recurse */
2481 if (current->flags & PF_WQ_WORKER)
2485 * Do not drain if one is already in progress unless it's specific to
2486 * a zone. Such callers are primarily CMA and memory hotplug and need
2487 * the drain to be complete when the call returns.
2489 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2492 mutex_lock(&pcpu_drain_mutex);
2496 * We don't care about racing with CPU hotplug event
2497 * as offline notification will cause the notified
2498 * cpu to drain that CPU pcps and on_each_cpu_mask
2499 * disables preemption as part of its processing
2501 for_each_online_cpu(cpu) {
2502 struct per_cpu_pageset *pcp;
2504 bool has_pcps = false;
2507 pcp = per_cpu_ptr(zone->pageset, cpu);
2511 for_each_populated_zone(z) {
2512 pcp = per_cpu_ptr(z->pageset, cpu);
2513 if (pcp->pcp.count) {
2521 cpumask_set_cpu(cpu, &cpus_with_pcps);
2523 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2526 for_each_cpu(cpu, &cpus_with_pcps) {
2527 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2528 INIT_WORK(work, drain_local_pages_wq);
2529 queue_work_on(cpu, mm_percpu_wq, work);
2531 for_each_cpu(cpu, &cpus_with_pcps)
2532 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2534 mutex_unlock(&pcpu_drain_mutex);
2537 #ifdef CONFIG_HIBERNATION
2540 * Touch the watchdog for every WD_PAGE_COUNT pages.
2542 #define WD_PAGE_COUNT (128*1024)
2544 void mark_free_pages(struct zone *zone)
2546 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2547 unsigned long flags;
2548 unsigned int order, t;
2551 if (zone_is_empty(zone))
2554 spin_lock_irqsave(&zone->lock, flags);
2556 max_zone_pfn = zone_end_pfn(zone);
2557 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2558 if (pfn_valid(pfn)) {
2559 page = pfn_to_page(pfn);
2561 if (!--page_count) {
2562 touch_nmi_watchdog();
2563 page_count = WD_PAGE_COUNT;
2566 if (page_zone(page) != zone)
2569 if (!swsusp_page_is_forbidden(page))
2570 swsusp_unset_page_free(page);
2573 for_each_migratetype_order(order, t) {
2574 list_for_each_entry(page,
2575 &zone->free_area[order].free_list[t], lru) {
2578 pfn = page_to_pfn(page);
2579 for (i = 0; i < (1UL << order); i++) {
2580 if (!--page_count) {
2581 touch_nmi_watchdog();
2582 page_count = WD_PAGE_COUNT;
2584 swsusp_set_page_free(pfn_to_page(pfn + i));
2588 spin_unlock_irqrestore(&zone->lock, flags);
2590 #endif /* CONFIG_PM */
2593 * Free a 0-order page
2594 * cold == true ? free a cold page : free a hot page
2596 void free_hot_cold_page(struct page *page, bool cold)
2598 struct zone *zone = page_zone(page);
2599 struct per_cpu_pages *pcp;
2600 unsigned long flags;
2601 unsigned long pfn = page_to_pfn(page);
2604 if (!free_pcp_prepare(page))
2607 migratetype = get_pfnblock_migratetype(page, pfn);
2608 set_pcppage_migratetype(page, migratetype);
2609 local_irq_save(flags);
2610 __count_vm_event(PGFREE);
2613 * We only track unmovable, reclaimable and movable on pcp lists.
2614 * Free ISOLATE pages back to the allocator because they are being
2615 * offlined but treat HIGHATOMIC as movable pages so we can get those
2616 * areas back if necessary. Otherwise, we may have to free
2617 * excessively into the page allocator
2619 if (migratetype >= MIGRATE_PCPTYPES) {
2620 if (unlikely(is_migrate_isolate(migratetype))) {
2621 free_one_page(zone, page, pfn, 0, migratetype);
2624 migratetype = MIGRATE_MOVABLE;
2627 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2629 list_add(&page->lru, &pcp->lists[migratetype]);
2631 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2633 if (pcp->count >= pcp->high) {
2634 unsigned long batch = READ_ONCE(pcp->batch);
2635 free_pcppages_bulk(zone, batch, pcp);
2636 pcp->count -= batch;
2640 local_irq_restore(flags);
2644 * Free a list of 0-order pages
2646 void free_hot_cold_page_list(struct list_head *list, bool cold)
2648 struct page *page, *next;
2650 list_for_each_entry_safe(page, next, list, lru) {
2651 trace_mm_page_free_batched(page, cold);
2652 free_hot_cold_page(page, cold);
2657 * split_page takes a non-compound higher-order page, and splits it into
2658 * n (1<<order) sub-pages: page[0..n]
2659 * Each sub-page must be freed individually.
2661 * Note: this is probably too low level an operation for use in drivers.
2662 * Please consult with lkml before using this in your driver.
2664 void split_page(struct page *page, unsigned int order)
2668 VM_BUG_ON_PAGE(PageCompound(page), page);
2669 VM_BUG_ON_PAGE(!page_count(page), page);
2671 #ifdef CONFIG_KMEMCHECK
2673 * Split shadow pages too, because free(page[0]) would
2674 * otherwise free the whole shadow.
2676 if (kmemcheck_page_is_tracked(page))
2677 split_page(virt_to_page(page[0].shadow), order);
2680 for (i = 1; i < (1 << order); i++)
2681 set_page_refcounted(page + i);
2682 split_page_owner(page, order);
2684 EXPORT_SYMBOL_GPL(split_page);
2686 int __isolate_free_page(struct page *page, unsigned int order)
2688 unsigned long watermark;
2692 BUG_ON(!PageBuddy(page));
2694 zone = page_zone(page);
2695 mt = get_pageblock_migratetype(page);
2697 if (!is_migrate_isolate(mt)) {
2699 * Obey watermarks as if the page was being allocated. We can
2700 * emulate a high-order watermark check with a raised order-0
2701 * watermark, because we already know our high-order page
2704 watermark = min_wmark_pages(zone) + (1UL << order);
2705 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2708 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2711 /* Remove page from free list */
2712 list_del(&page->lru);
2713 zone->free_area[order].nr_free--;
2714 rmv_page_order(page);
2717 * Set the pageblock if the isolated page is at least half of a
2720 if (order >= pageblock_order - 1) {
2721 struct page *endpage = page + (1 << order) - 1;
2722 for (; page < endpage; page += pageblock_nr_pages) {
2723 int mt = get_pageblock_migratetype(page);
2724 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2725 && !is_migrate_highatomic(mt))
2726 set_pageblock_migratetype(page,
2732 return 1UL << order;
2736 * Update NUMA hit/miss statistics
2738 * Must be called with interrupts disabled.
2740 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2743 enum zone_stat_item local_stat = NUMA_LOCAL;
2745 if (z->node != numa_node_id())
2746 local_stat = NUMA_OTHER;
2748 if (z->node == preferred_zone->node)
2749 __inc_zone_state(z, NUMA_HIT);
2751 __inc_zone_state(z, NUMA_MISS);
2752 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2754 __inc_zone_state(z, local_stat);
2758 /* Remove page from the per-cpu list, caller must protect the list */
2759 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2760 bool cold, struct per_cpu_pages *pcp,
2761 struct list_head *list)
2766 if (list_empty(list)) {
2767 pcp->count += rmqueue_bulk(zone, 0,
2770 if (unlikely(list_empty(list)))
2775 page = list_last_entry(list, struct page, lru);
2777 page = list_first_entry(list, struct page, lru);
2779 list_del(&page->lru);
2781 } while (check_new_pcp(page));
2786 /* Lock and remove page from the per-cpu list */
2787 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2788 struct zone *zone, unsigned int order,
2789 gfp_t gfp_flags, int migratetype)
2791 struct per_cpu_pages *pcp;
2792 struct list_head *list;
2793 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2795 unsigned long flags;
2797 local_irq_save(flags);
2798 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2799 list = &pcp->lists[migratetype];
2800 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2802 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2803 zone_statistics(preferred_zone, zone);
2805 local_irq_restore(flags);
2810 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2813 struct page *rmqueue(struct zone *preferred_zone,
2814 struct zone *zone, unsigned int order,
2815 gfp_t gfp_flags, unsigned int alloc_flags,
2818 unsigned long flags;
2821 if (likely(order == 0)) {
2822 page = rmqueue_pcplist(preferred_zone, zone, order,
2823 gfp_flags, migratetype);
2828 * We most definitely don't want callers attempting to
2829 * allocate greater than order-1 page units with __GFP_NOFAIL.
2831 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2832 spin_lock_irqsave(&zone->lock, flags);
2836 if (alloc_flags & ALLOC_HARDER) {
2837 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2839 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2842 page = __rmqueue(zone, order, migratetype);
2843 } while (page && check_new_pages(page, order));
2844 spin_unlock(&zone->lock);
2847 __mod_zone_freepage_state(zone, -(1 << order),
2848 get_pcppage_migratetype(page));
2850 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2851 zone_statistics(preferred_zone, zone);
2852 local_irq_restore(flags);
2855 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2859 local_irq_restore(flags);
2863 #ifdef CONFIG_FAIL_PAGE_ALLOC
2866 struct fault_attr attr;
2868 bool ignore_gfp_highmem;
2869 bool ignore_gfp_reclaim;
2871 } fail_page_alloc = {
2872 .attr = FAULT_ATTR_INITIALIZER,
2873 .ignore_gfp_reclaim = true,
2874 .ignore_gfp_highmem = true,
2878 static int __init setup_fail_page_alloc(char *str)
2880 return setup_fault_attr(&fail_page_alloc.attr, str);
2882 __setup("fail_page_alloc=", setup_fail_page_alloc);
2884 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2886 if (order < fail_page_alloc.min_order)
2888 if (gfp_mask & __GFP_NOFAIL)
2890 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2892 if (fail_page_alloc.ignore_gfp_reclaim &&
2893 (gfp_mask & __GFP_DIRECT_RECLAIM))
2896 return should_fail(&fail_page_alloc.attr, 1 << order);
2899 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2901 static int __init fail_page_alloc_debugfs(void)
2903 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2906 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2907 &fail_page_alloc.attr);
2909 return PTR_ERR(dir);
2911 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2912 &fail_page_alloc.ignore_gfp_reclaim))
2914 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2915 &fail_page_alloc.ignore_gfp_highmem))
2917 if (!debugfs_create_u32("min-order", mode, dir,
2918 &fail_page_alloc.min_order))
2923 debugfs_remove_recursive(dir);
2928 late_initcall(fail_page_alloc_debugfs);
2930 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2932 #else /* CONFIG_FAIL_PAGE_ALLOC */
2934 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2939 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2942 * Return true if free base pages are above 'mark'. For high-order checks it
2943 * will return true of the order-0 watermark is reached and there is at least
2944 * one free page of a suitable size. Checking now avoids taking the zone lock
2945 * to check in the allocation paths if no pages are free.
2947 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2948 int classzone_idx, unsigned int alloc_flags,
2953 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2955 /* free_pages may go negative - that's OK */
2956 free_pages -= (1 << order) - 1;
2958 if (alloc_flags & ALLOC_HIGH)
2962 * If the caller does not have rights to ALLOC_HARDER then subtract
2963 * the high-atomic reserves. This will over-estimate the size of the
2964 * atomic reserve but it avoids a search.
2966 if (likely(!alloc_harder))
2967 free_pages -= z->nr_reserved_highatomic;
2972 /* If allocation can't use CMA areas don't use free CMA pages */
2973 if (!(alloc_flags & ALLOC_CMA))
2974 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2978 * Check watermarks for an order-0 allocation request. If these
2979 * are not met, then a high-order request also cannot go ahead
2980 * even if a suitable page happened to be free.
2982 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2985 /* If this is an order-0 request then the watermark is fine */
2989 /* For a high-order request, check at least one suitable page is free */
2990 for (o = order; o < MAX_ORDER; o++) {
2991 struct free_area *area = &z->free_area[o];
3000 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3001 if (!list_empty(&area->free_list[mt]))
3006 if ((alloc_flags & ALLOC_CMA) &&
3007 !list_empty(&area->free_list[MIGRATE_CMA])) {
3015 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3016 int classzone_idx, unsigned int alloc_flags)
3018 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3019 zone_page_state(z, NR_FREE_PAGES));
3022 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3023 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3025 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3029 /* If allocation can't use CMA areas don't use free CMA pages */
3030 if (!(alloc_flags & ALLOC_CMA))
3031 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3035 * Fast check for order-0 only. If this fails then the reserves
3036 * need to be calculated. There is a corner case where the check
3037 * passes but only the high-order atomic reserve are free. If
3038 * the caller is !atomic then it'll uselessly search the free
3039 * list. That corner case is then slower but it is harmless.
3041 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3044 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3048 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3049 unsigned long mark, int classzone_idx)
3051 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3053 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3054 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3056 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3061 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3063 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3066 #else /* CONFIG_NUMA */
3067 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3071 #endif /* CONFIG_NUMA */
3074 * get_page_from_freelist goes through the zonelist trying to allocate
3077 static struct page *
3078 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3079 const struct alloc_context *ac)
3081 struct zoneref *z = ac->preferred_zoneref;
3083 struct pglist_data *last_pgdat_dirty_limit = NULL;
3086 * Scan zonelist, looking for a zone with enough free.
3087 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3089 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3094 if (cpusets_enabled() &&
3095 (alloc_flags & ALLOC_CPUSET) &&
3096 !__cpuset_zone_allowed(zone, gfp_mask))
3099 * When allocating a page cache page for writing, we
3100 * want to get it from a node that is within its dirty
3101 * limit, such that no single node holds more than its
3102 * proportional share of globally allowed dirty pages.
3103 * The dirty limits take into account the node's
3104 * lowmem reserves and high watermark so that kswapd
3105 * should be able to balance it without having to
3106 * write pages from its LRU list.
3108 * XXX: For now, allow allocations to potentially
3109 * exceed the per-node dirty limit in the slowpath
3110 * (spread_dirty_pages unset) before going into reclaim,
3111 * which is important when on a NUMA setup the allowed
3112 * nodes are together not big enough to reach the
3113 * global limit. The proper fix for these situations
3114 * will require awareness of nodes in the
3115 * dirty-throttling and the flusher threads.
3117 if (ac->spread_dirty_pages) {
3118 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3121 if (!node_dirty_ok(zone->zone_pgdat)) {
3122 last_pgdat_dirty_limit = zone->zone_pgdat;
3127 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3128 if (!zone_watermark_fast(zone, order, mark,
3129 ac_classzone_idx(ac), alloc_flags)) {
3132 /* Checked here to keep the fast path fast */
3133 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3134 if (alloc_flags & ALLOC_NO_WATERMARKS)
3137 if (node_reclaim_mode == 0 ||
3138 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3141 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3143 case NODE_RECLAIM_NOSCAN:
3146 case NODE_RECLAIM_FULL:
3147 /* scanned but unreclaimable */
3150 /* did we reclaim enough */
3151 if (zone_watermark_ok(zone, order, mark,
3152 ac_classzone_idx(ac), alloc_flags))
3160 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3161 gfp_mask, alloc_flags, ac->migratetype);
3163 prep_new_page(page, order, gfp_mask, alloc_flags);
3166 * If this is a high-order atomic allocation then check
3167 * if the pageblock should be reserved for the future
3169 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3170 reserve_highatomic_pageblock(page, zone, order);
3180 * Large machines with many possible nodes should not always dump per-node
3181 * meminfo in irq context.
3183 static inline bool should_suppress_show_mem(void)
3188 ret = in_interrupt();
3193 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3195 unsigned int filter = SHOW_MEM_FILTER_NODES;
3196 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3198 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3202 * This documents exceptions given to allocations in certain
3203 * contexts that are allowed to allocate outside current's set
3206 if (!(gfp_mask & __GFP_NOMEMALLOC))
3207 if (test_thread_flag(TIF_MEMDIE) ||
3208 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3209 filter &= ~SHOW_MEM_FILTER_NODES;
3210 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3211 filter &= ~SHOW_MEM_FILTER_NODES;
3213 show_mem(filter, nodemask);
3216 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3218 struct va_format vaf;
3220 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3221 DEFAULT_RATELIMIT_BURST);
3223 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3226 pr_warn("%s: ", current->comm);
3228 va_start(args, fmt);
3231 pr_cont("%pV", &vaf);
3234 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3236 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3238 pr_cont("(null)\n");
3240 cpuset_print_current_mems_allowed();
3243 warn_alloc_show_mem(gfp_mask, nodemask);
3246 static inline struct page *
3247 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3248 unsigned int alloc_flags,
3249 const struct alloc_context *ac)
3253 page = get_page_from_freelist(gfp_mask, order,
3254 alloc_flags|ALLOC_CPUSET, ac);
3256 * fallback to ignore cpuset restriction if our nodes
3260 page = get_page_from_freelist(gfp_mask, order,
3266 static inline struct page *
3267 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3268 const struct alloc_context *ac, unsigned long *did_some_progress)
3270 struct oom_control oc = {
3271 .zonelist = ac->zonelist,
3272 .nodemask = ac->nodemask,
3274 .gfp_mask = gfp_mask,
3279 *did_some_progress = 0;
3282 * Acquire the oom lock. If that fails, somebody else is
3283 * making progress for us.
3285 if (!mutex_trylock(&oom_lock)) {
3286 *did_some_progress = 1;
3287 schedule_timeout_uninterruptible(1);
3292 * Go through the zonelist yet one more time, keep very high watermark
3293 * here, this is only to catch a parallel oom killing, we must fail if
3294 * we're still under heavy pressure. But make sure that this reclaim
3295 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3296 * allocation which will never fail due to oom_lock already held.
3298 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3299 ~__GFP_DIRECT_RECLAIM, order,
3300 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3304 /* Coredumps can quickly deplete all memory reserves */
3305 if (current->flags & PF_DUMPCORE)
3307 /* The OOM killer will not help higher order allocs */
3308 if (order > PAGE_ALLOC_COSTLY_ORDER)
3311 * We have already exhausted all our reclaim opportunities without any
3312 * success so it is time to admit defeat. We will skip the OOM killer
3313 * because it is very likely that the caller has a more reasonable
3314 * fallback than shooting a random task.
3316 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3318 /* The OOM killer does not needlessly kill tasks for lowmem */
3319 if (ac->high_zoneidx < ZONE_NORMAL)
3321 if (pm_suspended_storage())
3324 * XXX: GFP_NOFS allocations should rather fail than rely on
3325 * other request to make a forward progress.
3326 * We are in an unfortunate situation where out_of_memory cannot
3327 * do much for this context but let's try it to at least get
3328 * access to memory reserved if the current task is killed (see
3329 * out_of_memory). Once filesystems are ready to handle allocation
3330 * failures more gracefully we should just bail out here.
3333 /* The OOM killer may not free memory on a specific node */
3334 if (gfp_mask & __GFP_THISNODE)
3337 /* Exhausted what can be done so it's blamo time */
3338 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3339 *did_some_progress = 1;
3342 * Help non-failing allocations by giving them access to memory
3345 if (gfp_mask & __GFP_NOFAIL)
3346 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3347 ALLOC_NO_WATERMARKS, ac);
3350 mutex_unlock(&oom_lock);
3355 * Maximum number of compaction retries wit a progress before OOM
3356 * killer is consider as the only way to move forward.
3358 #define MAX_COMPACT_RETRIES 16
3360 #ifdef CONFIG_COMPACTION
3361 /* Try memory compaction for high-order allocations before reclaim */
3362 static struct page *
3363 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3364 unsigned int alloc_flags, const struct alloc_context *ac,
3365 enum compact_priority prio, enum compact_result *compact_result)
3368 unsigned int noreclaim_flag;
3373 noreclaim_flag = memalloc_noreclaim_save();
3374 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3376 memalloc_noreclaim_restore(noreclaim_flag);
3378 if (*compact_result <= COMPACT_INACTIVE)
3382 * At least in one zone compaction wasn't deferred or skipped, so let's
3383 * count a compaction stall
3385 count_vm_event(COMPACTSTALL);
3387 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3390 struct zone *zone = page_zone(page);
3392 zone->compact_blockskip_flush = false;
3393 compaction_defer_reset(zone, order, true);
3394 count_vm_event(COMPACTSUCCESS);
3399 * It's bad if compaction run occurs and fails. The most likely reason
3400 * is that pages exist, but not enough to satisfy watermarks.
3402 count_vm_event(COMPACTFAIL);
3410 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3411 enum compact_result compact_result,
3412 enum compact_priority *compact_priority,
3413 int *compaction_retries)
3415 int max_retries = MAX_COMPACT_RETRIES;
3418 int retries = *compaction_retries;
3419 enum compact_priority priority = *compact_priority;
3424 if (compaction_made_progress(compact_result))
3425 (*compaction_retries)++;
3428 * compaction considers all the zone as desperately out of memory
3429 * so it doesn't really make much sense to retry except when the
3430 * failure could be caused by insufficient priority
3432 if (compaction_failed(compact_result))
3433 goto check_priority;
3436 * make sure the compaction wasn't deferred or didn't bail out early
3437 * due to locks contention before we declare that we should give up.
3438 * But do not retry if the given zonelist is not suitable for
3441 if (compaction_withdrawn(compact_result)) {
3442 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3447 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3448 * costly ones because they are de facto nofail and invoke OOM
3449 * killer to move on while costly can fail and users are ready
3450 * to cope with that. 1/4 retries is rather arbitrary but we
3451 * would need much more detailed feedback from compaction to
3452 * make a better decision.
3454 if (order > PAGE_ALLOC_COSTLY_ORDER)
3456 if (*compaction_retries <= max_retries) {
3462 * Make sure there are attempts at the highest priority if we exhausted
3463 * all retries or failed at the lower priorities.
3466 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3467 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3469 if (*compact_priority > min_priority) {
3470 (*compact_priority)--;
3471 *compaction_retries = 0;
3475 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3479 static inline struct page *
3480 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3481 unsigned int alloc_flags, const struct alloc_context *ac,
3482 enum compact_priority prio, enum compact_result *compact_result)
3484 *compact_result = COMPACT_SKIPPED;
3489 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3490 enum compact_result compact_result,
3491 enum compact_priority *compact_priority,
3492 int *compaction_retries)
3497 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3501 * There are setups with compaction disabled which would prefer to loop
3502 * inside the allocator rather than hit the oom killer prematurely.
3503 * Let's give them a good hope and keep retrying while the order-0
3504 * watermarks are OK.
3506 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3508 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3509 ac_classzone_idx(ac), alloc_flags))
3514 #endif /* CONFIG_COMPACTION */
3516 /* Perform direct synchronous page reclaim */
3518 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3519 const struct alloc_context *ac)
3521 struct reclaim_state reclaim_state;
3523 unsigned int noreclaim_flag;
3527 /* We now go into synchronous reclaim */
3528 cpuset_memory_pressure_bump();
3529 noreclaim_flag = memalloc_noreclaim_save();
3530 lockdep_set_current_reclaim_state(gfp_mask);
3531 reclaim_state.reclaimed_slab = 0;
3532 current->reclaim_state = &reclaim_state;
3534 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3537 current->reclaim_state = NULL;
3538 lockdep_clear_current_reclaim_state();
3539 memalloc_noreclaim_restore(noreclaim_flag);
3546 /* The really slow allocator path where we enter direct reclaim */
3547 static inline struct page *
3548 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3549 unsigned int alloc_flags, const struct alloc_context *ac,
3550 unsigned long *did_some_progress)
3552 struct page *page = NULL;
3553 bool drained = false;
3555 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3556 if (unlikely(!(*did_some_progress)))
3560 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3563 * If an allocation failed after direct reclaim, it could be because
3564 * pages are pinned on the per-cpu lists or in high alloc reserves.
3565 * Shrink them them and try again
3567 if (!page && !drained) {
3568 unreserve_highatomic_pageblock(ac, false);
3569 drain_all_pages(NULL);
3577 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3581 pg_data_t *last_pgdat = NULL;
3583 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3584 ac->high_zoneidx, ac->nodemask) {
3585 if (last_pgdat != zone->zone_pgdat)
3586 wakeup_kswapd(zone, order, ac->high_zoneidx);
3587 last_pgdat = zone->zone_pgdat;
3591 static inline unsigned int
3592 gfp_to_alloc_flags(gfp_t gfp_mask)
3594 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3596 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3597 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3600 * The caller may dip into page reserves a bit more if the caller
3601 * cannot run direct reclaim, or if the caller has realtime scheduling
3602 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3603 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3605 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3607 if (gfp_mask & __GFP_ATOMIC) {
3609 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3610 * if it can't schedule.
3612 if (!(gfp_mask & __GFP_NOMEMALLOC))
3613 alloc_flags |= ALLOC_HARDER;
3615 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3616 * comment for __cpuset_node_allowed().
3618 alloc_flags &= ~ALLOC_CPUSET;
3619 } else if (unlikely(rt_task(current)) && !in_interrupt())
3620 alloc_flags |= ALLOC_HARDER;
3623 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3624 alloc_flags |= ALLOC_CMA;
3629 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3631 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3634 if (gfp_mask & __GFP_MEMALLOC)
3636 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3638 if (!in_interrupt() &&
3639 ((current->flags & PF_MEMALLOC) ||
3640 unlikely(test_thread_flag(TIF_MEMDIE))))
3647 * Checks whether it makes sense to retry the reclaim to make a forward progress
3648 * for the given allocation request.
3650 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3651 * without success, or when we couldn't even meet the watermark if we
3652 * reclaimed all remaining pages on the LRU lists.
3654 * Returns true if a retry is viable or false to enter the oom path.
3657 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3658 struct alloc_context *ac, int alloc_flags,
3659 bool did_some_progress, int *no_progress_loops)
3665 * Costly allocations might have made a progress but this doesn't mean
3666 * their order will become available due to high fragmentation so
3667 * always increment the no progress counter for them
3669 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3670 *no_progress_loops = 0;
3672 (*no_progress_loops)++;
3675 * Make sure we converge to OOM if we cannot make any progress
3676 * several times in the row.
3678 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3679 /* Before OOM, exhaust highatomic_reserve */
3680 return unreserve_highatomic_pageblock(ac, true);
3684 * Keep reclaiming pages while there is a chance this will lead
3685 * somewhere. If none of the target zones can satisfy our allocation
3686 * request even if all reclaimable pages are considered then we are
3687 * screwed and have to go OOM.
3689 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3691 unsigned long available;
3692 unsigned long reclaimable;
3693 unsigned long min_wmark = min_wmark_pages(zone);
3696 available = reclaimable = zone_reclaimable_pages(zone);
3697 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3700 * Would the allocation succeed if we reclaimed all
3701 * reclaimable pages?
3703 wmark = __zone_watermark_ok(zone, order, min_wmark,
3704 ac_classzone_idx(ac), alloc_flags, available);
3705 trace_reclaim_retry_zone(z, order, reclaimable,
3706 available, min_wmark, *no_progress_loops, wmark);
3709 * If we didn't make any progress and have a lot of
3710 * dirty + writeback pages then we should wait for
3711 * an IO to complete to slow down the reclaim and
3712 * prevent from pre mature OOM
3714 if (!did_some_progress) {
3715 unsigned long write_pending;
3717 write_pending = zone_page_state_snapshot(zone,
3718 NR_ZONE_WRITE_PENDING);
3720 if (2 * write_pending > reclaimable) {
3721 congestion_wait(BLK_RW_ASYNC, HZ/10);
3727 * Memory allocation/reclaim might be called from a WQ
3728 * context and the current implementation of the WQ
3729 * concurrency control doesn't recognize that
3730 * a particular WQ is congested if the worker thread is
3731 * looping without ever sleeping. Therefore we have to
3732 * do a short sleep here rather than calling
3735 if (current->flags & PF_WQ_WORKER)
3736 schedule_timeout_uninterruptible(1);
3748 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3751 * It's possible that cpuset's mems_allowed and the nodemask from
3752 * mempolicy don't intersect. This should be normally dealt with by
3753 * policy_nodemask(), but it's possible to race with cpuset update in
3754 * such a way the check therein was true, and then it became false
3755 * before we got our cpuset_mems_cookie here.
3756 * This assumes that for all allocations, ac->nodemask can come only
3757 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3758 * when it does not intersect with the cpuset restrictions) or the
3759 * caller can deal with a violated nodemask.
3761 if (cpusets_enabled() && ac->nodemask &&
3762 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3763 ac->nodemask = NULL;
3768 * When updating a task's mems_allowed or mempolicy nodemask, it is
3769 * possible to race with parallel threads in such a way that our
3770 * allocation can fail while the mask is being updated. If we are about
3771 * to fail, check if the cpuset changed during allocation and if so,
3774 if (read_mems_allowed_retry(cpuset_mems_cookie))
3780 static inline struct page *
3781 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3782 struct alloc_context *ac)
3784 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3785 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3786 struct page *page = NULL;
3787 unsigned int alloc_flags;
3788 unsigned long did_some_progress;
3789 enum compact_priority compact_priority;
3790 enum compact_result compact_result;
3791 int compaction_retries;
3792 int no_progress_loops;
3793 unsigned long alloc_start = jiffies;
3794 unsigned int stall_timeout = 10 * HZ;
3795 unsigned int cpuset_mems_cookie;
3798 * In the slowpath, we sanity check order to avoid ever trying to
3799 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3800 * be using allocators in order of preference for an area that is
3803 if (order >= MAX_ORDER) {
3804 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3809 * We also sanity check to catch abuse of atomic reserves being used by
3810 * callers that are not in atomic context.
3812 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3813 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3814 gfp_mask &= ~__GFP_ATOMIC;
3817 compaction_retries = 0;
3818 no_progress_loops = 0;
3819 compact_priority = DEF_COMPACT_PRIORITY;
3820 cpuset_mems_cookie = read_mems_allowed_begin();
3823 * The fast path uses conservative alloc_flags to succeed only until
3824 * kswapd needs to be woken up, and to avoid the cost of setting up
3825 * alloc_flags precisely. So we do that now.
3827 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3830 * We need to recalculate the starting point for the zonelist iterator
3831 * because we might have used different nodemask in the fast path, or
3832 * there was a cpuset modification and we are retrying - otherwise we
3833 * could end up iterating over non-eligible zones endlessly.
3835 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3836 ac->high_zoneidx, ac->nodemask);
3837 if (!ac->preferred_zoneref->zone)
3840 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3841 wake_all_kswapds(order, ac);
3844 * The adjusted alloc_flags might result in immediate success, so try
3847 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3852 * For costly allocations, try direct compaction first, as it's likely
3853 * that we have enough base pages and don't need to reclaim. For non-
3854 * movable high-order allocations, do that as well, as compaction will
3855 * try prevent permanent fragmentation by migrating from blocks of the
3857 * Don't try this for allocations that are allowed to ignore
3858 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3860 if (can_direct_reclaim &&
3862 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3863 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3864 page = __alloc_pages_direct_compact(gfp_mask, order,
3866 INIT_COMPACT_PRIORITY,
3872 * Checks for costly allocations with __GFP_NORETRY, which
3873 * includes THP page fault allocations
3875 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3877 * If compaction is deferred for high-order allocations,
3878 * it is because sync compaction recently failed. If
3879 * this is the case and the caller requested a THP
3880 * allocation, we do not want to heavily disrupt the
3881 * system, so we fail the allocation instead of entering
3884 if (compact_result == COMPACT_DEFERRED)
3888 * Looks like reclaim/compaction is worth trying, but
3889 * sync compaction could be very expensive, so keep
3890 * using async compaction.
3892 compact_priority = INIT_COMPACT_PRIORITY;
3897 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3898 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3899 wake_all_kswapds(order, ac);
3901 if (gfp_pfmemalloc_allowed(gfp_mask))
3902 alloc_flags = ALLOC_NO_WATERMARKS;
3905 * Reset the zonelist iterators if memory policies can be ignored.
3906 * These allocations are high priority and system rather than user
3909 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3910 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3911 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3912 ac->high_zoneidx, ac->nodemask);
3915 /* Attempt with potentially adjusted zonelist and alloc_flags */
3916 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3920 /* Caller is not willing to reclaim, we can't balance anything */
3921 if (!can_direct_reclaim)
3924 /* Make sure we know about allocations which stall for too long */
3925 if (time_after(jiffies, alloc_start + stall_timeout)) {
3926 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
3927 "page allocation stalls for %ums, order:%u",
3928 jiffies_to_msecs(jiffies-alloc_start), order);
3929 stall_timeout += 10 * HZ;
3932 /* Avoid recursion of direct reclaim */
3933 if (current->flags & PF_MEMALLOC)
3936 /* Try direct reclaim and then allocating */
3937 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3938 &did_some_progress);
3942 /* Try direct compaction and then allocating */
3943 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3944 compact_priority, &compact_result);
3948 /* Do not loop if specifically requested */
3949 if (gfp_mask & __GFP_NORETRY)
3953 * Do not retry costly high order allocations unless they are
3954 * __GFP_RETRY_MAYFAIL
3956 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
3959 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3960 did_some_progress > 0, &no_progress_loops))
3964 * It doesn't make any sense to retry for the compaction if the order-0
3965 * reclaim is not able to make any progress because the current
3966 * implementation of the compaction depends on the sufficient amount
3967 * of free memory (see __compaction_suitable)
3969 if (did_some_progress > 0 &&
3970 should_compact_retry(ac, order, alloc_flags,
3971 compact_result, &compact_priority,
3972 &compaction_retries))
3976 /* Deal with possible cpuset update races before we start OOM killing */
3977 if (check_retry_cpuset(cpuset_mems_cookie, ac))
3980 /* Reclaim has failed us, start killing things */
3981 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3985 /* Avoid allocations with no watermarks from looping endlessly */
3986 if (test_thread_flag(TIF_MEMDIE) &&
3987 (alloc_flags == ALLOC_NO_WATERMARKS ||
3988 (gfp_mask & __GFP_NOMEMALLOC)))
3991 /* Retry as long as the OOM killer is making progress */
3992 if (did_some_progress) {
3993 no_progress_loops = 0;
3998 /* Deal with possible cpuset update races before we fail */
3999 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4003 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4006 if (gfp_mask & __GFP_NOFAIL) {
4008 * All existing users of the __GFP_NOFAIL are blockable, so warn
4009 * of any new users that actually require GFP_NOWAIT
4011 if (WARN_ON_ONCE(!can_direct_reclaim))
4015 * PF_MEMALLOC request from this context is rather bizarre
4016 * because we cannot reclaim anything and only can loop waiting
4017 * for somebody to do a work for us
4019 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4022 * non failing costly orders are a hard requirement which we
4023 * are not prepared for much so let's warn about these users
4024 * so that we can identify them and convert them to something
4027 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4030 * Help non-failing allocations by giving them access to memory
4031 * reserves but do not use ALLOC_NO_WATERMARKS because this
4032 * could deplete whole memory reserves which would just make
4033 * the situation worse
4035 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4043 warn_alloc(gfp_mask, ac->nodemask,
4044 "page allocation failure: order:%u", order);
4049 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4050 int preferred_nid, nodemask_t *nodemask,
4051 struct alloc_context *ac, gfp_t *alloc_mask,
4052 unsigned int *alloc_flags)
4054 ac->high_zoneidx = gfp_zone(gfp_mask);
4055 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4056 ac->nodemask = nodemask;
4057 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4059 if (cpusets_enabled()) {
4060 *alloc_mask |= __GFP_HARDWALL;
4062 ac->nodemask = &cpuset_current_mems_allowed;
4064 *alloc_flags |= ALLOC_CPUSET;
4067 lockdep_trace_alloc(gfp_mask);
4069 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4071 if (should_fail_alloc_page(gfp_mask, order))
4074 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4075 *alloc_flags |= ALLOC_CMA;
4080 /* Determine whether to spread dirty pages and what the first usable zone */
4081 static inline void finalise_ac(gfp_t gfp_mask,
4082 unsigned int order, struct alloc_context *ac)
4084 /* Dirty zone balancing only done in the fast path */
4085 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4088 * The preferred zone is used for statistics but crucially it is
4089 * also used as the starting point for the zonelist iterator. It
4090 * may get reset for allocations that ignore memory policies.
4092 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4093 ac->high_zoneidx, ac->nodemask);
4097 * This is the 'heart' of the zoned buddy allocator.
4100 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4101 nodemask_t *nodemask)
4104 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4105 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
4106 struct alloc_context ac = { };
4108 gfp_mask &= gfp_allowed_mask;
4109 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4112 finalise_ac(gfp_mask, order, &ac);
4114 /* First allocation attempt */
4115 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4120 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4121 * resp. GFP_NOIO which has to be inherited for all allocation requests
4122 * from a particular context which has been marked by
4123 * memalloc_no{fs,io}_{save,restore}.
4125 alloc_mask = current_gfp_context(gfp_mask);
4126 ac.spread_dirty_pages = false;
4129 * Restore the original nodemask if it was potentially replaced with
4130 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4132 if (unlikely(ac.nodemask != nodemask))
4133 ac.nodemask = nodemask;
4135 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4138 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4139 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4140 __free_pages(page, order);
4144 if (kmemcheck_enabled && page)
4145 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4147 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4151 EXPORT_SYMBOL(__alloc_pages_nodemask);
4154 * Common helper functions.
4156 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4161 * __get_free_pages() returns a 32-bit address, which cannot represent
4164 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4166 page = alloc_pages(gfp_mask, order);
4169 return (unsigned long) page_address(page);
4171 EXPORT_SYMBOL(__get_free_pages);
4173 unsigned long get_zeroed_page(gfp_t gfp_mask)
4175 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4177 EXPORT_SYMBOL(get_zeroed_page);
4179 void __free_pages(struct page *page, unsigned int order)
4181 if (put_page_testzero(page)) {
4183 free_hot_cold_page(page, false);
4185 __free_pages_ok(page, order);
4189 EXPORT_SYMBOL(__free_pages);
4191 void free_pages(unsigned long addr, unsigned int order)
4194 VM_BUG_ON(!virt_addr_valid((void *)addr));
4195 __free_pages(virt_to_page((void *)addr), order);
4199 EXPORT_SYMBOL(free_pages);
4203 * An arbitrary-length arbitrary-offset area of memory which resides
4204 * within a 0 or higher order page. Multiple fragments within that page
4205 * are individually refcounted, in the page's reference counter.
4207 * The page_frag functions below provide a simple allocation framework for
4208 * page fragments. This is used by the network stack and network device
4209 * drivers to provide a backing region of memory for use as either an
4210 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4212 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4215 struct page *page = NULL;
4216 gfp_t gfp = gfp_mask;
4218 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4219 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4221 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4222 PAGE_FRAG_CACHE_MAX_ORDER);
4223 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4225 if (unlikely(!page))
4226 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4228 nc->va = page ? page_address(page) : NULL;
4233 void __page_frag_cache_drain(struct page *page, unsigned int count)
4235 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4237 if (page_ref_sub_and_test(page, count)) {
4238 unsigned int order = compound_order(page);
4241 free_hot_cold_page(page, false);
4243 __free_pages_ok(page, order);
4246 EXPORT_SYMBOL(__page_frag_cache_drain);
4248 void *page_frag_alloc(struct page_frag_cache *nc,
4249 unsigned int fragsz, gfp_t gfp_mask)
4251 unsigned int size = PAGE_SIZE;
4255 if (unlikely(!nc->va)) {
4257 page = __page_frag_cache_refill(nc, gfp_mask);
4261 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4262 /* if size can vary use size else just use PAGE_SIZE */
4265 /* Even if we own the page, we do not use atomic_set().
4266 * This would break get_page_unless_zero() users.
4268 page_ref_add(page, size - 1);
4270 /* reset page count bias and offset to start of new frag */
4271 nc->pfmemalloc = page_is_pfmemalloc(page);
4272 nc->pagecnt_bias = size;
4276 offset = nc->offset - fragsz;
4277 if (unlikely(offset < 0)) {
4278 page = virt_to_page(nc->va);
4280 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4283 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4284 /* if size can vary use size else just use PAGE_SIZE */
4287 /* OK, page count is 0, we can safely set it */
4288 set_page_count(page, size);
4290 /* reset page count bias and offset to start of new frag */
4291 nc->pagecnt_bias = size;
4292 offset = size - fragsz;
4296 nc->offset = offset;
4298 return nc->va + offset;
4300 EXPORT_SYMBOL(page_frag_alloc);
4303 * Frees a page fragment allocated out of either a compound or order 0 page.
4305 void page_frag_free(void *addr)
4307 struct page *page = virt_to_head_page(addr);
4309 if (unlikely(put_page_testzero(page)))
4310 __free_pages_ok(page, compound_order(page));
4312 EXPORT_SYMBOL(page_frag_free);
4314 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4318 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4319 unsigned long used = addr + PAGE_ALIGN(size);
4321 split_page(virt_to_page((void *)addr), order);
4322 while (used < alloc_end) {
4327 return (void *)addr;
4331 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4332 * @size: the number of bytes to allocate
4333 * @gfp_mask: GFP flags for the allocation
4335 * This function is similar to alloc_pages(), except that it allocates the
4336 * minimum number of pages to satisfy the request. alloc_pages() can only
4337 * allocate memory in power-of-two pages.
4339 * This function is also limited by MAX_ORDER.
4341 * Memory allocated by this function must be released by free_pages_exact().
4343 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4345 unsigned int order = get_order(size);
4348 addr = __get_free_pages(gfp_mask, order);
4349 return make_alloc_exact(addr, order, size);
4351 EXPORT_SYMBOL(alloc_pages_exact);
4354 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4356 * @nid: the preferred node ID where memory should be allocated
4357 * @size: the number of bytes to allocate
4358 * @gfp_mask: GFP flags for the allocation
4360 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4363 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4365 unsigned int order = get_order(size);
4366 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4369 return make_alloc_exact((unsigned long)page_address(p), order, size);
4373 * free_pages_exact - release memory allocated via alloc_pages_exact()
4374 * @virt: the value returned by alloc_pages_exact.
4375 * @size: size of allocation, same value as passed to alloc_pages_exact().
4377 * Release the memory allocated by a previous call to alloc_pages_exact.
4379 void free_pages_exact(void *virt, size_t size)
4381 unsigned long addr = (unsigned long)virt;
4382 unsigned long end = addr + PAGE_ALIGN(size);
4384 while (addr < end) {
4389 EXPORT_SYMBOL(free_pages_exact);
4392 * nr_free_zone_pages - count number of pages beyond high watermark
4393 * @offset: The zone index of the highest zone
4395 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4396 * high watermark within all zones at or below a given zone index. For each
4397 * zone, the number of pages is calculated as:
4399 * nr_free_zone_pages = managed_pages - high_pages
4401 static unsigned long nr_free_zone_pages(int offset)
4406 /* Just pick one node, since fallback list is circular */
4407 unsigned long sum = 0;
4409 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4411 for_each_zone_zonelist(zone, z, zonelist, offset) {
4412 unsigned long size = zone->managed_pages;
4413 unsigned long high = high_wmark_pages(zone);
4422 * nr_free_buffer_pages - count number of pages beyond high watermark
4424 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4425 * watermark within ZONE_DMA and ZONE_NORMAL.
4427 unsigned long nr_free_buffer_pages(void)
4429 return nr_free_zone_pages(gfp_zone(GFP_USER));
4431 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4434 * nr_free_pagecache_pages - count number of pages beyond high watermark
4436 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4437 * high watermark within all zones.
4439 unsigned long nr_free_pagecache_pages(void)
4441 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4444 static inline void show_node(struct zone *zone)
4446 if (IS_ENABLED(CONFIG_NUMA))
4447 printk("Node %d ", zone_to_nid(zone));
4450 long si_mem_available(void)
4453 unsigned long pagecache;
4454 unsigned long wmark_low = 0;
4455 unsigned long pages[NR_LRU_LISTS];
4459 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4460 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4463 wmark_low += zone->watermark[WMARK_LOW];
4466 * Estimate the amount of memory available for userspace allocations,
4467 * without causing swapping.
4469 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4472 * Not all the page cache can be freed, otherwise the system will
4473 * start swapping. Assume at least half of the page cache, or the
4474 * low watermark worth of cache, needs to stay.
4476 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4477 pagecache -= min(pagecache / 2, wmark_low);
4478 available += pagecache;
4481 * Part of the reclaimable slab consists of items that are in use,
4482 * and cannot be freed. Cap this estimate at the low watermark.
4484 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4485 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4492 EXPORT_SYMBOL_GPL(si_mem_available);
4494 void si_meminfo(struct sysinfo *val)
4496 val->totalram = totalram_pages;
4497 val->sharedram = global_node_page_state(NR_SHMEM);
4498 val->freeram = global_page_state(NR_FREE_PAGES);
4499 val->bufferram = nr_blockdev_pages();
4500 val->totalhigh = totalhigh_pages;
4501 val->freehigh = nr_free_highpages();
4502 val->mem_unit = PAGE_SIZE;
4505 EXPORT_SYMBOL(si_meminfo);
4508 void si_meminfo_node(struct sysinfo *val, int nid)
4510 int zone_type; /* needs to be signed */
4511 unsigned long managed_pages = 0;
4512 unsigned long managed_highpages = 0;
4513 unsigned long free_highpages = 0;
4514 pg_data_t *pgdat = NODE_DATA(nid);
4516 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4517 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4518 val->totalram = managed_pages;
4519 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4520 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4521 #ifdef CONFIG_HIGHMEM
4522 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4523 struct zone *zone = &pgdat->node_zones[zone_type];
4525 if (is_highmem(zone)) {
4526 managed_highpages += zone->managed_pages;
4527 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4530 val->totalhigh = managed_highpages;
4531 val->freehigh = free_highpages;
4533 val->totalhigh = managed_highpages;
4534 val->freehigh = free_highpages;
4536 val->mem_unit = PAGE_SIZE;
4541 * Determine whether the node should be displayed or not, depending on whether
4542 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4544 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4546 if (!(flags & SHOW_MEM_FILTER_NODES))
4550 * no node mask - aka implicit memory numa policy. Do not bother with
4551 * the synchronization - read_mems_allowed_begin - because we do not
4552 * have to be precise here.
4555 nodemask = &cpuset_current_mems_allowed;
4557 return !node_isset(nid, *nodemask);
4560 #define K(x) ((x) << (PAGE_SHIFT-10))
4562 static void show_migration_types(unsigned char type)
4564 static const char types[MIGRATE_TYPES] = {
4565 [MIGRATE_UNMOVABLE] = 'U',
4566 [MIGRATE_MOVABLE] = 'M',
4567 [MIGRATE_RECLAIMABLE] = 'E',
4568 [MIGRATE_HIGHATOMIC] = 'H',
4570 [MIGRATE_CMA] = 'C',
4572 #ifdef CONFIG_MEMORY_ISOLATION
4573 [MIGRATE_ISOLATE] = 'I',
4576 char tmp[MIGRATE_TYPES + 1];
4580 for (i = 0; i < MIGRATE_TYPES; i++) {
4581 if (type & (1 << i))
4586 printk(KERN_CONT "(%s) ", tmp);
4590 * Show free area list (used inside shift_scroll-lock stuff)
4591 * We also calculate the percentage fragmentation. We do this by counting the
4592 * memory on each free list with the exception of the first item on the list.
4595 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4598 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4600 unsigned long free_pcp = 0;
4605 for_each_populated_zone(zone) {
4606 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4609 for_each_online_cpu(cpu)
4610 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4613 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4614 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4615 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4616 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4617 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4618 " free:%lu free_pcp:%lu free_cma:%lu\n",
4619 global_node_page_state(NR_ACTIVE_ANON),
4620 global_node_page_state(NR_INACTIVE_ANON),
4621 global_node_page_state(NR_ISOLATED_ANON),
4622 global_node_page_state(NR_ACTIVE_FILE),
4623 global_node_page_state(NR_INACTIVE_FILE),
4624 global_node_page_state(NR_ISOLATED_FILE),
4625 global_node_page_state(NR_UNEVICTABLE),
4626 global_node_page_state(NR_FILE_DIRTY),
4627 global_node_page_state(NR_WRITEBACK),
4628 global_node_page_state(NR_UNSTABLE_NFS),
4629 global_node_page_state(NR_SLAB_RECLAIMABLE),
4630 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4631 global_node_page_state(NR_FILE_MAPPED),
4632 global_node_page_state(NR_SHMEM),
4633 global_page_state(NR_PAGETABLE),
4634 global_page_state(NR_BOUNCE),
4635 global_page_state(NR_FREE_PAGES),
4637 global_page_state(NR_FREE_CMA_PAGES));
4639 for_each_online_pgdat(pgdat) {
4640 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4644 " active_anon:%lukB"
4645 " inactive_anon:%lukB"
4646 " active_file:%lukB"
4647 " inactive_file:%lukB"
4648 " unevictable:%lukB"
4649 " isolated(anon):%lukB"
4650 " isolated(file):%lukB"
4655 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4657 " shmem_pmdmapped: %lukB"
4660 " writeback_tmp:%lukB"
4662 " all_unreclaimable? %s"
4665 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4666 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4667 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4668 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4669 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4670 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4671 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4672 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4673 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4674 K(node_page_state(pgdat, NR_WRITEBACK)),
4675 K(node_page_state(pgdat, NR_SHMEM)),
4676 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4677 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4678 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4680 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4682 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4683 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4684 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4688 for_each_populated_zone(zone) {
4691 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4695 for_each_online_cpu(cpu)
4696 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4705 " active_anon:%lukB"
4706 " inactive_anon:%lukB"
4707 " active_file:%lukB"
4708 " inactive_file:%lukB"
4709 " unevictable:%lukB"
4710 " writepending:%lukB"
4714 " kernel_stack:%lukB"
4722 K(zone_page_state(zone, NR_FREE_PAGES)),
4723 K(min_wmark_pages(zone)),
4724 K(low_wmark_pages(zone)),
4725 K(high_wmark_pages(zone)),
4726 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4727 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4728 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4729 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4730 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4731 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4732 K(zone->present_pages),
4733 K(zone->managed_pages),
4734 K(zone_page_state(zone, NR_MLOCK)),
4735 zone_page_state(zone, NR_KERNEL_STACK_KB),
4736 K(zone_page_state(zone, NR_PAGETABLE)),
4737 K(zone_page_state(zone, NR_BOUNCE)),
4739 K(this_cpu_read(zone->pageset->pcp.count)),
4740 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4741 printk("lowmem_reserve[]:");
4742 for (i = 0; i < MAX_NR_ZONES; i++)
4743 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4744 printk(KERN_CONT "\n");
4747 for_each_populated_zone(zone) {
4749 unsigned long nr[MAX_ORDER], flags, total = 0;
4750 unsigned char types[MAX_ORDER];
4752 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4755 printk(KERN_CONT "%s: ", zone->name);
4757 spin_lock_irqsave(&zone->lock, flags);
4758 for (order = 0; order < MAX_ORDER; order++) {
4759 struct free_area *area = &zone->free_area[order];
4762 nr[order] = area->nr_free;
4763 total += nr[order] << order;
4766 for (type = 0; type < MIGRATE_TYPES; type++) {
4767 if (!list_empty(&area->free_list[type]))
4768 types[order] |= 1 << type;
4771 spin_unlock_irqrestore(&zone->lock, flags);
4772 for (order = 0; order < MAX_ORDER; order++) {
4773 printk(KERN_CONT "%lu*%lukB ",
4774 nr[order], K(1UL) << order);
4776 show_migration_types(types[order]);
4778 printk(KERN_CONT "= %lukB\n", K(total));
4781 hugetlb_show_meminfo();
4783 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4785 show_swap_cache_info();
4788 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4790 zoneref->zone = zone;
4791 zoneref->zone_idx = zone_idx(zone);
4795 * Builds allocation fallback zone lists.
4797 * Add all populated zones of a node to the zonelist.
4799 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4803 enum zone_type zone_type = MAX_NR_ZONES;
4807 zone = pgdat->node_zones + zone_type;
4808 if (managed_zone(zone)) {
4809 zoneref_set_zone(zone,
4810 &zonelist->_zonerefs[nr_zones++]);
4811 check_highest_zone(zone_type);
4813 } while (zone_type);
4821 * 0 = automatic detection of better ordering.
4822 * 1 = order by ([node] distance, -zonetype)
4823 * 2 = order by (-zonetype, [node] distance)
4825 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4826 * the same zonelist. So only NUMA can configure this param.
4828 #define ZONELIST_ORDER_DEFAULT 0
4829 #define ZONELIST_ORDER_NODE 1
4830 #define ZONELIST_ORDER_ZONE 2
4832 /* zonelist order in the kernel.
4833 * set_zonelist_order() will set this to NODE or ZONE.
4835 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4836 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4840 /* The value user specified ....changed by config */
4841 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4842 /* string for sysctl */
4843 #define NUMA_ZONELIST_ORDER_LEN 16
4844 char numa_zonelist_order[16] = "default";
4847 * interface for configure zonelist ordering.
4848 * command line option "numa_zonelist_order"
4849 * = "[dD]efault - default, automatic configuration.
4850 * = "[nN]ode - order by node locality, then by zone within node
4851 * = "[zZ]one - order by zone, then by locality within zone
4854 static int __parse_numa_zonelist_order(char *s)
4856 if (*s == 'd' || *s == 'D') {
4857 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4858 } else if (*s == 'n' || *s == 'N') {
4859 user_zonelist_order = ZONELIST_ORDER_NODE;
4860 } else if (*s == 'z' || *s == 'Z') {
4861 user_zonelist_order = ZONELIST_ORDER_ZONE;
4863 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4869 static __init int setup_numa_zonelist_order(char *s)
4876 ret = __parse_numa_zonelist_order(s);
4878 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4882 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4885 * sysctl handler for numa_zonelist_order
4887 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4888 void __user *buffer, size_t *length,
4891 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4893 static DEFINE_MUTEX(zl_order_mutex);
4895 mutex_lock(&zl_order_mutex);
4897 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4901 strcpy(saved_string, (char *)table->data);
4903 ret = proc_dostring(table, write, buffer, length, ppos);
4907 int oldval = user_zonelist_order;
4909 ret = __parse_numa_zonelist_order((char *)table->data);
4912 * bogus value. restore saved string
4914 strncpy((char *)table->data, saved_string,
4915 NUMA_ZONELIST_ORDER_LEN);
4916 user_zonelist_order = oldval;
4917 } else if (oldval != user_zonelist_order) {
4918 mem_hotplug_begin();
4919 mutex_lock(&zonelists_mutex);
4920 build_all_zonelists(NULL, NULL);
4921 mutex_unlock(&zonelists_mutex);
4926 mutex_unlock(&zl_order_mutex);
4931 #define MAX_NODE_LOAD (nr_online_nodes)
4932 static int node_load[MAX_NUMNODES];
4935 * find_next_best_node - find the next node that should appear in a given node's fallback list
4936 * @node: node whose fallback list we're appending
4937 * @used_node_mask: nodemask_t of already used nodes
4939 * We use a number of factors to determine which is the next node that should
4940 * appear on a given node's fallback list. The node should not have appeared
4941 * already in @node's fallback list, and it should be the next closest node
4942 * according to the distance array (which contains arbitrary distance values
4943 * from each node to each node in the system), and should also prefer nodes
4944 * with no CPUs, since presumably they'll have very little allocation pressure
4945 * on them otherwise.
4946 * It returns -1 if no node is found.
4948 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4951 int min_val = INT_MAX;
4952 int best_node = NUMA_NO_NODE;
4953 const struct cpumask *tmp = cpumask_of_node(0);
4955 /* Use the local node if we haven't already */
4956 if (!node_isset(node, *used_node_mask)) {
4957 node_set(node, *used_node_mask);
4961 for_each_node_state(n, N_MEMORY) {
4963 /* Don't want a node to appear more than once */
4964 if (node_isset(n, *used_node_mask))
4967 /* Use the distance array to find the distance */
4968 val = node_distance(node, n);
4970 /* Penalize nodes under us ("prefer the next node") */
4973 /* Give preference to headless and unused nodes */
4974 tmp = cpumask_of_node(n);
4975 if (!cpumask_empty(tmp))
4976 val += PENALTY_FOR_NODE_WITH_CPUS;
4978 /* Slight preference for less loaded node */
4979 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4980 val += node_load[n];
4982 if (val < min_val) {
4989 node_set(best_node, *used_node_mask);
4996 * Build zonelists ordered by node and zones within node.
4997 * This results in maximum locality--normal zone overflows into local
4998 * DMA zone, if any--but risks exhausting DMA zone.
5000 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
5003 struct zonelist *zonelist;
5005 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5006 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
5008 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5009 zonelist->_zonerefs[j].zone = NULL;
5010 zonelist->_zonerefs[j].zone_idx = 0;
5014 * Build gfp_thisnode zonelists
5016 static void build_thisnode_zonelists(pg_data_t *pgdat)
5019 struct zonelist *zonelist;
5021 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
5022 j = build_zonelists_node(pgdat, zonelist, 0);
5023 zonelist->_zonerefs[j].zone = NULL;
5024 zonelist->_zonerefs[j].zone_idx = 0;
5028 * Build zonelists ordered by zone and nodes within zones.
5029 * This results in conserving DMA zone[s] until all Normal memory is
5030 * exhausted, but results in overflowing to remote node while memory
5031 * may still exist in local DMA zone.
5033 static int node_order[MAX_NUMNODES];
5035 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
5038 int zone_type; /* needs to be signed */
5040 struct zonelist *zonelist;
5042 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5044 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
5045 for (j = 0; j < nr_nodes; j++) {
5046 node = node_order[j];
5047 z = &NODE_DATA(node)->node_zones[zone_type];
5048 if (managed_zone(z)) {
5050 &zonelist->_zonerefs[pos++]);
5051 check_highest_zone(zone_type);
5055 zonelist->_zonerefs[pos].zone = NULL;
5056 zonelist->_zonerefs[pos].zone_idx = 0;
5059 #if defined(CONFIG_64BIT)
5061 * Devices that require DMA32/DMA are relatively rare and do not justify a
5062 * penalty to every machine in case the specialised case applies. Default
5063 * to Node-ordering on 64-bit NUMA machines
5065 static int default_zonelist_order(void)
5067 return ZONELIST_ORDER_NODE;
5071 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
5072 * by the kernel. If processes running on node 0 deplete the low memory zone
5073 * then reclaim will occur more frequency increasing stalls and potentially
5074 * be easier to OOM if a large percentage of the zone is under writeback or
5075 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
5076 * Hence, default to zone ordering on 32-bit.
5078 static int default_zonelist_order(void)
5080 return ZONELIST_ORDER_ZONE;
5082 #endif /* CONFIG_64BIT */
5084 static void set_zonelist_order(void)
5086 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
5087 current_zonelist_order = default_zonelist_order();
5089 current_zonelist_order = user_zonelist_order;
5092 static void build_zonelists(pg_data_t *pgdat)
5095 nodemask_t used_mask;
5096 int local_node, prev_node;
5097 struct zonelist *zonelist;
5098 unsigned int order = current_zonelist_order;
5100 /* initialize zonelists */
5101 for (i = 0; i < MAX_ZONELISTS; i++) {
5102 zonelist = pgdat->node_zonelists + i;
5103 zonelist->_zonerefs[0].zone = NULL;
5104 zonelist->_zonerefs[0].zone_idx = 0;
5107 /* NUMA-aware ordering of nodes */
5108 local_node = pgdat->node_id;
5109 load = nr_online_nodes;
5110 prev_node = local_node;
5111 nodes_clear(used_mask);
5113 memset(node_order, 0, sizeof(node_order));
5116 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5118 * We don't want to pressure a particular node.
5119 * So adding penalty to the first node in same
5120 * distance group to make it round-robin.
5122 if (node_distance(local_node, node) !=
5123 node_distance(local_node, prev_node))
5124 node_load[node] = load;
5128 if (order == ZONELIST_ORDER_NODE)
5129 build_zonelists_in_node_order(pgdat, node);
5131 node_order[i++] = node; /* remember order */
5134 if (order == ZONELIST_ORDER_ZONE) {
5135 /* calculate node order -- i.e., DMA last! */
5136 build_zonelists_in_zone_order(pgdat, i);
5139 build_thisnode_zonelists(pgdat);
5142 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5144 * Return node id of node used for "local" allocations.
5145 * I.e., first node id of first zone in arg node's generic zonelist.
5146 * Used for initializing percpu 'numa_mem', which is used primarily
5147 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5149 int local_memory_node(int node)
5153 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5154 gfp_zone(GFP_KERNEL),
5156 return z->zone->node;
5160 static void setup_min_unmapped_ratio(void);
5161 static void setup_min_slab_ratio(void);
5162 #else /* CONFIG_NUMA */
5164 static void set_zonelist_order(void)
5166 current_zonelist_order = ZONELIST_ORDER_ZONE;
5169 static void build_zonelists(pg_data_t *pgdat)
5171 int node, local_node;
5173 struct zonelist *zonelist;
5175 local_node = pgdat->node_id;
5177 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5178 j = build_zonelists_node(pgdat, zonelist, 0);
5181 * Now we build the zonelist so that it contains the zones
5182 * of all the other nodes.
5183 * We don't want to pressure a particular node, so when
5184 * building the zones for node N, we make sure that the
5185 * zones coming right after the local ones are those from
5186 * node N+1 (modulo N)
5188 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5189 if (!node_online(node))
5191 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5193 for (node = 0; node < local_node; node++) {
5194 if (!node_online(node))
5196 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5199 zonelist->_zonerefs[j].zone = NULL;
5200 zonelist->_zonerefs[j].zone_idx = 0;
5203 #endif /* CONFIG_NUMA */
5206 * Boot pageset table. One per cpu which is going to be used for all
5207 * zones and all nodes. The parameters will be set in such a way
5208 * that an item put on a list will immediately be handed over to
5209 * the buddy list. This is safe since pageset manipulation is done
5210 * with interrupts disabled.
5212 * The boot_pagesets must be kept even after bootup is complete for
5213 * unused processors and/or zones. They do play a role for bootstrapping
5214 * hotplugged processors.
5216 * zoneinfo_show() and maybe other functions do
5217 * not check if the processor is online before following the pageset pointer.
5218 * Other parts of the kernel may not check if the zone is available.
5220 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5221 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5222 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5223 static void setup_zone_pageset(struct zone *zone);
5226 * Global mutex to protect against size modification of zonelists
5227 * as well as to serialize pageset setup for the new populated zone.
5229 DEFINE_MUTEX(zonelists_mutex);
5231 /* return values int ....just for stop_machine() */
5232 static int __build_all_zonelists(void *data)
5236 pg_data_t *self = data;
5239 memset(node_load, 0, sizeof(node_load));
5242 if (self && !node_online(self->node_id)) {
5243 build_zonelists(self);
5246 for_each_online_node(nid) {
5247 pg_data_t *pgdat = NODE_DATA(nid);
5249 build_zonelists(pgdat);
5253 * Initialize the boot_pagesets that are going to be used
5254 * for bootstrapping processors. The real pagesets for
5255 * each zone will be allocated later when the per cpu
5256 * allocator is available.
5258 * boot_pagesets are used also for bootstrapping offline
5259 * cpus if the system is already booted because the pagesets
5260 * are needed to initialize allocators on a specific cpu too.
5261 * F.e. the percpu allocator needs the page allocator which
5262 * needs the percpu allocator in order to allocate its pagesets
5263 * (a chicken-egg dilemma).
5265 for_each_possible_cpu(cpu) {
5266 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5268 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5270 * We now know the "local memory node" for each node--
5271 * i.e., the node of the first zone in the generic zonelist.
5272 * Set up numa_mem percpu variable for on-line cpus. During
5273 * boot, only the boot cpu should be on-line; we'll init the
5274 * secondary cpus' numa_mem as they come on-line. During
5275 * node/memory hotplug, we'll fixup all on-line cpus.
5277 if (cpu_online(cpu))
5278 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5285 static noinline void __init
5286 build_all_zonelists_init(void)
5288 __build_all_zonelists(NULL);
5289 mminit_verify_zonelist();
5290 cpuset_init_current_mems_allowed();
5294 * Called with zonelists_mutex held always
5295 * unless system_state == SYSTEM_BOOTING.
5297 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5298 * [we're only called with non-NULL zone through __meminit paths] and
5299 * (2) call of __init annotated helper build_all_zonelists_init
5300 * [protected by SYSTEM_BOOTING].
5302 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5304 set_zonelist_order();
5306 if (system_state == SYSTEM_BOOTING) {
5307 build_all_zonelists_init();
5309 #ifdef CONFIG_MEMORY_HOTPLUG
5311 setup_zone_pageset(zone);
5313 /* we have to stop all cpus to guarantee there is no user
5315 stop_machine_cpuslocked(__build_all_zonelists, pgdat, NULL);
5316 /* cpuset refresh routine should be here */
5318 vm_total_pages = nr_free_pagecache_pages();
5320 * Disable grouping by mobility if the number of pages in the
5321 * system is too low to allow the mechanism to work. It would be
5322 * more accurate, but expensive to check per-zone. This check is
5323 * made on memory-hotadd so a system can start with mobility
5324 * disabled and enable it later
5326 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5327 page_group_by_mobility_disabled = 1;
5329 page_group_by_mobility_disabled = 0;
5331 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5333 zonelist_order_name[current_zonelist_order],
5334 page_group_by_mobility_disabled ? "off" : "on",
5337 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5342 * Initially all pages are reserved - free ones are freed
5343 * up by free_all_bootmem() once the early boot process is
5344 * done. Non-atomic initialization, single-pass.
5346 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5347 unsigned long start_pfn, enum memmap_context context)
5349 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5350 unsigned long end_pfn = start_pfn + size;
5351 pg_data_t *pgdat = NODE_DATA(nid);
5353 unsigned long nr_initialised = 0;
5354 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5355 struct memblock_region *r = NULL, *tmp;
5358 if (highest_memmap_pfn < end_pfn - 1)
5359 highest_memmap_pfn = end_pfn - 1;
5362 * Honor reservation requested by the driver for this ZONE_DEVICE
5365 if (altmap && start_pfn == altmap->base_pfn)
5366 start_pfn += altmap->reserve;
5368 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5370 * There can be holes in boot-time mem_map[]s handed to this
5371 * function. They do not exist on hotplugged memory.
5373 if (context != MEMMAP_EARLY)
5376 if (!early_pfn_valid(pfn)) {
5377 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5379 * Skip to the pfn preceding the next valid one (or
5380 * end_pfn), such that we hit a valid pfn (or end_pfn)
5381 * on our next iteration of the loop.
5383 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5387 if (!early_pfn_in_nid(pfn, nid))
5389 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5392 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5394 * Check given memblock attribute by firmware which can affect
5395 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5396 * mirrored, it's an overlapped memmap init. skip it.
5398 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5399 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5400 for_each_memblock(memory, tmp)
5401 if (pfn < memblock_region_memory_end_pfn(tmp))
5405 if (pfn >= memblock_region_memory_base_pfn(r) &&
5406 memblock_is_mirror(r)) {
5407 /* already initialized as NORMAL */
5408 pfn = memblock_region_memory_end_pfn(r);
5416 * Mark the block movable so that blocks are reserved for
5417 * movable at startup. This will force kernel allocations
5418 * to reserve their blocks rather than leaking throughout
5419 * the address space during boot when many long-lived
5420 * kernel allocations are made.
5422 * bitmap is created for zone's valid pfn range. but memmap
5423 * can be created for invalid pages (for alignment)
5424 * check here not to call set_pageblock_migratetype() against
5427 if (!(pfn & (pageblock_nr_pages - 1))) {
5428 struct page *page = pfn_to_page(pfn);
5430 __init_single_page(page, pfn, zone, nid);
5431 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5433 __init_single_pfn(pfn, zone, nid);
5438 static void __meminit zone_init_free_lists(struct zone *zone)
5440 unsigned int order, t;
5441 for_each_migratetype_order(order, t) {
5442 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5443 zone->free_area[order].nr_free = 0;
5447 #ifndef __HAVE_ARCH_MEMMAP_INIT
5448 #define memmap_init(size, nid, zone, start_pfn) \
5449 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5452 static int zone_batchsize(struct zone *zone)
5458 * The per-cpu-pages pools are set to around 1000th of the
5459 * size of the zone. But no more than 1/2 of a meg.
5461 * OK, so we don't know how big the cache is. So guess.
5463 batch = zone->managed_pages / 1024;
5464 if (batch * PAGE_SIZE > 512 * 1024)
5465 batch = (512 * 1024) / PAGE_SIZE;
5466 batch /= 4; /* We effectively *= 4 below */
5471 * Clamp the batch to a 2^n - 1 value. Having a power
5472 * of 2 value was found to be more likely to have
5473 * suboptimal cache aliasing properties in some cases.
5475 * For example if 2 tasks are alternately allocating
5476 * batches of pages, one task can end up with a lot
5477 * of pages of one half of the possible page colors
5478 * and the other with pages of the other colors.
5480 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5485 /* The deferral and batching of frees should be suppressed under NOMMU
5488 * The problem is that NOMMU needs to be able to allocate large chunks
5489 * of contiguous memory as there's no hardware page translation to
5490 * assemble apparent contiguous memory from discontiguous pages.
5492 * Queueing large contiguous runs of pages for batching, however,
5493 * causes the pages to actually be freed in smaller chunks. As there
5494 * can be a significant delay between the individual batches being
5495 * recycled, this leads to the once large chunks of space being
5496 * fragmented and becoming unavailable for high-order allocations.
5503 * pcp->high and pcp->batch values are related and dependent on one another:
5504 * ->batch must never be higher then ->high.
5505 * The following function updates them in a safe manner without read side
5508 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5509 * those fields changing asynchronously (acording the the above rule).
5511 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5512 * outside of boot time (or some other assurance that no concurrent updaters
5515 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5516 unsigned long batch)
5518 /* start with a fail safe value for batch */
5522 /* Update high, then batch, in order */
5529 /* a companion to pageset_set_high() */
5530 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5532 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5535 static void pageset_init(struct per_cpu_pageset *p)
5537 struct per_cpu_pages *pcp;
5540 memset(p, 0, sizeof(*p));
5544 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5545 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5548 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5551 pageset_set_batch(p, batch);
5555 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5556 * to the value high for the pageset p.
5558 static void pageset_set_high(struct per_cpu_pageset *p,
5561 unsigned long batch = max(1UL, high / 4);
5562 if ((high / 4) > (PAGE_SHIFT * 8))
5563 batch = PAGE_SHIFT * 8;
5565 pageset_update(&p->pcp, high, batch);
5568 static void pageset_set_high_and_batch(struct zone *zone,
5569 struct per_cpu_pageset *pcp)
5571 if (percpu_pagelist_fraction)
5572 pageset_set_high(pcp,
5573 (zone->managed_pages /
5574 percpu_pagelist_fraction));
5576 pageset_set_batch(pcp, zone_batchsize(zone));
5579 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5581 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5584 pageset_set_high_and_batch(zone, pcp);
5587 static void __meminit setup_zone_pageset(struct zone *zone)
5590 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5591 for_each_possible_cpu(cpu)
5592 zone_pageset_init(zone, cpu);
5596 * Allocate per cpu pagesets and initialize them.
5597 * Before this call only boot pagesets were available.
5599 void __init setup_per_cpu_pageset(void)
5601 struct pglist_data *pgdat;
5604 for_each_populated_zone(zone)
5605 setup_zone_pageset(zone);
5607 for_each_online_pgdat(pgdat)
5608 pgdat->per_cpu_nodestats =
5609 alloc_percpu(struct per_cpu_nodestat);
5612 static __meminit void zone_pcp_init(struct zone *zone)
5615 * per cpu subsystem is not up at this point. The following code
5616 * relies on the ability of the linker to provide the
5617 * offset of a (static) per cpu variable into the per cpu area.
5619 zone->pageset = &boot_pageset;
5621 if (populated_zone(zone))
5622 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5623 zone->name, zone->present_pages,
5624 zone_batchsize(zone));
5627 void __meminit init_currently_empty_zone(struct zone *zone,
5628 unsigned long zone_start_pfn,
5631 struct pglist_data *pgdat = zone->zone_pgdat;
5633 pgdat->nr_zones = zone_idx(zone) + 1;
5635 zone->zone_start_pfn = zone_start_pfn;
5637 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5638 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5640 (unsigned long)zone_idx(zone),
5641 zone_start_pfn, (zone_start_pfn + size));
5643 zone_init_free_lists(zone);
5644 zone->initialized = 1;
5647 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5648 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5651 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5653 int __meminit __early_pfn_to_nid(unsigned long pfn,
5654 struct mminit_pfnnid_cache *state)
5656 unsigned long start_pfn, end_pfn;
5659 if (state->last_start <= pfn && pfn < state->last_end)
5660 return state->last_nid;
5662 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5664 state->last_start = start_pfn;
5665 state->last_end = end_pfn;
5666 state->last_nid = nid;
5671 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5674 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5675 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5676 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5678 * If an architecture guarantees that all ranges registered contain no holes
5679 * and may be freed, this this function may be used instead of calling
5680 * memblock_free_early_nid() manually.
5682 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5684 unsigned long start_pfn, end_pfn;
5687 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5688 start_pfn = min(start_pfn, max_low_pfn);
5689 end_pfn = min(end_pfn, max_low_pfn);
5691 if (start_pfn < end_pfn)
5692 memblock_free_early_nid(PFN_PHYS(start_pfn),
5693 (end_pfn - start_pfn) << PAGE_SHIFT,
5699 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5700 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5702 * If an architecture guarantees that all ranges registered contain no holes and may
5703 * be freed, this function may be used instead of calling memory_present() manually.
5705 void __init sparse_memory_present_with_active_regions(int nid)
5707 unsigned long start_pfn, end_pfn;
5710 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5711 memory_present(this_nid, start_pfn, end_pfn);
5715 * get_pfn_range_for_nid - Return the start and end page frames for a node
5716 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5717 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5718 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5720 * It returns the start and end page frame of a node based on information
5721 * provided by memblock_set_node(). If called for a node
5722 * with no available memory, a warning is printed and the start and end
5725 void __meminit get_pfn_range_for_nid(unsigned int nid,
5726 unsigned long *start_pfn, unsigned long *end_pfn)
5728 unsigned long this_start_pfn, this_end_pfn;
5734 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5735 *start_pfn = min(*start_pfn, this_start_pfn);
5736 *end_pfn = max(*end_pfn, this_end_pfn);
5739 if (*start_pfn == -1UL)
5744 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5745 * assumption is made that zones within a node are ordered in monotonic
5746 * increasing memory addresses so that the "highest" populated zone is used
5748 static void __init find_usable_zone_for_movable(void)
5751 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5752 if (zone_index == ZONE_MOVABLE)
5755 if (arch_zone_highest_possible_pfn[zone_index] >
5756 arch_zone_lowest_possible_pfn[zone_index])
5760 VM_BUG_ON(zone_index == -1);
5761 movable_zone = zone_index;
5765 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5766 * because it is sized independent of architecture. Unlike the other zones,
5767 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5768 * in each node depending on the size of each node and how evenly kernelcore
5769 * is distributed. This helper function adjusts the zone ranges
5770 * provided by the architecture for a given node by using the end of the
5771 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5772 * zones within a node are in order of monotonic increases memory addresses
5774 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5775 unsigned long zone_type,
5776 unsigned long node_start_pfn,
5777 unsigned long node_end_pfn,
5778 unsigned long *zone_start_pfn,
5779 unsigned long *zone_end_pfn)
5781 /* Only adjust if ZONE_MOVABLE is on this node */
5782 if (zone_movable_pfn[nid]) {
5783 /* Size ZONE_MOVABLE */
5784 if (zone_type == ZONE_MOVABLE) {
5785 *zone_start_pfn = zone_movable_pfn[nid];
5786 *zone_end_pfn = min(node_end_pfn,
5787 arch_zone_highest_possible_pfn[movable_zone]);
5789 /* Adjust for ZONE_MOVABLE starting within this range */
5790 } else if (!mirrored_kernelcore &&
5791 *zone_start_pfn < zone_movable_pfn[nid] &&
5792 *zone_end_pfn > zone_movable_pfn[nid]) {
5793 *zone_end_pfn = zone_movable_pfn[nid];
5795 /* Check if this whole range is within ZONE_MOVABLE */
5796 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5797 *zone_start_pfn = *zone_end_pfn;
5802 * Return the number of pages a zone spans in a node, including holes
5803 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5805 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5806 unsigned long zone_type,
5807 unsigned long node_start_pfn,
5808 unsigned long node_end_pfn,
5809 unsigned long *zone_start_pfn,
5810 unsigned long *zone_end_pfn,
5811 unsigned long *ignored)
5813 /* When hotadd a new node from cpu_up(), the node should be empty */
5814 if (!node_start_pfn && !node_end_pfn)
5817 /* Get the start and end of the zone */
5818 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5819 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5820 adjust_zone_range_for_zone_movable(nid, zone_type,
5821 node_start_pfn, node_end_pfn,
5822 zone_start_pfn, zone_end_pfn);
5824 /* Check that this node has pages within the zone's required range */
5825 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5828 /* Move the zone boundaries inside the node if necessary */
5829 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5830 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5832 /* Return the spanned pages */
5833 return *zone_end_pfn - *zone_start_pfn;
5837 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5838 * then all holes in the requested range will be accounted for.
5840 unsigned long __meminit __absent_pages_in_range(int nid,
5841 unsigned long range_start_pfn,
5842 unsigned long range_end_pfn)
5844 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5845 unsigned long start_pfn, end_pfn;
5848 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5849 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5850 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5851 nr_absent -= end_pfn - start_pfn;
5857 * absent_pages_in_range - Return number of page frames in holes within a range
5858 * @start_pfn: The start PFN to start searching for holes
5859 * @end_pfn: The end PFN to stop searching for holes
5861 * It returns the number of pages frames in memory holes within a range.
5863 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5864 unsigned long end_pfn)
5866 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5869 /* Return the number of page frames in holes in a zone on a node */
5870 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5871 unsigned long zone_type,
5872 unsigned long node_start_pfn,
5873 unsigned long node_end_pfn,
5874 unsigned long *ignored)
5876 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5877 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5878 unsigned long zone_start_pfn, zone_end_pfn;
5879 unsigned long nr_absent;
5881 /* When hotadd a new node from cpu_up(), the node should be empty */
5882 if (!node_start_pfn && !node_end_pfn)
5885 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5886 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5888 adjust_zone_range_for_zone_movable(nid, zone_type,
5889 node_start_pfn, node_end_pfn,
5890 &zone_start_pfn, &zone_end_pfn);
5891 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5894 * ZONE_MOVABLE handling.
5895 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5898 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5899 unsigned long start_pfn, end_pfn;
5900 struct memblock_region *r;
5902 for_each_memblock(memory, r) {
5903 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5904 zone_start_pfn, zone_end_pfn);
5905 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5906 zone_start_pfn, zone_end_pfn);
5908 if (zone_type == ZONE_MOVABLE &&
5909 memblock_is_mirror(r))
5910 nr_absent += end_pfn - start_pfn;
5912 if (zone_type == ZONE_NORMAL &&
5913 !memblock_is_mirror(r))
5914 nr_absent += end_pfn - start_pfn;
5921 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5922 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5923 unsigned long zone_type,
5924 unsigned long node_start_pfn,
5925 unsigned long node_end_pfn,
5926 unsigned long *zone_start_pfn,
5927 unsigned long *zone_end_pfn,
5928 unsigned long *zones_size)
5932 *zone_start_pfn = node_start_pfn;
5933 for (zone = 0; zone < zone_type; zone++)
5934 *zone_start_pfn += zones_size[zone];
5936 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5938 return zones_size[zone_type];
5941 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5942 unsigned long zone_type,
5943 unsigned long node_start_pfn,
5944 unsigned long node_end_pfn,
5945 unsigned long *zholes_size)
5950 return zholes_size[zone_type];
5953 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5955 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5956 unsigned long node_start_pfn,
5957 unsigned long node_end_pfn,
5958 unsigned long *zones_size,
5959 unsigned long *zholes_size)
5961 unsigned long realtotalpages = 0, totalpages = 0;
5964 for (i = 0; i < MAX_NR_ZONES; i++) {
5965 struct zone *zone = pgdat->node_zones + i;
5966 unsigned long zone_start_pfn, zone_end_pfn;
5967 unsigned long size, real_size;
5969 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5975 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5976 node_start_pfn, node_end_pfn,
5979 zone->zone_start_pfn = zone_start_pfn;
5981 zone->zone_start_pfn = 0;
5982 zone->spanned_pages = size;
5983 zone->present_pages = real_size;
5986 realtotalpages += real_size;
5989 pgdat->node_spanned_pages = totalpages;
5990 pgdat->node_present_pages = realtotalpages;
5991 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5995 #ifndef CONFIG_SPARSEMEM
5997 * Calculate the size of the zone->blockflags rounded to an unsigned long
5998 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5999 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6000 * round what is now in bits to nearest long in bits, then return it in
6003 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6005 unsigned long usemapsize;
6007 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6008 usemapsize = roundup(zonesize, pageblock_nr_pages);
6009 usemapsize = usemapsize >> pageblock_order;
6010 usemapsize *= NR_PAGEBLOCK_BITS;
6011 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6013 return usemapsize / 8;
6016 static void __init setup_usemap(struct pglist_data *pgdat,
6018 unsigned long zone_start_pfn,
6019 unsigned long zonesize)
6021 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6022 zone->pageblock_flags = NULL;
6024 zone->pageblock_flags =
6025 memblock_virt_alloc_node_nopanic(usemapsize,
6029 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6030 unsigned long zone_start_pfn, unsigned long zonesize) {}
6031 #endif /* CONFIG_SPARSEMEM */
6033 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6035 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6036 void __paginginit set_pageblock_order(void)
6040 /* Check that pageblock_nr_pages has not already been setup */
6041 if (pageblock_order)
6044 if (HPAGE_SHIFT > PAGE_SHIFT)
6045 order = HUGETLB_PAGE_ORDER;
6047 order = MAX_ORDER - 1;
6050 * Assume the largest contiguous order of interest is a huge page.
6051 * This value may be variable depending on boot parameters on IA64 and
6054 pageblock_order = order;
6056 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6059 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6060 * is unused as pageblock_order is set at compile-time. See
6061 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6064 void __paginginit set_pageblock_order(void)
6068 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6070 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6071 unsigned long present_pages)
6073 unsigned long pages = spanned_pages;
6076 * Provide a more accurate estimation if there are holes within
6077 * the zone and SPARSEMEM is in use. If there are holes within the
6078 * zone, each populated memory region may cost us one or two extra
6079 * memmap pages due to alignment because memmap pages for each
6080 * populated regions may not be naturally aligned on page boundary.
6081 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6083 if (spanned_pages > present_pages + (present_pages >> 4) &&
6084 IS_ENABLED(CONFIG_SPARSEMEM))
6085 pages = present_pages;
6087 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6091 * Set up the zone data structures:
6092 * - mark all pages reserved
6093 * - mark all memory queues empty
6094 * - clear the memory bitmaps
6096 * NOTE: pgdat should get zeroed by caller.
6098 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6101 int nid = pgdat->node_id;
6103 pgdat_resize_init(pgdat);
6104 #ifdef CONFIG_NUMA_BALANCING
6105 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6106 pgdat->numabalancing_migrate_nr_pages = 0;
6107 pgdat->numabalancing_migrate_next_window = jiffies;
6109 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6110 spin_lock_init(&pgdat->split_queue_lock);
6111 INIT_LIST_HEAD(&pgdat->split_queue);
6112 pgdat->split_queue_len = 0;
6114 init_waitqueue_head(&pgdat->kswapd_wait);
6115 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6116 #ifdef CONFIG_COMPACTION
6117 init_waitqueue_head(&pgdat->kcompactd_wait);
6119 pgdat_page_ext_init(pgdat);
6120 spin_lock_init(&pgdat->lru_lock);
6121 lruvec_init(node_lruvec(pgdat));
6123 pgdat->per_cpu_nodestats = &boot_nodestats;
6125 for (j = 0; j < MAX_NR_ZONES; j++) {
6126 struct zone *zone = pgdat->node_zones + j;
6127 unsigned long size, realsize, freesize, memmap_pages;
6128 unsigned long zone_start_pfn = zone->zone_start_pfn;
6130 size = zone->spanned_pages;
6131 realsize = freesize = zone->present_pages;
6134 * Adjust freesize so that it accounts for how much memory
6135 * is used by this zone for memmap. This affects the watermark
6136 * and per-cpu initialisations
6138 memmap_pages = calc_memmap_size(size, realsize);
6139 if (!is_highmem_idx(j)) {
6140 if (freesize >= memmap_pages) {
6141 freesize -= memmap_pages;
6144 " %s zone: %lu pages used for memmap\n",
6145 zone_names[j], memmap_pages);
6147 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6148 zone_names[j], memmap_pages, freesize);
6151 /* Account for reserved pages */
6152 if (j == 0 && freesize > dma_reserve) {
6153 freesize -= dma_reserve;
6154 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6155 zone_names[0], dma_reserve);
6158 if (!is_highmem_idx(j))
6159 nr_kernel_pages += freesize;
6160 /* Charge for highmem memmap if there are enough kernel pages */
6161 else if (nr_kernel_pages > memmap_pages * 2)
6162 nr_kernel_pages -= memmap_pages;
6163 nr_all_pages += freesize;
6166 * Set an approximate value for lowmem here, it will be adjusted
6167 * when the bootmem allocator frees pages into the buddy system.
6168 * And all highmem pages will be managed by the buddy system.
6170 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6174 zone->name = zone_names[j];
6175 zone->zone_pgdat = pgdat;
6176 spin_lock_init(&zone->lock);
6177 zone_seqlock_init(zone);
6178 zone_pcp_init(zone);
6183 set_pageblock_order();
6184 setup_usemap(pgdat, zone, zone_start_pfn, size);
6185 init_currently_empty_zone(zone, zone_start_pfn, size);
6186 memmap_init(size, nid, j, zone_start_pfn);
6190 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6192 unsigned long __maybe_unused start = 0;
6193 unsigned long __maybe_unused offset = 0;
6195 /* Skip empty nodes */
6196 if (!pgdat->node_spanned_pages)
6199 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6200 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6201 offset = pgdat->node_start_pfn - start;
6202 /* ia64 gets its own node_mem_map, before this, without bootmem */
6203 if (!pgdat->node_mem_map) {
6204 unsigned long size, end;
6208 * The zone's endpoints aren't required to be MAX_ORDER
6209 * aligned but the node_mem_map endpoints must be in order
6210 * for the buddy allocator to function correctly.
6212 end = pgdat_end_pfn(pgdat);
6213 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6214 size = (end - start) * sizeof(struct page);
6215 map = alloc_remap(pgdat->node_id, size);
6217 map = memblock_virt_alloc_node_nopanic(size,
6219 pgdat->node_mem_map = map + offset;
6221 #ifndef CONFIG_NEED_MULTIPLE_NODES
6223 * With no DISCONTIG, the global mem_map is just set as node 0's
6225 if (pgdat == NODE_DATA(0)) {
6226 mem_map = NODE_DATA(0)->node_mem_map;
6227 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6228 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6230 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6233 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6236 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6237 unsigned long node_start_pfn, unsigned long *zholes_size)
6239 pg_data_t *pgdat = NODE_DATA(nid);
6240 unsigned long start_pfn = 0;
6241 unsigned long end_pfn = 0;
6243 /* pg_data_t should be reset to zero when it's allocated */
6244 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6246 pgdat->node_id = nid;
6247 pgdat->node_start_pfn = node_start_pfn;
6248 pgdat->per_cpu_nodestats = NULL;
6249 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6250 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6251 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6252 (u64)start_pfn << PAGE_SHIFT,
6253 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6255 start_pfn = node_start_pfn;
6257 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6258 zones_size, zholes_size);
6260 alloc_node_mem_map(pgdat);
6261 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6262 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6263 nid, (unsigned long)pgdat,
6264 (unsigned long)pgdat->node_mem_map);
6267 reset_deferred_meminit(pgdat);
6268 free_area_init_core(pgdat);
6271 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6273 #if MAX_NUMNODES > 1
6275 * Figure out the number of possible node ids.
6277 void __init setup_nr_node_ids(void)
6279 unsigned int highest;
6281 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6282 nr_node_ids = highest + 1;
6287 * node_map_pfn_alignment - determine the maximum internode alignment
6289 * This function should be called after node map is populated and sorted.
6290 * It calculates the maximum power of two alignment which can distinguish
6293 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6294 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6295 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6296 * shifted, 1GiB is enough and this function will indicate so.
6298 * This is used to test whether pfn -> nid mapping of the chosen memory
6299 * model has fine enough granularity to avoid incorrect mapping for the
6300 * populated node map.
6302 * Returns the determined alignment in pfn's. 0 if there is no alignment
6303 * requirement (single node).
6305 unsigned long __init node_map_pfn_alignment(void)
6307 unsigned long accl_mask = 0, last_end = 0;
6308 unsigned long start, end, mask;
6312 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6313 if (!start || last_nid < 0 || last_nid == nid) {
6320 * Start with a mask granular enough to pin-point to the
6321 * start pfn and tick off bits one-by-one until it becomes
6322 * too coarse to separate the current node from the last.
6324 mask = ~((1 << __ffs(start)) - 1);
6325 while (mask && last_end <= (start & (mask << 1)))
6328 /* accumulate all internode masks */
6332 /* convert mask to number of pages */
6333 return ~accl_mask + 1;
6336 /* Find the lowest pfn for a node */
6337 static unsigned long __init find_min_pfn_for_node(int nid)
6339 unsigned long min_pfn = ULONG_MAX;
6340 unsigned long start_pfn;
6343 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6344 min_pfn = min(min_pfn, start_pfn);
6346 if (min_pfn == ULONG_MAX) {
6347 pr_warn("Could not find start_pfn for node %d\n", nid);
6355 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6357 * It returns the minimum PFN based on information provided via
6358 * memblock_set_node().
6360 unsigned long __init find_min_pfn_with_active_regions(void)
6362 return find_min_pfn_for_node(MAX_NUMNODES);
6366 * early_calculate_totalpages()
6367 * Sum pages in active regions for movable zone.
6368 * Populate N_MEMORY for calculating usable_nodes.
6370 static unsigned long __init early_calculate_totalpages(void)
6372 unsigned long totalpages = 0;
6373 unsigned long start_pfn, end_pfn;
6376 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6377 unsigned long pages = end_pfn - start_pfn;
6379 totalpages += pages;
6381 node_set_state(nid, N_MEMORY);
6387 * Find the PFN the Movable zone begins in each node. Kernel memory
6388 * is spread evenly between nodes as long as the nodes have enough
6389 * memory. When they don't, some nodes will have more kernelcore than
6392 static void __init find_zone_movable_pfns_for_nodes(void)
6395 unsigned long usable_startpfn;
6396 unsigned long kernelcore_node, kernelcore_remaining;
6397 /* save the state before borrow the nodemask */
6398 nodemask_t saved_node_state = node_states[N_MEMORY];
6399 unsigned long totalpages = early_calculate_totalpages();
6400 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6401 struct memblock_region *r;
6403 /* Need to find movable_zone earlier when movable_node is specified. */
6404 find_usable_zone_for_movable();
6407 * If movable_node is specified, ignore kernelcore and movablecore
6410 if (movable_node_is_enabled()) {
6411 for_each_memblock(memory, r) {
6412 if (!memblock_is_hotpluggable(r))
6417 usable_startpfn = PFN_DOWN(r->base);
6418 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6419 min(usable_startpfn, zone_movable_pfn[nid]) :
6427 * If kernelcore=mirror is specified, ignore movablecore option
6429 if (mirrored_kernelcore) {
6430 bool mem_below_4gb_not_mirrored = false;
6432 for_each_memblock(memory, r) {
6433 if (memblock_is_mirror(r))
6438 usable_startpfn = memblock_region_memory_base_pfn(r);
6440 if (usable_startpfn < 0x100000) {
6441 mem_below_4gb_not_mirrored = true;
6445 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6446 min(usable_startpfn, zone_movable_pfn[nid]) :
6450 if (mem_below_4gb_not_mirrored)
6451 pr_warn("This configuration results in unmirrored kernel memory.");
6457 * If movablecore=nn[KMG] was specified, calculate what size of
6458 * kernelcore that corresponds so that memory usable for
6459 * any allocation type is evenly spread. If both kernelcore
6460 * and movablecore are specified, then the value of kernelcore
6461 * will be used for required_kernelcore if it's greater than
6462 * what movablecore would have allowed.
6464 if (required_movablecore) {
6465 unsigned long corepages;
6468 * Round-up so that ZONE_MOVABLE is at least as large as what
6469 * was requested by the user
6471 required_movablecore =
6472 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6473 required_movablecore = min(totalpages, required_movablecore);
6474 corepages = totalpages - required_movablecore;
6476 required_kernelcore = max(required_kernelcore, corepages);
6480 * If kernelcore was not specified or kernelcore size is larger
6481 * than totalpages, there is no ZONE_MOVABLE.
6483 if (!required_kernelcore || required_kernelcore >= totalpages)
6486 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6487 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6490 /* Spread kernelcore memory as evenly as possible throughout nodes */
6491 kernelcore_node = required_kernelcore / usable_nodes;
6492 for_each_node_state(nid, N_MEMORY) {
6493 unsigned long start_pfn, end_pfn;
6496 * Recalculate kernelcore_node if the division per node
6497 * now exceeds what is necessary to satisfy the requested
6498 * amount of memory for the kernel
6500 if (required_kernelcore < kernelcore_node)
6501 kernelcore_node = required_kernelcore / usable_nodes;
6504 * As the map is walked, we track how much memory is usable
6505 * by the kernel using kernelcore_remaining. When it is
6506 * 0, the rest of the node is usable by ZONE_MOVABLE
6508 kernelcore_remaining = kernelcore_node;
6510 /* Go through each range of PFNs within this node */
6511 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6512 unsigned long size_pages;
6514 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6515 if (start_pfn >= end_pfn)
6518 /* Account for what is only usable for kernelcore */
6519 if (start_pfn < usable_startpfn) {
6520 unsigned long kernel_pages;
6521 kernel_pages = min(end_pfn, usable_startpfn)
6524 kernelcore_remaining -= min(kernel_pages,
6525 kernelcore_remaining);
6526 required_kernelcore -= min(kernel_pages,
6527 required_kernelcore);
6529 /* Continue if range is now fully accounted */
6530 if (end_pfn <= usable_startpfn) {
6533 * Push zone_movable_pfn to the end so
6534 * that if we have to rebalance
6535 * kernelcore across nodes, we will
6536 * not double account here
6538 zone_movable_pfn[nid] = end_pfn;
6541 start_pfn = usable_startpfn;
6545 * The usable PFN range for ZONE_MOVABLE is from
6546 * start_pfn->end_pfn. Calculate size_pages as the
6547 * number of pages used as kernelcore
6549 size_pages = end_pfn - start_pfn;
6550 if (size_pages > kernelcore_remaining)
6551 size_pages = kernelcore_remaining;
6552 zone_movable_pfn[nid] = start_pfn + size_pages;
6555 * Some kernelcore has been met, update counts and
6556 * break if the kernelcore for this node has been
6559 required_kernelcore -= min(required_kernelcore,
6561 kernelcore_remaining -= size_pages;
6562 if (!kernelcore_remaining)
6568 * If there is still required_kernelcore, we do another pass with one
6569 * less node in the count. This will push zone_movable_pfn[nid] further
6570 * along on the nodes that still have memory until kernelcore is
6574 if (usable_nodes && required_kernelcore > usable_nodes)
6578 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6579 for (nid = 0; nid < MAX_NUMNODES; nid++)
6580 zone_movable_pfn[nid] =
6581 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6584 /* restore the node_state */
6585 node_states[N_MEMORY] = saved_node_state;
6588 /* Any regular or high memory on that node ? */
6589 static void check_for_memory(pg_data_t *pgdat, int nid)
6591 enum zone_type zone_type;
6593 if (N_MEMORY == N_NORMAL_MEMORY)
6596 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6597 struct zone *zone = &pgdat->node_zones[zone_type];
6598 if (populated_zone(zone)) {
6599 node_set_state(nid, N_HIGH_MEMORY);
6600 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6601 zone_type <= ZONE_NORMAL)
6602 node_set_state(nid, N_NORMAL_MEMORY);
6609 * free_area_init_nodes - Initialise all pg_data_t and zone data
6610 * @max_zone_pfn: an array of max PFNs for each zone
6612 * This will call free_area_init_node() for each active node in the system.
6613 * Using the page ranges provided by memblock_set_node(), the size of each
6614 * zone in each node and their holes is calculated. If the maximum PFN
6615 * between two adjacent zones match, it is assumed that the zone is empty.
6616 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6617 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6618 * starts where the previous one ended. For example, ZONE_DMA32 starts
6619 * at arch_max_dma_pfn.
6621 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6623 unsigned long start_pfn, end_pfn;
6626 /* Record where the zone boundaries are */
6627 memset(arch_zone_lowest_possible_pfn, 0,
6628 sizeof(arch_zone_lowest_possible_pfn));
6629 memset(arch_zone_highest_possible_pfn, 0,
6630 sizeof(arch_zone_highest_possible_pfn));
6632 start_pfn = find_min_pfn_with_active_regions();
6634 for (i = 0; i < MAX_NR_ZONES; i++) {
6635 if (i == ZONE_MOVABLE)
6638 end_pfn = max(max_zone_pfn[i], start_pfn);
6639 arch_zone_lowest_possible_pfn[i] = start_pfn;
6640 arch_zone_highest_possible_pfn[i] = end_pfn;
6642 start_pfn = end_pfn;
6645 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6646 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6647 find_zone_movable_pfns_for_nodes();
6649 /* Print out the zone ranges */
6650 pr_info("Zone ranges:\n");
6651 for (i = 0; i < MAX_NR_ZONES; i++) {
6652 if (i == ZONE_MOVABLE)
6654 pr_info(" %-8s ", zone_names[i]);
6655 if (arch_zone_lowest_possible_pfn[i] ==
6656 arch_zone_highest_possible_pfn[i])
6659 pr_cont("[mem %#018Lx-%#018Lx]\n",
6660 (u64)arch_zone_lowest_possible_pfn[i]
6662 ((u64)arch_zone_highest_possible_pfn[i]
6663 << PAGE_SHIFT) - 1);
6666 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6667 pr_info("Movable zone start for each node\n");
6668 for (i = 0; i < MAX_NUMNODES; i++) {
6669 if (zone_movable_pfn[i])
6670 pr_info(" Node %d: %#018Lx\n", i,
6671 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6674 /* Print out the early node map */
6675 pr_info("Early memory node ranges\n");
6676 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6677 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6678 (u64)start_pfn << PAGE_SHIFT,
6679 ((u64)end_pfn << PAGE_SHIFT) - 1);
6681 /* Initialise every node */
6682 mminit_verify_pageflags_layout();
6683 setup_nr_node_ids();
6684 for_each_online_node(nid) {
6685 pg_data_t *pgdat = NODE_DATA(nid);
6686 free_area_init_node(nid, NULL,
6687 find_min_pfn_for_node(nid), NULL);
6689 /* Any memory on that node */
6690 if (pgdat->node_present_pages)
6691 node_set_state(nid, N_MEMORY);
6692 check_for_memory(pgdat, nid);
6696 static int __init cmdline_parse_core(char *p, unsigned long *core)
6698 unsigned long long coremem;
6702 coremem = memparse(p, &p);
6703 *core = coremem >> PAGE_SHIFT;
6705 /* Paranoid check that UL is enough for the coremem value */
6706 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6712 * kernelcore=size sets the amount of memory for use for allocations that
6713 * cannot be reclaimed or migrated.
6715 static int __init cmdline_parse_kernelcore(char *p)
6717 /* parse kernelcore=mirror */
6718 if (parse_option_str(p, "mirror")) {
6719 mirrored_kernelcore = true;
6723 return cmdline_parse_core(p, &required_kernelcore);
6727 * movablecore=size sets the amount of memory for use for allocations that
6728 * can be reclaimed or migrated.
6730 static int __init cmdline_parse_movablecore(char *p)
6732 return cmdline_parse_core(p, &required_movablecore);
6735 early_param("kernelcore", cmdline_parse_kernelcore);
6736 early_param("movablecore", cmdline_parse_movablecore);
6738 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6740 void adjust_managed_page_count(struct page *page, long count)
6742 spin_lock(&managed_page_count_lock);
6743 page_zone(page)->managed_pages += count;
6744 totalram_pages += count;
6745 #ifdef CONFIG_HIGHMEM
6746 if (PageHighMem(page))
6747 totalhigh_pages += count;
6749 spin_unlock(&managed_page_count_lock);
6751 EXPORT_SYMBOL(adjust_managed_page_count);
6753 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6756 unsigned long pages = 0;
6758 start = (void *)PAGE_ALIGN((unsigned long)start);
6759 end = (void *)((unsigned long)end & PAGE_MASK);
6760 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6761 if ((unsigned int)poison <= 0xFF)
6762 memset(pos, poison, PAGE_SIZE);
6763 free_reserved_page(virt_to_page(pos));
6767 pr_info("Freeing %s memory: %ldK\n",
6768 s, pages << (PAGE_SHIFT - 10));
6772 EXPORT_SYMBOL(free_reserved_area);
6774 #ifdef CONFIG_HIGHMEM
6775 void free_highmem_page(struct page *page)
6777 __free_reserved_page(page);
6779 page_zone(page)->managed_pages++;
6785 void __init mem_init_print_info(const char *str)
6787 unsigned long physpages, codesize, datasize, rosize, bss_size;
6788 unsigned long init_code_size, init_data_size;
6790 physpages = get_num_physpages();
6791 codesize = _etext - _stext;
6792 datasize = _edata - _sdata;
6793 rosize = __end_rodata - __start_rodata;
6794 bss_size = __bss_stop - __bss_start;
6795 init_data_size = __init_end - __init_begin;
6796 init_code_size = _einittext - _sinittext;
6799 * Detect special cases and adjust section sizes accordingly:
6800 * 1) .init.* may be embedded into .data sections
6801 * 2) .init.text.* may be out of [__init_begin, __init_end],
6802 * please refer to arch/tile/kernel/vmlinux.lds.S.
6803 * 3) .rodata.* may be embedded into .text or .data sections.
6805 #define adj_init_size(start, end, size, pos, adj) \
6807 if (start <= pos && pos < end && size > adj) \
6811 adj_init_size(__init_begin, __init_end, init_data_size,
6812 _sinittext, init_code_size);
6813 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6814 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6815 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6816 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6818 #undef adj_init_size
6820 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6821 #ifdef CONFIG_HIGHMEM
6825 nr_free_pages() << (PAGE_SHIFT - 10),
6826 physpages << (PAGE_SHIFT - 10),
6827 codesize >> 10, datasize >> 10, rosize >> 10,
6828 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6829 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6830 totalcma_pages << (PAGE_SHIFT - 10),
6831 #ifdef CONFIG_HIGHMEM
6832 totalhigh_pages << (PAGE_SHIFT - 10),
6834 str ? ", " : "", str ? str : "");
6838 * set_dma_reserve - set the specified number of pages reserved in the first zone
6839 * @new_dma_reserve: The number of pages to mark reserved
6841 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6842 * In the DMA zone, a significant percentage may be consumed by kernel image
6843 * and other unfreeable allocations which can skew the watermarks badly. This
6844 * function may optionally be used to account for unfreeable pages in the
6845 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6846 * smaller per-cpu batchsize.
6848 void __init set_dma_reserve(unsigned long new_dma_reserve)
6850 dma_reserve = new_dma_reserve;
6853 void __init free_area_init(unsigned long *zones_size)
6855 free_area_init_node(0, zones_size,
6856 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6859 static int page_alloc_cpu_dead(unsigned int cpu)
6862 lru_add_drain_cpu(cpu);
6866 * Spill the event counters of the dead processor
6867 * into the current processors event counters.
6868 * This artificially elevates the count of the current
6871 vm_events_fold_cpu(cpu);
6874 * Zero the differential counters of the dead processor
6875 * so that the vm statistics are consistent.
6877 * This is only okay since the processor is dead and cannot
6878 * race with what we are doing.
6880 cpu_vm_stats_fold(cpu);
6884 void __init page_alloc_init(void)
6888 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6889 "mm/page_alloc:dead", NULL,
6890 page_alloc_cpu_dead);
6895 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6896 * or min_free_kbytes changes.
6898 static void calculate_totalreserve_pages(void)
6900 struct pglist_data *pgdat;
6901 unsigned long reserve_pages = 0;
6902 enum zone_type i, j;
6904 for_each_online_pgdat(pgdat) {
6906 pgdat->totalreserve_pages = 0;
6908 for (i = 0; i < MAX_NR_ZONES; i++) {
6909 struct zone *zone = pgdat->node_zones + i;
6912 /* Find valid and maximum lowmem_reserve in the zone */
6913 for (j = i; j < MAX_NR_ZONES; j++) {
6914 if (zone->lowmem_reserve[j] > max)
6915 max = zone->lowmem_reserve[j];
6918 /* we treat the high watermark as reserved pages. */
6919 max += high_wmark_pages(zone);
6921 if (max > zone->managed_pages)
6922 max = zone->managed_pages;
6924 pgdat->totalreserve_pages += max;
6926 reserve_pages += max;
6929 totalreserve_pages = reserve_pages;
6933 * setup_per_zone_lowmem_reserve - called whenever
6934 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6935 * has a correct pages reserved value, so an adequate number of
6936 * pages are left in the zone after a successful __alloc_pages().
6938 static void setup_per_zone_lowmem_reserve(void)
6940 struct pglist_data *pgdat;
6941 enum zone_type j, idx;
6943 for_each_online_pgdat(pgdat) {
6944 for (j = 0; j < MAX_NR_ZONES; j++) {
6945 struct zone *zone = pgdat->node_zones + j;
6946 unsigned long managed_pages = zone->managed_pages;
6948 zone->lowmem_reserve[j] = 0;
6952 struct zone *lower_zone;
6956 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6957 sysctl_lowmem_reserve_ratio[idx] = 1;
6959 lower_zone = pgdat->node_zones + idx;
6960 lower_zone->lowmem_reserve[j] = managed_pages /
6961 sysctl_lowmem_reserve_ratio[idx];
6962 managed_pages += lower_zone->managed_pages;
6967 /* update totalreserve_pages */
6968 calculate_totalreserve_pages();
6971 static void __setup_per_zone_wmarks(void)
6973 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6974 unsigned long lowmem_pages = 0;
6976 unsigned long flags;
6978 /* Calculate total number of !ZONE_HIGHMEM pages */
6979 for_each_zone(zone) {
6980 if (!is_highmem(zone))
6981 lowmem_pages += zone->managed_pages;
6984 for_each_zone(zone) {
6987 spin_lock_irqsave(&zone->lock, flags);
6988 tmp = (u64)pages_min * zone->managed_pages;
6989 do_div(tmp, lowmem_pages);
6990 if (is_highmem(zone)) {
6992 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6993 * need highmem pages, so cap pages_min to a small
6996 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6997 * deltas control asynch page reclaim, and so should
6998 * not be capped for highmem.
7000 unsigned long min_pages;
7002 min_pages = zone->managed_pages / 1024;
7003 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7004 zone->watermark[WMARK_MIN] = min_pages;
7007 * If it's a lowmem zone, reserve a number of pages
7008 * proportionate to the zone's size.
7010 zone->watermark[WMARK_MIN] = tmp;
7014 * Set the kswapd watermarks distance according to the
7015 * scale factor in proportion to available memory, but
7016 * ensure a minimum size on small systems.
7018 tmp = max_t(u64, tmp >> 2,
7019 mult_frac(zone->managed_pages,
7020 watermark_scale_factor, 10000));
7022 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7023 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7025 spin_unlock_irqrestore(&zone->lock, flags);
7028 /* update totalreserve_pages */
7029 calculate_totalreserve_pages();
7033 * setup_per_zone_wmarks - called when min_free_kbytes changes
7034 * or when memory is hot-{added|removed}
7036 * Ensures that the watermark[min,low,high] values for each zone are set
7037 * correctly with respect to min_free_kbytes.
7039 void setup_per_zone_wmarks(void)
7041 mutex_lock(&zonelists_mutex);
7042 __setup_per_zone_wmarks();
7043 mutex_unlock(&zonelists_mutex);
7047 * Initialise min_free_kbytes.
7049 * For small machines we want it small (128k min). For large machines
7050 * we want it large (64MB max). But it is not linear, because network
7051 * bandwidth does not increase linearly with machine size. We use
7053 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7054 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7070 int __meminit init_per_zone_wmark_min(void)
7072 unsigned long lowmem_kbytes;
7073 int new_min_free_kbytes;
7075 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7076 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7078 if (new_min_free_kbytes > user_min_free_kbytes) {
7079 min_free_kbytes = new_min_free_kbytes;
7080 if (min_free_kbytes < 128)
7081 min_free_kbytes = 128;
7082 if (min_free_kbytes > 65536)
7083 min_free_kbytes = 65536;
7085 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7086 new_min_free_kbytes, user_min_free_kbytes);
7088 setup_per_zone_wmarks();
7089 refresh_zone_stat_thresholds();
7090 setup_per_zone_lowmem_reserve();
7093 setup_min_unmapped_ratio();
7094 setup_min_slab_ratio();
7099 core_initcall(init_per_zone_wmark_min)
7102 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7103 * that we can call two helper functions whenever min_free_kbytes
7106 int min_free_kbytes_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);
7116 user_min_free_kbytes = min_free_kbytes;
7117 setup_per_zone_wmarks();
7122 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7123 void __user *buffer, size_t *length, loff_t *ppos)
7127 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7132 setup_per_zone_wmarks();
7138 static void setup_min_unmapped_ratio(void)
7143 for_each_online_pgdat(pgdat)
7144 pgdat->min_unmapped_pages = 0;
7147 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7148 sysctl_min_unmapped_ratio) / 100;
7152 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7153 void __user *buffer, size_t *length, loff_t *ppos)
7157 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7161 setup_min_unmapped_ratio();
7166 static void setup_min_slab_ratio(void)
7171 for_each_online_pgdat(pgdat)
7172 pgdat->min_slab_pages = 0;
7175 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7176 sysctl_min_slab_ratio) / 100;
7179 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7180 void __user *buffer, size_t *length, loff_t *ppos)
7184 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7188 setup_min_slab_ratio();
7195 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7196 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7197 * whenever sysctl_lowmem_reserve_ratio changes.
7199 * The reserve ratio obviously has absolutely no relation with the
7200 * minimum watermarks. The lowmem reserve ratio can only make sense
7201 * if in function of the boot time zone sizes.
7203 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7204 void __user *buffer, size_t *length, loff_t *ppos)
7206 proc_dointvec_minmax(table, write, buffer, length, ppos);
7207 setup_per_zone_lowmem_reserve();
7212 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7213 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7214 * pagelist can have before it gets flushed back to buddy allocator.
7216 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7217 void __user *buffer, size_t *length, loff_t *ppos)
7220 int old_percpu_pagelist_fraction;
7223 mutex_lock(&pcp_batch_high_lock);
7224 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7226 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7227 if (!write || ret < 0)
7230 /* Sanity checking to avoid pcp imbalance */
7231 if (percpu_pagelist_fraction &&
7232 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7233 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7239 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7242 for_each_populated_zone(zone) {
7245 for_each_possible_cpu(cpu)
7246 pageset_set_high_and_batch(zone,
7247 per_cpu_ptr(zone->pageset, cpu));
7250 mutex_unlock(&pcp_batch_high_lock);
7255 int hashdist = HASHDIST_DEFAULT;
7257 static int __init set_hashdist(char *str)
7261 hashdist = simple_strtoul(str, &str, 0);
7264 __setup("hashdist=", set_hashdist);
7267 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7269 * Returns the number of pages that arch has reserved but
7270 * is not known to alloc_large_system_hash().
7272 static unsigned long __init arch_reserved_kernel_pages(void)
7279 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7280 * machines. As memory size is increased the scale is also increased but at
7281 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7282 * quadruples the scale is increased by one, which means the size of hash table
7283 * only doubles, instead of quadrupling as well.
7284 * Because 32-bit systems cannot have large physical memory, where this scaling
7285 * makes sense, it is disabled on such platforms.
7287 #if __BITS_PER_LONG > 32
7288 #define ADAPT_SCALE_BASE (64ul << 30)
7289 #define ADAPT_SCALE_SHIFT 2
7290 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7294 * allocate a large system hash table from bootmem
7295 * - it is assumed that the hash table must contain an exact power-of-2
7296 * quantity of entries
7297 * - limit is the number of hash buckets, not the total allocation size
7299 void *__init alloc_large_system_hash(const char *tablename,
7300 unsigned long bucketsize,
7301 unsigned long numentries,
7304 unsigned int *_hash_shift,
7305 unsigned int *_hash_mask,
7306 unsigned long low_limit,
7307 unsigned long high_limit)
7309 unsigned long long max = high_limit;
7310 unsigned long log2qty, size;
7314 /* allow the kernel cmdline to have a say */
7316 /* round applicable memory size up to nearest megabyte */
7317 numentries = nr_kernel_pages;
7318 numentries -= arch_reserved_kernel_pages();
7320 /* It isn't necessary when PAGE_SIZE >= 1MB */
7321 if (PAGE_SHIFT < 20)
7322 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7324 #if __BITS_PER_LONG > 32
7326 unsigned long adapt;
7328 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7329 adapt <<= ADAPT_SCALE_SHIFT)
7334 /* limit to 1 bucket per 2^scale bytes of low memory */
7335 if (scale > PAGE_SHIFT)
7336 numentries >>= (scale - PAGE_SHIFT);
7338 numentries <<= (PAGE_SHIFT - scale);
7340 /* Make sure we've got at least a 0-order allocation.. */
7341 if (unlikely(flags & HASH_SMALL)) {
7342 /* Makes no sense without HASH_EARLY */
7343 WARN_ON(!(flags & HASH_EARLY));
7344 if (!(numentries >> *_hash_shift)) {
7345 numentries = 1UL << *_hash_shift;
7346 BUG_ON(!numentries);
7348 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7349 numentries = PAGE_SIZE / bucketsize;
7351 numentries = roundup_pow_of_two(numentries);
7353 /* limit allocation size to 1/16 total memory by default */
7355 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7356 do_div(max, bucketsize);
7358 max = min(max, 0x80000000ULL);
7360 if (numentries < low_limit)
7361 numentries = low_limit;
7362 if (numentries > max)
7365 log2qty = ilog2(numentries);
7368 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7369 * currently not used when HASH_EARLY is specified.
7371 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7373 size = bucketsize << log2qty;
7374 if (flags & HASH_EARLY)
7375 table = memblock_virt_alloc_nopanic(size, 0);
7377 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7380 * If bucketsize is not a power-of-two, we may free
7381 * some pages at the end of hash table which
7382 * alloc_pages_exact() automatically does
7384 if (get_order(size) < MAX_ORDER) {
7385 table = alloc_pages_exact(size, gfp_flags);
7386 kmemleak_alloc(table, size, 1, gfp_flags);
7389 } while (!table && size > PAGE_SIZE && --log2qty);
7392 panic("Failed to allocate %s hash table\n", tablename);
7394 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7395 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7398 *_hash_shift = log2qty;
7400 *_hash_mask = (1 << log2qty) - 1;
7406 * This function checks whether pageblock includes unmovable pages or not.
7407 * If @count is not zero, it is okay to include less @count unmovable pages
7409 * PageLRU check without isolation or lru_lock could race so that
7410 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7411 * check without lock_page also may miss some movable non-lru pages at
7412 * race condition. So you can't expect this function should be exact.
7414 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7415 bool skip_hwpoisoned_pages)
7417 unsigned long pfn, iter, found;
7421 * For avoiding noise data, lru_add_drain_all() should be called
7422 * If ZONE_MOVABLE, the zone never contains unmovable pages
7424 if (zone_idx(zone) == ZONE_MOVABLE)
7426 mt = get_pageblock_migratetype(page);
7427 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7430 pfn = page_to_pfn(page);
7431 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7432 unsigned long check = pfn + iter;
7434 if (!pfn_valid_within(check))
7437 page = pfn_to_page(check);
7440 * Hugepages are not in LRU lists, but they're movable.
7441 * We need not scan over tail pages bacause we don't
7442 * handle each tail page individually in migration.
7444 if (PageHuge(page)) {
7445 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7450 * We can't use page_count without pin a page
7451 * because another CPU can free compound page.
7452 * This check already skips compound tails of THP
7453 * because their page->_refcount is zero at all time.
7455 if (!page_ref_count(page)) {
7456 if (PageBuddy(page))
7457 iter += (1 << page_order(page)) - 1;
7462 * The HWPoisoned page may be not in buddy system, and
7463 * page_count() is not 0.
7465 if (skip_hwpoisoned_pages && PageHWPoison(page))
7468 if (__PageMovable(page))
7474 * If there are RECLAIMABLE pages, we need to check
7475 * it. But now, memory offline itself doesn't call
7476 * shrink_node_slabs() and it still to be fixed.
7479 * If the page is not RAM, page_count()should be 0.
7480 * we don't need more check. This is an _used_ not-movable page.
7482 * The problematic thing here is PG_reserved pages. PG_reserved
7483 * is set to both of a memory hole page and a _used_ kernel
7492 bool is_pageblock_removable_nolock(struct page *page)
7498 * We have to be careful here because we are iterating over memory
7499 * sections which are not zone aware so we might end up outside of
7500 * the zone but still within the section.
7501 * We have to take care about the node as well. If the node is offline
7502 * its NODE_DATA will be NULL - see page_zone.
7504 if (!node_online(page_to_nid(page)))
7507 zone = page_zone(page);
7508 pfn = page_to_pfn(page);
7509 if (!zone_spans_pfn(zone, pfn))
7512 return !has_unmovable_pages(zone, page, 0, true);
7515 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7517 static unsigned long pfn_max_align_down(unsigned long pfn)
7519 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7520 pageblock_nr_pages) - 1);
7523 static unsigned long pfn_max_align_up(unsigned long pfn)
7525 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7526 pageblock_nr_pages));
7529 /* [start, end) must belong to a single zone. */
7530 static int __alloc_contig_migrate_range(struct compact_control *cc,
7531 unsigned long start, unsigned long end)
7533 /* This function is based on compact_zone() from compaction.c. */
7534 unsigned long nr_reclaimed;
7535 unsigned long pfn = start;
7536 unsigned int tries = 0;
7541 while (pfn < end || !list_empty(&cc->migratepages)) {
7542 if (fatal_signal_pending(current)) {
7547 if (list_empty(&cc->migratepages)) {
7548 cc->nr_migratepages = 0;
7549 pfn = isolate_migratepages_range(cc, pfn, end);
7555 } else if (++tries == 5) {
7556 ret = ret < 0 ? ret : -EBUSY;
7560 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7562 cc->nr_migratepages -= nr_reclaimed;
7564 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7565 NULL, 0, cc->mode, MR_CMA);
7568 putback_movable_pages(&cc->migratepages);
7575 * alloc_contig_range() -- tries to allocate given range of pages
7576 * @start: start PFN to allocate
7577 * @end: one-past-the-last PFN to allocate
7578 * @migratetype: migratetype of the underlaying pageblocks (either
7579 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7580 * in range must have the same migratetype and it must
7581 * be either of the two.
7582 * @gfp_mask: GFP mask to use during compaction
7584 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7585 * aligned, however it's the caller's responsibility to guarantee that
7586 * we are the only thread that changes migrate type of pageblocks the
7589 * The PFN range must belong to a single zone.
7591 * Returns zero on success or negative error code. On success all
7592 * pages which PFN is in [start, end) are allocated for the caller and
7593 * need to be freed with free_contig_range().
7595 int alloc_contig_range(unsigned long start, unsigned long end,
7596 unsigned migratetype, gfp_t gfp_mask)
7598 unsigned long outer_start, outer_end;
7602 struct compact_control cc = {
7603 .nr_migratepages = 0,
7605 .zone = page_zone(pfn_to_page(start)),
7606 .mode = MIGRATE_SYNC,
7607 .ignore_skip_hint = true,
7608 .gfp_mask = current_gfp_context(gfp_mask),
7610 INIT_LIST_HEAD(&cc.migratepages);
7613 * What we do here is we mark all pageblocks in range as
7614 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7615 * have different sizes, and due to the way page allocator
7616 * work, we align the range to biggest of the two pages so
7617 * that page allocator won't try to merge buddies from
7618 * different pageblocks and change MIGRATE_ISOLATE to some
7619 * other migration type.
7621 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7622 * migrate the pages from an unaligned range (ie. pages that
7623 * we are interested in). This will put all the pages in
7624 * range back to page allocator as MIGRATE_ISOLATE.
7626 * When this is done, we take the pages in range from page
7627 * allocator removing them from the buddy system. This way
7628 * page allocator will never consider using them.
7630 * This lets us mark the pageblocks back as
7631 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7632 * aligned range but not in the unaligned, original range are
7633 * put back to page allocator so that buddy can use them.
7636 ret = start_isolate_page_range(pfn_max_align_down(start),
7637 pfn_max_align_up(end), migratetype,
7643 * In case of -EBUSY, we'd like to know which page causes problem.
7644 * So, just fall through. We will check it in test_pages_isolated().
7646 ret = __alloc_contig_migrate_range(&cc, start, end);
7647 if (ret && ret != -EBUSY)
7651 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7652 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7653 * more, all pages in [start, end) are free in page allocator.
7654 * What we are going to do is to allocate all pages from
7655 * [start, end) (that is remove them from page allocator).
7657 * The only problem is that pages at the beginning and at the
7658 * end of interesting range may be not aligned with pages that
7659 * page allocator holds, ie. they can be part of higher order
7660 * pages. Because of this, we reserve the bigger range and
7661 * once this is done free the pages we are not interested in.
7663 * We don't have to hold zone->lock here because the pages are
7664 * isolated thus they won't get removed from buddy.
7667 lru_add_drain_all();
7668 drain_all_pages(cc.zone);
7671 outer_start = start;
7672 while (!PageBuddy(pfn_to_page(outer_start))) {
7673 if (++order >= MAX_ORDER) {
7674 outer_start = start;
7677 outer_start &= ~0UL << order;
7680 if (outer_start != start) {
7681 order = page_order(pfn_to_page(outer_start));
7684 * outer_start page could be small order buddy page and
7685 * it doesn't include start page. Adjust outer_start
7686 * in this case to report failed page properly
7687 * on tracepoint in test_pages_isolated()
7689 if (outer_start + (1UL << order) <= start)
7690 outer_start = start;
7693 /* Make sure the range is really isolated. */
7694 if (test_pages_isolated(outer_start, end, false)) {
7695 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7696 __func__, outer_start, end);
7701 /* Grab isolated pages from freelists. */
7702 outer_end = isolate_freepages_range(&cc, outer_start, end);
7708 /* Free head and tail (if any) */
7709 if (start != outer_start)
7710 free_contig_range(outer_start, start - outer_start);
7711 if (end != outer_end)
7712 free_contig_range(end, outer_end - end);
7715 undo_isolate_page_range(pfn_max_align_down(start),
7716 pfn_max_align_up(end), migratetype);
7720 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7722 unsigned int count = 0;
7724 for (; nr_pages--; pfn++) {
7725 struct page *page = pfn_to_page(pfn);
7727 count += page_count(page) != 1;
7730 WARN(count != 0, "%d pages are still in use!\n", count);
7734 #ifdef CONFIG_MEMORY_HOTPLUG
7736 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7737 * page high values need to be recalulated.
7739 void __meminit zone_pcp_update(struct zone *zone)
7742 mutex_lock(&pcp_batch_high_lock);
7743 for_each_possible_cpu(cpu)
7744 pageset_set_high_and_batch(zone,
7745 per_cpu_ptr(zone->pageset, cpu));
7746 mutex_unlock(&pcp_batch_high_lock);
7750 void zone_pcp_reset(struct zone *zone)
7752 unsigned long flags;
7754 struct per_cpu_pageset *pset;
7756 /* avoid races with drain_pages() */
7757 local_irq_save(flags);
7758 if (zone->pageset != &boot_pageset) {
7759 for_each_online_cpu(cpu) {
7760 pset = per_cpu_ptr(zone->pageset, cpu);
7761 drain_zonestat(zone, pset);
7763 free_percpu(zone->pageset);
7764 zone->pageset = &boot_pageset;
7766 local_irq_restore(flags);
7769 #ifdef CONFIG_MEMORY_HOTREMOVE
7771 * All pages in the range must be in a single zone and isolated
7772 * before calling this.
7775 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7779 unsigned int order, i;
7781 unsigned long flags;
7782 /* find the first valid pfn */
7783 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7788 offline_mem_sections(pfn, end_pfn);
7789 zone = page_zone(pfn_to_page(pfn));
7790 spin_lock_irqsave(&zone->lock, flags);
7792 while (pfn < end_pfn) {
7793 if (!pfn_valid(pfn)) {
7797 page = pfn_to_page(pfn);
7799 * The HWPoisoned page may be not in buddy system, and
7800 * page_count() is not 0.
7802 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7804 SetPageReserved(page);
7808 BUG_ON(page_count(page));
7809 BUG_ON(!PageBuddy(page));
7810 order = page_order(page);
7811 #ifdef CONFIG_DEBUG_VM
7812 pr_info("remove from free list %lx %d %lx\n",
7813 pfn, 1 << order, end_pfn);
7815 list_del(&page->lru);
7816 rmv_page_order(page);
7817 zone->free_area[order].nr_free--;
7818 for (i = 0; i < (1 << order); i++)
7819 SetPageReserved((page+i));
7820 pfn += (1 << order);
7822 spin_unlock_irqrestore(&zone->lock, flags);
7826 bool is_free_buddy_page(struct page *page)
7828 struct zone *zone = page_zone(page);
7829 unsigned long pfn = page_to_pfn(page);
7830 unsigned long flags;
7833 spin_lock_irqsave(&zone->lock, flags);
7834 for (order = 0; order < MAX_ORDER; order++) {
7835 struct page *page_head = page - (pfn & ((1 << order) - 1));
7837 if (PageBuddy(page_head) && page_order(page_head) >= order)
7840 spin_unlock_irqrestore(&zone->lock, flags);
7842 return order < MAX_ORDER;