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/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node);
79 EXPORT_PER_CPU_SYMBOL(numa_node);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91 int _node_numa_mem_[MAX_NUMNODES];
94 /* work_structs for global per-cpu drains */
95 DEFINE_MUTEX(pcpu_drain_mutex);
96 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
98 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
99 volatile unsigned long latent_entropy __latent_entropy;
100 EXPORT_SYMBOL(latent_entropy);
104 * Array of node states.
106 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
107 [N_POSSIBLE] = NODE_MASK_ALL,
108 [N_ONLINE] = { { [0] = 1UL } },
110 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
111 #ifdef CONFIG_HIGHMEM
112 [N_HIGH_MEMORY] = { { [0] = 1UL } },
114 #ifdef CONFIG_MOVABLE_NODE
115 [N_MEMORY] = { { [0] = 1UL } },
117 [N_CPU] = { { [0] = 1UL } },
120 EXPORT_SYMBOL(node_states);
122 /* Protect totalram_pages and zone->managed_pages */
123 static DEFINE_SPINLOCK(managed_page_count_lock);
125 unsigned long totalram_pages __read_mostly;
126 unsigned long totalreserve_pages __read_mostly;
127 unsigned long totalcma_pages __read_mostly;
129 int percpu_pagelist_fraction;
130 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
133 * A cached value of the page's pageblock's migratetype, used when the page is
134 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
135 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
136 * Also the migratetype set in the page does not necessarily match the pcplist
137 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
138 * other index - this ensures that it will be put on the correct CMA freelist.
140 static inline int get_pcppage_migratetype(struct page *page)
145 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
147 page->index = migratetype;
150 #ifdef CONFIG_PM_SLEEP
152 * The following functions are used by the suspend/hibernate code to temporarily
153 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
154 * while devices are suspended. To avoid races with the suspend/hibernate code,
155 * they should always be called with pm_mutex held (gfp_allowed_mask also should
156 * only be modified with pm_mutex held, unless the suspend/hibernate code is
157 * guaranteed not to run in parallel with that modification).
160 static gfp_t saved_gfp_mask;
162 void pm_restore_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex));
165 if (saved_gfp_mask) {
166 gfp_allowed_mask = saved_gfp_mask;
171 void pm_restrict_gfp_mask(void)
173 WARN_ON(!mutex_is_locked(&pm_mutex));
174 WARN_ON(saved_gfp_mask);
175 saved_gfp_mask = gfp_allowed_mask;
176 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
179 bool pm_suspended_storage(void)
181 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 #endif /* CONFIG_PM_SLEEP */
187 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
188 unsigned int pageblock_order __read_mostly;
191 static void __free_pages_ok(struct page *page, unsigned int order);
194 * results with 256, 32 in the lowmem_reserve sysctl:
195 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
196 * 1G machine -> (16M dma, 784M normal, 224M high)
197 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
198 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
199 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
201 * TBD: should special case ZONE_DMA32 machines here - in those we normally
202 * don't need any ZONE_NORMAL reservation
204 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
205 #ifdef CONFIG_ZONE_DMA
208 #ifdef CONFIG_ZONE_DMA32
211 #ifdef CONFIG_HIGHMEM
217 EXPORT_SYMBOL(totalram_pages);
219 static char * const zone_names[MAX_NR_ZONES] = {
220 #ifdef CONFIG_ZONE_DMA
223 #ifdef CONFIG_ZONE_DMA32
227 #ifdef CONFIG_HIGHMEM
231 #ifdef CONFIG_ZONE_DEVICE
236 char * const migratetype_names[MIGRATE_TYPES] = {
244 #ifdef CONFIG_MEMORY_ISOLATION
249 compound_page_dtor * const compound_page_dtors[] = {
252 #ifdef CONFIG_HUGETLB_PAGE
255 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 int min_free_kbytes = 1024;
261 int user_min_free_kbytes = -1;
262 int watermark_scale_factor = 10;
264 static unsigned long __meminitdata nr_kernel_pages;
265 static unsigned long __meminitdata nr_all_pages;
266 static unsigned long __meminitdata dma_reserve;
268 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
269 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
270 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
271 static unsigned long __initdata required_kernelcore;
272 static unsigned long __initdata required_movablecore;
273 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
274 static bool mirrored_kernelcore;
276 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
278 EXPORT_SYMBOL(movable_zone);
279 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
282 int nr_node_ids __read_mostly = MAX_NUMNODES;
283 int nr_online_nodes __read_mostly = 1;
284 EXPORT_SYMBOL(nr_node_ids);
285 EXPORT_SYMBOL(nr_online_nodes);
288 int page_group_by_mobility_disabled __read_mostly;
290 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
291 static inline void reset_deferred_meminit(pg_data_t *pgdat)
293 pgdat->first_deferred_pfn = ULONG_MAX;
296 /* Returns true if the struct page for the pfn is uninitialised */
297 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
299 int nid = early_pfn_to_nid(pfn);
301 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
308 * Returns false when the remaining initialisation should be deferred until
309 * later in the boot cycle when it can be parallelised.
311 static inline bool update_defer_init(pg_data_t *pgdat,
312 unsigned long pfn, unsigned long zone_end,
313 unsigned long *nr_initialised)
315 unsigned long max_initialise;
317 /* Always populate low zones for address-contrained allocations */
318 if (zone_end < pgdat_end_pfn(pgdat))
321 * Initialise at least 2G of a node but also take into account that
322 * two large system hashes that can take up 1GB for 0.25TB/node.
324 max_initialise = max(2UL << (30 - PAGE_SHIFT),
325 (pgdat->node_spanned_pages >> 8));
328 if ((*nr_initialised > max_initialise) &&
329 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
330 pgdat->first_deferred_pfn = pfn;
337 static inline void reset_deferred_meminit(pg_data_t *pgdat)
341 static inline bool early_page_uninitialised(unsigned long pfn)
346 static inline bool update_defer_init(pg_data_t *pgdat,
347 unsigned long pfn, unsigned long zone_end,
348 unsigned long *nr_initialised)
354 /* Return a pointer to the bitmap storing bits affecting a block of pages */
355 static inline unsigned long *get_pageblock_bitmap(struct page *page,
358 #ifdef CONFIG_SPARSEMEM
359 return __pfn_to_section(pfn)->pageblock_flags;
361 return page_zone(page)->pageblock_flags;
362 #endif /* CONFIG_SPARSEMEM */
365 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
367 #ifdef CONFIG_SPARSEMEM
368 pfn &= (PAGES_PER_SECTION-1);
369 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
371 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
372 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
373 #endif /* CONFIG_SPARSEMEM */
377 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
378 * @page: The page within the block of interest
379 * @pfn: The target page frame number
380 * @end_bitidx: The last bit of interest to retrieve
381 * @mask: mask of bits that the caller is interested in
383 * Return: pageblock_bits flags
385 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
387 unsigned long end_bitidx,
390 unsigned long *bitmap;
391 unsigned long bitidx, word_bitidx;
394 bitmap = get_pageblock_bitmap(page, pfn);
395 bitidx = pfn_to_bitidx(page, pfn);
396 word_bitidx = bitidx / BITS_PER_LONG;
397 bitidx &= (BITS_PER_LONG-1);
399 word = bitmap[word_bitidx];
400 bitidx += end_bitidx;
401 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
404 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
405 unsigned long end_bitidx,
408 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
411 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
413 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
417 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
418 * @page: The page within the block of interest
419 * @flags: The flags to set
420 * @pfn: The target page frame number
421 * @end_bitidx: The last bit of interest
422 * @mask: mask of bits that the caller is interested in
424 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
426 unsigned long end_bitidx,
429 unsigned long *bitmap;
430 unsigned long bitidx, word_bitidx;
431 unsigned long old_word, word;
433 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
435 bitmap = get_pageblock_bitmap(page, pfn);
436 bitidx = pfn_to_bitidx(page, pfn);
437 word_bitidx = bitidx / BITS_PER_LONG;
438 bitidx &= (BITS_PER_LONG-1);
440 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
442 bitidx += end_bitidx;
443 mask <<= (BITS_PER_LONG - bitidx - 1);
444 flags <<= (BITS_PER_LONG - bitidx - 1);
446 word = READ_ONCE(bitmap[word_bitidx]);
448 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
449 if (word == old_word)
455 void set_pageblock_migratetype(struct page *page, int migratetype)
457 if (unlikely(page_group_by_mobility_disabled &&
458 migratetype < MIGRATE_PCPTYPES))
459 migratetype = MIGRATE_UNMOVABLE;
461 set_pageblock_flags_group(page, (unsigned long)migratetype,
462 PB_migrate, PB_migrate_end);
465 #ifdef CONFIG_DEBUG_VM
466 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
470 unsigned long pfn = page_to_pfn(page);
471 unsigned long sp, start_pfn;
474 seq = zone_span_seqbegin(zone);
475 start_pfn = zone->zone_start_pfn;
476 sp = zone->spanned_pages;
477 if (!zone_spans_pfn(zone, pfn))
479 } while (zone_span_seqretry(zone, seq));
482 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
483 pfn, zone_to_nid(zone), zone->name,
484 start_pfn, start_pfn + sp);
489 static int page_is_consistent(struct zone *zone, struct page *page)
491 if (!pfn_valid_within(page_to_pfn(page)))
493 if (zone != page_zone(page))
499 * Temporary debugging check for pages not lying within a given zone.
501 static int bad_range(struct zone *zone, struct page *page)
503 if (page_outside_zone_boundaries(zone, page))
505 if (!page_is_consistent(zone, page))
511 static inline int bad_range(struct zone *zone, struct page *page)
517 static void bad_page(struct page *page, const char *reason,
518 unsigned long bad_flags)
520 static unsigned long resume;
521 static unsigned long nr_shown;
522 static unsigned long nr_unshown;
525 * Allow a burst of 60 reports, then keep quiet for that minute;
526 * or allow a steady drip of one report per second.
528 if (nr_shown == 60) {
529 if (time_before(jiffies, resume)) {
535 "BUG: Bad page state: %lu messages suppressed\n",
542 resume = jiffies + 60 * HZ;
544 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
545 current->comm, page_to_pfn(page));
546 __dump_page(page, reason);
547 bad_flags &= page->flags;
549 pr_alert("bad because of flags: %#lx(%pGp)\n",
550 bad_flags, &bad_flags);
551 dump_page_owner(page);
556 /* Leave bad fields for debug, except PageBuddy could make trouble */
557 page_mapcount_reset(page); /* remove PageBuddy */
558 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
562 * Higher-order pages are called "compound pages". They are structured thusly:
564 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
566 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
567 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
569 * The first tail page's ->compound_dtor holds the offset in array of compound
570 * page destructors. See compound_page_dtors.
572 * The first tail page's ->compound_order holds the order of allocation.
573 * This usage means that zero-order pages may not be compound.
576 void free_compound_page(struct page *page)
578 __free_pages_ok(page, compound_order(page));
581 void prep_compound_page(struct page *page, unsigned int order)
584 int nr_pages = 1 << order;
586 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
587 set_compound_order(page, order);
589 for (i = 1; i < nr_pages; i++) {
590 struct page *p = page + i;
591 set_page_count(p, 0);
592 p->mapping = TAIL_MAPPING;
593 set_compound_head(p, page);
595 atomic_set(compound_mapcount_ptr(page), -1);
598 #ifdef CONFIG_DEBUG_PAGEALLOC
599 unsigned int _debug_guardpage_minorder;
600 bool _debug_pagealloc_enabled __read_mostly
601 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
602 EXPORT_SYMBOL(_debug_pagealloc_enabled);
603 bool _debug_guardpage_enabled __read_mostly;
605 static int __init early_debug_pagealloc(char *buf)
609 return kstrtobool(buf, &_debug_pagealloc_enabled);
611 early_param("debug_pagealloc", early_debug_pagealloc);
613 static bool need_debug_guardpage(void)
615 /* If we don't use debug_pagealloc, we don't need guard page */
616 if (!debug_pagealloc_enabled())
619 if (!debug_guardpage_minorder())
625 static void init_debug_guardpage(void)
627 if (!debug_pagealloc_enabled())
630 if (!debug_guardpage_minorder())
633 _debug_guardpage_enabled = true;
636 struct page_ext_operations debug_guardpage_ops = {
637 .need = need_debug_guardpage,
638 .init = init_debug_guardpage,
641 static int __init debug_guardpage_minorder_setup(char *buf)
645 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
646 pr_err("Bad debug_guardpage_minorder value\n");
649 _debug_guardpage_minorder = res;
650 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
653 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
655 static inline bool set_page_guard(struct zone *zone, struct page *page,
656 unsigned int order, int migratetype)
658 struct page_ext *page_ext;
660 if (!debug_guardpage_enabled())
663 if (order >= debug_guardpage_minorder())
666 page_ext = lookup_page_ext(page);
667 if (unlikely(!page_ext))
670 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
672 INIT_LIST_HEAD(&page->lru);
673 set_page_private(page, order);
674 /* Guard pages are not available for any usage */
675 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
680 static inline void clear_page_guard(struct zone *zone, struct page *page,
681 unsigned int order, int migratetype)
683 struct page_ext *page_ext;
685 if (!debug_guardpage_enabled())
688 page_ext = lookup_page_ext(page);
689 if (unlikely(!page_ext))
692 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
694 set_page_private(page, 0);
695 if (!is_migrate_isolate(migratetype))
696 __mod_zone_freepage_state(zone, (1 << order), migratetype);
699 struct page_ext_operations debug_guardpage_ops;
700 static inline bool set_page_guard(struct zone *zone, struct page *page,
701 unsigned int order, int migratetype) { return false; }
702 static inline void clear_page_guard(struct zone *zone, struct page *page,
703 unsigned int order, int migratetype) {}
706 static inline void set_page_order(struct page *page, unsigned int order)
708 set_page_private(page, order);
709 __SetPageBuddy(page);
712 static inline void rmv_page_order(struct page *page)
714 __ClearPageBuddy(page);
715 set_page_private(page, 0);
719 * This function checks whether a page is free && is the buddy
720 * we can do coalesce a page and its buddy if
721 * (a) the buddy is not in a hole (check before calling!) &&
722 * (b) the buddy is in the buddy system &&
723 * (c) a page and its buddy have the same order &&
724 * (d) a page and its buddy are in the same zone.
726 * For recording whether a page is in the buddy system, we set ->_mapcount
727 * PAGE_BUDDY_MAPCOUNT_VALUE.
728 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
729 * serialized by zone->lock.
731 * For recording page's order, we use page_private(page).
733 static inline int page_is_buddy(struct page *page, struct page *buddy,
736 if (page_is_guard(buddy) && page_order(buddy) == order) {
737 if (page_zone_id(page) != page_zone_id(buddy))
740 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
745 if (PageBuddy(buddy) && page_order(buddy) == order) {
747 * zone check is done late to avoid uselessly
748 * calculating zone/node ids for pages that could
751 if (page_zone_id(page) != page_zone_id(buddy))
754 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
762 * Freeing function for a buddy system allocator.
764 * The concept of a buddy system is to maintain direct-mapped table
765 * (containing bit values) for memory blocks of various "orders".
766 * The bottom level table contains the map for the smallest allocatable
767 * units of memory (here, pages), and each level above it describes
768 * pairs of units from the levels below, hence, "buddies".
769 * At a high level, all that happens here is marking the table entry
770 * at the bottom level available, and propagating the changes upward
771 * as necessary, plus some accounting needed to play nicely with other
772 * parts of the VM system.
773 * At each level, we keep a list of pages, which are heads of continuous
774 * free pages of length of (1 << order) and marked with _mapcount
775 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
777 * So when we are allocating or freeing one, we can derive the state of the
778 * other. That is, if we allocate a small block, and both were
779 * free, the remainder of the region must be split into blocks.
780 * If a block is freed, and its buddy is also free, then this
781 * triggers coalescing into a block of larger size.
786 static inline void __free_one_page(struct page *page,
788 struct zone *zone, unsigned int order,
791 unsigned long combined_pfn;
792 unsigned long uninitialized_var(buddy_pfn);
794 unsigned int max_order;
796 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
798 VM_BUG_ON(!zone_is_initialized(zone));
799 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
801 VM_BUG_ON(migratetype == -1);
802 if (likely(!is_migrate_isolate(migratetype)))
803 __mod_zone_freepage_state(zone, 1 << order, migratetype);
805 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
806 VM_BUG_ON_PAGE(bad_range(zone, page), page);
809 while (order < max_order - 1) {
810 buddy_pfn = __find_buddy_pfn(pfn, order);
811 buddy = page + (buddy_pfn - pfn);
813 if (!pfn_valid_within(buddy_pfn))
815 if (!page_is_buddy(page, buddy, order))
818 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
819 * merge with it and move up one order.
821 if (page_is_guard(buddy)) {
822 clear_page_guard(zone, buddy, order, migratetype);
824 list_del(&buddy->lru);
825 zone->free_area[order].nr_free--;
826 rmv_page_order(buddy);
828 combined_pfn = buddy_pfn & pfn;
829 page = page + (combined_pfn - pfn);
833 if (max_order < MAX_ORDER) {
834 /* If we are here, it means order is >= pageblock_order.
835 * We want to prevent merge between freepages on isolate
836 * pageblock and normal pageblock. Without this, pageblock
837 * isolation could cause incorrect freepage or CMA accounting.
839 * We don't want to hit this code for the more frequent
842 if (unlikely(has_isolate_pageblock(zone))) {
845 buddy_pfn = __find_buddy_pfn(pfn, order);
846 buddy = page + (buddy_pfn - pfn);
847 buddy_mt = get_pageblock_migratetype(buddy);
849 if (migratetype != buddy_mt
850 && (is_migrate_isolate(migratetype) ||
851 is_migrate_isolate(buddy_mt)))
855 goto continue_merging;
859 set_page_order(page, order);
862 * If this is not the largest possible page, check if the buddy
863 * of the next-highest order is free. If it is, it's possible
864 * that pages are being freed that will coalesce soon. In case,
865 * that is happening, add the free page to the tail of the list
866 * so it's less likely to be used soon and more likely to be merged
867 * as a higher order page
869 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
870 struct page *higher_page, *higher_buddy;
871 combined_pfn = buddy_pfn & pfn;
872 higher_page = page + (combined_pfn - pfn);
873 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
874 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
875 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
876 list_add_tail(&page->lru,
877 &zone->free_area[order].free_list[migratetype]);
882 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
884 zone->free_area[order].nr_free++;
888 * A bad page could be due to a number of fields. Instead of multiple branches,
889 * try and check multiple fields with one check. The caller must do a detailed
890 * check if necessary.
892 static inline bool page_expected_state(struct page *page,
893 unsigned long check_flags)
895 if (unlikely(atomic_read(&page->_mapcount) != -1))
898 if (unlikely((unsigned long)page->mapping |
899 page_ref_count(page) |
901 (unsigned long)page->mem_cgroup |
903 (page->flags & check_flags)))
909 static void free_pages_check_bad(struct page *page)
911 const char *bad_reason;
912 unsigned long bad_flags;
917 if (unlikely(atomic_read(&page->_mapcount) != -1))
918 bad_reason = "nonzero mapcount";
919 if (unlikely(page->mapping != NULL))
920 bad_reason = "non-NULL mapping";
921 if (unlikely(page_ref_count(page) != 0))
922 bad_reason = "nonzero _refcount";
923 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
924 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
925 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
928 if (unlikely(page->mem_cgroup))
929 bad_reason = "page still charged to cgroup";
931 bad_page(page, bad_reason, bad_flags);
934 static inline int free_pages_check(struct page *page)
936 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
939 /* Something has gone sideways, find it */
940 free_pages_check_bad(page);
944 static int free_tail_pages_check(struct page *head_page, struct page *page)
949 * We rely page->lru.next never has bit 0 set, unless the page
950 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
952 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
954 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
958 switch (page - head_page) {
960 /* the first tail page: ->mapping is compound_mapcount() */
961 if (unlikely(compound_mapcount(page))) {
962 bad_page(page, "nonzero compound_mapcount", 0);
968 * the second tail page: ->mapping is
969 * page_deferred_list().next -- ignore value.
973 if (page->mapping != TAIL_MAPPING) {
974 bad_page(page, "corrupted mapping in tail page", 0);
979 if (unlikely(!PageTail(page))) {
980 bad_page(page, "PageTail not set", 0);
983 if (unlikely(compound_head(page) != head_page)) {
984 bad_page(page, "compound_head not consistent", 0);
989 page->mapping = NULL;
990 clear_compound_head(page);
994 static __always_inline bool free_pages_prepare(struct page *page,
995 unsigned int order, bool check_free)
999 VM_BUG_ON_PAGE(PageTail(page), page);
1001 trace_mm_page_free(page, order);
1002 kmemcheck_free_shadow(page, order);
1005 * Check tail pages before head page information is cleared to
1006 * avoid checking PageCompound for order-0 pages.
1008 if (unlikely(order)) {
1009 bool compound = PageCompound(page);
1012 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1015 ClearPageDoubleMap(page);
1016 for (i = 1; i < (1 << order); i++) {
1018 bad += free_tail_pages_check(page, page + i);
1019 if (unlikely(free_pages_check(page + i))) {
1023 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1026 if (PageMappingFlags(page))
1027 page->mapping = NULL;
1028 if (memcg_kmem_enabled() && PageKmemcg(page))
1029 memcg_kmem_uncharge(page, order);
1031 bad += free_pages_check(page);
1035 page_cpupid_reset_last(page);
1036 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1037 reset_page_owner(page, order);
1039 if (!PageHighMem(page)) {
1040 debug_check_no_locks_freed(page_address(page),
1041 PAGE_SIZE << order);
1042 debug_check_no_obj_freed(page_address(page),
1043 PAGE_SIZE << order);
1045 arch_free_page(page, order);
1046 kernel_poison_pages(page, 1 << order, 0);
1047 kernel_map_pages(page, 1 << order, 0);
1048 kasan_free_pages(page, order);
1053 #ifdef CONFIG_DEBUG_VM
1054 static inline bool free_pcp_prepare(struct page *page)
1056 return free_pages_prepare(page, 0, true);
1059 static inline bool bulkfree_pcp_prepare(struct page *page)
1064 static bool free_pcp_prepare(struct page *page)
1066 return free_pages_prepare(page, 0, false);
1069 static bool bulkfree_pcp_prepare(struct page *page)
1071 return free_pages_check(page);
1073 #endif /* CONFIG_DEBUG_VM */
1076 * Frees a number of pages from the PCP lists
1077 * Assumes all pages on list are in same zone, and of same order.
1078 * count is the number of pages to free.
1080 * If the zone was previously in an "all pages pinned" state then look to
1081 * see if this freeing clears that state.
1083 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1084 * pinned" detection logic.
1086 static void free_pcppages_bulk(struct zone *zone, int count,
1087 struct per_cpu_pages *pcp)
1089 int migratetype = 0;
1091 unsigned long nr_scanned, flags;
1092 bool isolated_pageblocks;
1094 spin_lock_irqsave(&zone->lock, flags);
1095 isolated_pageblocks = has_isolate_pageblock(zone);
1096 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1098 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1102 struct list_head *list;
1105 * Remove pages from lists in a round-robin fashion. A
1106 * batch_free count is maintained that is incremented when an
1107 * empty list is encountered. This is so more pages are freed
1108 * off fuller lists instead of spinning excessively around empty
1113 if (++migratetype == MIGRATE_PCPTYPES)
1115 list = &pcp->lists[migratetype];
1116 } while (list_empty(list));
1118 /* This is the only non-empty list. Free them all. */
1119 if (batch_free == MIGRATE_PCPTYPES)
1123 int mt; /* migratetype of the to-be-freed page */
1125 page = list_last_entry(list, struct page, lru);
1126 /* must delete as __free_one_page list manipulates */
1127 list_del(&page->lru);
1129 mt = get_pcppage_migratetype(page);
1130 /* MIGRATE_ISOLATE page should not go to pcplists */
1131 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1132 /* Pageblock could have been isolated meanwhile */
1133 if (unlikely(isolated_pageblocks))
1134 mt = get_pageblock_migratetype(page);
1136 if (bulkfree_pcp_prepare(page))
1139 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1140 trace_mm_page_pcpu_drain(page, 0, mt);
1141 } while (--count && --batch_free && !list_empty(list));
1143 spin_unlock_irqrestore(&zone->lock, flags);
1146 static void free_one_page(struct zone *zone,
1147 struct page *page, unsigned long pfn,
1151 unsigned long nr_scanned, flags;
1152 spin_lock_irqsave(&zone->lock, flags);
1153 __count_vm_events(PGFREE, 1 << order);
1154 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1156 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1158 if (unlikely(has_isolate_pageblock(zone) ||
1159 is_migrate_isolate(migratetype))) {
1160 migratetype = get_pfnblock_migratetype(page, pfn);
1162 __free_one_page(page, pfn, zone, order, migratetype);
1163 spin_unlock_irqrestore(&zone->lock, flags);
1166 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1167 unsigned long zone, int nid)
1169 set_page_links(page, zone, nid, pfn);
1170 init_page_count(page);
1171 page_mapcount_reset(page);
1172 page_cpupid_reset_last(page);
1174 INIT_LIST_HEAD(&page->lru);
1175 #ifdef WANT_PAGE_VIRTUAL
1176 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1177 if (!is_highmem_idx(zone))
1178 set_page_address(page, __va(pfn << PAGE_SHIFT));
1182 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1185 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1188 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1189 static void init_reserved_page(unsigned long pfn)
1194 if (!early_page_uninitialised(pfn))
1197 nid = early_pfn_to_nid(pfn);
1198 pgdat = NODE_DATA(nid);
1200 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1201 struct zone *zone = &pgdat->node_zones[zid];
1203 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1206 __init_single_pfn(pfn, zid, nid);
1209 static inline void init_reserved_page(unsigned long pfn)
1212 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1215 * Initialised pages do not have PageReserved set. This function is
1216 * called for each range allocated by the bootmem allocator and
1217 * marks the pages PageReserved. The remaining valid pages are later
1218 * sent to the buddy page allocator.
1220 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1222 unsigned long start_pfn = PFN_DOWN(start);
1223 unsigned long end_pfn = PFN_UP(end);
1225 for (; start_pfn < end_pfn; start_pfn++) {
1226 if (pfn_valid(start_pfn)) {
1227 struct page *page = pfn_to_page(start_pfn);
1229 init_reserved_page(start_pfn);
1231 /* Avoid false-positive PageTail() */
1232 INIT_LIST_HEAD(&page->lru);
1234 SetPageReserved(page);
1239 static void __free_pages_ok(struct page *page, unsigned int order)
1242 unsigned long pfn = page_to_pfn(page);
1244 if (!free_pages_prepare(page, order, true))
1247 migratetype = get_pfnblock_migratetype(page, pfn);
1248 free_one_page(page_zone(page), page, pfn, order, migratetype);
1251 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1253 unsigned int nr_pages = 1 << order;
1254 struct page *p = page;
1258 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1260 __ClearPageReserved(p);
1261 set_page_count(p, 0);
1263 __ClearPageReserved(p);
1264 set_page_count(p, 0);
1266 page_zone(page)->managed_pages += nr_pages;
1267 set_page_refcounted(page);
1268 __free_pages(page, order);
1271 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1272 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1274 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1276 int __meminit early_pfn_to_nid(unsigned long pfn)
1278 static DEFINE_SPINLOCK(early_pfn_lock);
1281 spin_lock(&early_pfn_lock);
1282 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1284 nid = first_online_node;
1285 spin_unlock(&early_pfn_lock);
1291 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1292 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1293 struct mminit_pfnnid_cache *state)
1297 nid = __early_pfn_to_nid(pfn, state);
1298 if (nid >= 0 && nid != node)
1303 /* Only safe to use early in boot when initialisation is single-threaded */
1304 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1306 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1315 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1316 struct mminit_pfnnid_cache *state)
1323 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1326 if (early_page_uninitialised(pfn))
1328 return __free_pages_boot_core(page, order);
1332 * Check that the whole (or subset of) a pageblock given by the interval of
1333 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1334 * with the migration of free compaction scanner. The scanners then need to
1335 * use only pfn_valid_within() check for arches that allow holes within
1338 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1340 * It's possible on some configurations to have a setup like node0 node1 node0
1341 * i.e. it's possible that all pages within a zones range of pages do not
1342 * belong to a single zone. We assume that a border between node0 and node1
1343 * can occur within a single pageblock, but not a node0 node1 node0
1344 * interleaving within a single pageblock. It is therefore sufficient to check
1345 * the first and last page of a pageblock and avoid checking each individual
1346 * page in a pageblock.
1348 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1349 unsigned long end_pfn, struct zone *zone)
1351 struct page *start_page;
1352 struct page *end_page;
1354 /* end_pfn is one past the range we are checking */
1357 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1360 start_page = pfn_to_page(start_pfn);
1362 if (page_zone(start_page) != zone)
1365 end_page = pfn_to_page(end_pfn);
1367 /* This gives a shorter code than deriving page_zone(end_page) */
1368 if (page_zone_id(start_page) != page_zone_id(end_page))
1374 void set_zone_contiguous(struct zone *zone)
1376 unsigned long block_start_pfn = zone->zone_start_pfn;
1377 unsigned long block_end_pfn;
1379 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1380 for (; block_start_pfn < zone_end_pfn(zone);
1381 block_start_pfn = block_end_pfn,
1382 block_end_pfn += pageblock_nr_pages) {
1384 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1386 if (!__pageblock_pfn_to_page(block_start_pfn,
1387 block_end_pfn, zone))
1391 /* We confirm that there is no hole */
1392 zone->contiguous = true;
1395 void clear_zone_contiguous(struct zone *zone)
1397 zone->contiguous = false;
1400 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1401 static void __init deferred_free_range(struct page *page,
1402 unsigned long pfn, int nr_pages)
1409 /* Free a large naturally-aligned chunk if possible */
1410 if (nr_pages == pageblock_nr_pages &&
1411 (pfn & (pageblock_nr_pages - 1)) == 0) {
1412 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1413 __free_pages_boot_core(page, pageblock_order);
1417 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1418 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1419 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1420 __free_pages_boot_core(page, 0);
1424 /* Completion tracking for deferred_init_memmap() threads */
1425 static atomic_t pgdat_init_n_undone __initdata;
1426 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1428 static inline void __init pgdat_init_report_one_done(void)
1430 if (atomic_dec_and_test(&pgdat_init_n_undone))
1431 complete(&pgdat_init_all_done_comp);
1434 /* Initialise remaining memory on a node */
1435 static int __init deferred_init_memmap(void *data)
1437 pg_data_t *pgdat = data;
1438 int nid = pgdat->node_id;
1439 struct mminit_pfnnid_cache nid_init_state = { };
1440 unsigned long start = jiffies;
1441 unsigned long nr_pages = 0;
1442 unsigned long walk_start, walk_end;
1445 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1446 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1448 if (first_init_pfn == ULONG_MAX) {
1449 pgdat_init_report_one_done();
1453 /* Bind memory initialisation thread to a local node if possible */
1454 if (!cpumask_empty(cpumask))
1455 set_cpus_allowed_ptr(current, cpumask);
1457 /* Sanity check boundaries */
1458 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1459 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1460 pgdat->first_deferred_pfn = ULONG_MAX;
1462 /* Only the highest zone is deferred so find it */
1463 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1464 zone = pgdat->node_zones + zid;
1465 if (first_init_pfn < zone_end_pfn(zone))
1469 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1470 unsigned long pfn, end_pfn;
1471 struct page *page = NULL;
1472 struct page *free_base_page = NULL;
1473 unsigned long free_base_pfn = 0;
1476 end_pfn = min(walk_end, zone_end_pfn(zone));
1477 pfn = first_init_pfn;
1478 if (pfn < walk_start)
1480 if (pfn < zone->zone_start_pfn)
1481 pfn = zone->zone_start_pfn;
1483 for (; pfn < end_pfn; pfn++) {
1484 if (!pfn_valid_within(pfn))
1488 * Ensure pfn_valid is checked every
1489 * pageblock_nr_pages for memory holes
1491 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1492 if (!pfn_valid(pfn)) {
1498 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1503 /* Minimise pfn page lookups and scheduler checks */
1504 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1507 nr_pages += nr_to_free;
1508 deferred_free_range(free_base_page,
1509 free_base_pfn, nr_to_free);
1510 free_base_page = NULL;
1511 free_base_pfn = nr_to_free = 0;
1513 page = pfn_to_page(pfn);
1518 VM_BUG_ON(page_zone(page) != zone);
1522 __init_single_page(page, pfn, zid, nid);
1523 if (!free_base_page) {
1524 free_base_page = page;
1525 free_base_pfn = pfn;
1530 /* Where possible, batch up pages for a single free */
1533 /* Free the current block of pages to allocator */
1534 nr_pages += nr_to_free;
1535 deferred_free_range(free_base_page, free_base_pfn,
1537 free_base_page = NULL;
1538 free_base_pfn = nr_to_free = 0;
1540 /* Free the last block of pages to allocator */
1541 nr_pages += nr_to_free;
1542 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1544 first_init_pfn = max(end_pfn, first_init_pfn);
1547 /* Sanity check that the next zone really is unpopulated */
1548 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1550 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1551 jiffies_to_msecs(jiffies - start));
1553 pgdat_init_report_one_done();
1556 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1558 void __init page_alloc_init_late(void)
1562 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1565 /* There will be num_node_state(N_MEMORY) threads */
1566 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1567 for_each_node_state(nid, N_MEMORY) {
1568 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1571 /* Block until all are initialised */
1572 wait_for_completion(&pgdat_init_all_done_comp);
1574 /* Reinit limits that are based on free pages after the kernel is up */
1575 files_maxfiles_init();
1578 for_each_populated_zone(zone)
1579 set_zone_contiguous(zone);
1583 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1584 void __init init_cma_reserved_pageblock(struct page *page)
1586 unsigned i = pageblock_nr_pages;
1587 struct page *p = page;
1590 __ClearPageReserved(p);
1591 set_page_count(p, 0);
1594 set_pageblock_migratetype(page, MIGRATE_CMA);
1596 if (pageblock_order >= MAX_ORDER) {
1597 i = pageblock_nr_pages;
1600 set_page_refcounted(p);
1601 __free_pages(p, MAX_ORDER - 1);
1602 p += MAX_ORDER_NR_PAGES;
1603 } while (i -= MAX_ORDER_NR_PAGES);
1605 set_page_refcounted(page);
1606 __free_pages(page, pageblock_order);
1609 adjust_managed_page_count(page, pageblock_nr_pages);
1614 * The order of subdivision here is critical for the IO subsystem.
1615 * Please do not alter this order without good reasons and regression
1616 * testing. Specifically, as large blocks of memory are subdivided,
1617 * the order in which smaller blocks are delivered depends on the order
1618 * they're subdivided in this function. This is the primary factor
1619 * influencing the order in which pages are delivered to the IO
1620 * subsystem according to empirical testing, and this is also justified
1621 * by considering the behavior of a buddy system containing a single
1622 * large block of memory acted on by a series of small allocations.
1623 * This behavior is a critical factor in sglist merging's success.
1627 static inline void expand(struct zone *zone, struct page *page,
1628 int low, int high, struct free_area *area,
1631 unsigned long size = 1 << high;
1633 while (high > low) {
1637 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1640 * Mark as guard pages (or page), that will allow to
1641 * merge back to allocator when buddy will be freed.
1642 * Corresponding page table entries will not be touched,
1643 * pages will stay not present in virtual address space
1645 if (set_page_guard(zone, &page[size], high, migratetype))
1648 list_add(&page[size].lru, &area->free_list[migratetype]);
1650 set_page_order(&page[size], high);
1654 static void check_new_page_bad(struct page *page)
1656 const char *bad_reason = NULL;
1657 unsigned long bad_flags = 0;
1659 if (unlikely(atomic_read(&page->_mapcount) != -1))
1660 bad_reason = "nonzero mapcount";
1661 if (unlikely(page->mapping != NULL))
1662 bad_reason = "non-NULL mapping";
1663 if (unlikely(page_ref_count(page) != 0))
1664 bad_reason = "nonzero _count";
1665 if (unlikely(page->flags & __PG_HWPOISON)) {
1666 bad_reason = "HWPoisoned (hardware-corrupted)";
1667 bad_flags = __PG_HWPOISON;
1668 /* Don't complain about hwpoisoned pages */
1669 page_mapcount_reset(page); /* remove PageBuddy */
1672 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1673 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1674 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1677 if (unlikely(page->mem_cgroup))
1678 bad_reason = "page still charged to cgroup";
1680 bad_page(page, bad_reason, bad_flags);
1684 * This page is about to be returned from the page allocator
1686 static inline int check_new_page(struct page *page)
1688 if (likely(page_expected_state(page,
1689 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1692 check_new_page_bad(page);
1696 static inline bool free_pages_prezeroed(bool poisoned)
1698 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1699 page_poisoning_enabled() && poisoned;
1702 #ifdef CONFIG_DEBUG_VM
1703 static bool check_pcp_refill(struct page *page)
1708 static bool check_new_pcp(struct page *page)
1710 return check_new_page(page);
1713 static bool check_pcp_refill(struct page *page)
1715 return check_new_page(page);
1717 static bool check_new_pcp(struct page *page)
1721 #endif /* CONFIG_DEBUG_VM */
1723 static bool check_new_pages(struct page *page, unsigned int order)
1726 for (i = 0; i < (1 << order); i++) {
1727 struct page *p = page + i;
1729 if (unlikely(check_new_page(p)))
1736 inline void post_alloc_hook(struct page *page, unsigned int order,
1739 set_page_private(page, 0);
1740 set_page_refcounted(page);
1742 arch_alloc_page(page, order);
1743 kernel_map_pages(page, 1 << order, 1);
1744 kernel_poison_pages(page, 1 << order, 1);
1745 kasan_alloc_pages(page, order);
1746 set_page_owner(page, order, gfp_flags);
1749 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1750 unsigned int alloc_flags)
1753 bool poisoned = true;
1755 for (i = 0; i < (1 << order); i++) {
1756 struct page *p = page + i;
1758 poisoned &= page_is_poisoned(p);
1761 post_alloc_hook(page, order, gfp_flags);
1763 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1764 for (i = 0; i < (1 << order); i++)
1765 clear_highpage(page + i);
1767 if (order && (gfp_flags & __GFP_COMP))
1768 prep_compound_page(page, order);
1771 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1772 * allocate the page. The expectation is that the caller is taking
1773 * steps that will free more memory. The caller should avoid the page
1774 * being used for !PFMEMALLOC purposes.
1776 if (alloc_flags & ALLOC_NO_WATERMARKS)
1777 set_page_pfmemalloc(page);
1779 clear_page_pfmemalloc(page);
1783 * Go through the free lists for the given migratetype and remove
1784 * the smallest available page from the freelists
1787 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1790 unsigned int current_order;
1791 struct free_area *area;
1794 /* Find a page of the appropriate size in the preferred list */
1795 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1796 area = &(zone->free_area[current_order]);
1797 page = list_first_entry_or_null(&area->free_list[migratetype],
1801 list_del(&page->lru);
1802 rmv_page_order(page);
1804 expand(zone, page, order, current_order, area, migratetype);
1805 set_pcppage_migratetype(page, migratetype);
1814 * This array describes the order lists are fallen back to when
1815 * the free lists for the desirable migrate type are depleted
1817 static int fallbacks[MIGRATE_TYPES][4] = {
1818 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1819 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1820 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1822 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1824 #ifdef CONFIG_MEMORY_ISOLATION
1825 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1830 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1833 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1836 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1837 unsigned int order) { return NULL; }
1841 * Move the free pages in a range to the free lists of the requested type.
1842 * Note that start_page and end_pages are not aligned on a pageblock
1843 * boundary. If alignment is required, use move_freepages_block()
1845 int move_freepages(struct zone *zone,
1846 struct page *start_page, struct page *end_page,
1851 int pages_moved = 0;
1853 #ifndef CONFIG_HOLES_IN_ZONE
1855 * page_zone is not safe to call in this context when
1856 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1857 * anyway as we check zone boundaries in move_freepages_block().
1858 * Remove at a later date when no bug reports exist related to
1859 * grouping pages by mobility
1861 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1864 for (page = start_page; page <= end_page;) {
1865 if (!pfn_valid_within(page_to_pfn(page))) {
1870 /* Make sure we are not inadvertently changing nodes */
1871 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1873 if (!PageBuddy(page)) {
1878 order = page_order(page);
1879 list_move(&page->lru,
1880 &zone->free_area[order].free_list[migratetype]);
1882 pages_moved += 1 << order;
1888 int move_freepages_block(struct zone *zone, struct page *page,
1891 unsigned long start_pfn, end_pfn;
1892 struct page *start_page, *end_page;
1894 start_pfn = page_to_pfn(page);
1895 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1896 start_page = pfn_to_page(start_pfn);
1897 end_page = start_page + pageblock_nr_pages - 1;
1898 end_pfn = start_pfn + pageblock_nr_pages - 1;
1900 /* Do not cross zone boundaries */
1901 if (!zone_spans_pfn(zone, start_pfn))
1903 if (!zone_spans_pfn(zone, end_pfn))
1906 return move_freepages(zone, start_page, end_page, migratetype);
1909 static void change_pageblock_range(struct page *pageblock_page,
1910 int start_order, int migratetype)
1912 int nr_pageblocks = 1 << (start_order - pageblock_order);
1914 while (nr_pageblocks--) {
1915 set_pageblock_migratetype(pageblock_page, migratetype);
1916 pageblock_page += pageblock_nr_pages;
1921 * When we are falling back to another migratetype during allocation, try to
1922 * steal extra free pages from the same pageblocks to satisfy further
1923 * allocations, instead of polluting multiple pageblocks.
1925 * If we are stealing a relatively large buddy page, it is likely there will
1926 * be more free pages in the pageblock, so try to steal them all. For
1927 * reclaimable and unmovable allocations, we steal regardless of page size,
1928 * as fragmentation caused by those allocations polluting movable pageblocks
1929 * is worse than movable allocations stealing from unmovable and reclaimable
1932 static bool can_steal_fallback(unsigned int order, int start_mt)
1935 * Leaving this order check is intended, although there is
1936 * relaxed order check in next check. The reason is that
1937 * we can actually steal whole pageblock if this condition met,
1938 * but, below check doesn't guarantee it and that is just heuristic
1939 * so could be changed anytime.
1941 if (order >= pageblock_order)
1944 if (order >= pageblock_order / 2 ||
1945 start_mt == MIGRATE_RECLAIMABLE ||
1946 start_mt == MIGRATE_UNMOVABLE ||
1947 page_group_by_mobility_disabled)
1954 * This function implements actual steal behaviour. If order is large enough,
1955 * we can steal whole pageblock. If not, we first move freepages in this
1956 * pageblock and check whether half of pages are moved or not. If half of
1957 * pages are moved, we can change migratetype of pageblock and permanently
1958 * use it's pages as requested migratetype in the future.
1960 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1963 unsigned int current_order = page_order(page);
1966 /* Take ownership for orders >= pageblock_order */
1967 if (current_order >= pageblock_order) {
1968 change_pageblock_range(page, current_order, start_type);
1972 pages = move_freepages_block(zone, page, start_type);
1974 /* Claim the whole block if over half of it is free */
1975 if (pages >= (1 << (pageblock_order-1)) ||
1976 page_group_by_mobility_disabled)
1977 set_pageblock_migratetype(page, start_type);
1981 * Check whether there is a suitable fallback freepage with requested order.
1982 * If only_stealable is true, this function returns fallback_mt only if
1983 * we can steal other freepages all together. This would help to reduce
1984 * fragmentation due to mixed migratetype pages in one pageblock.
1986 int find_suitable_fallback(struct free_area *area, unsigned int order,
1987 int migratetype, bool only_stealable, bool *can_steal)
1992 if (area->nr_free == 0)
1997 fallback_mt = fallbacks[migratetype][i];
1998 if (fallback_mt == MIGRATE_TYPES)
2001 if (list_empty(&area->free_list[fallback_mt]))
2004 if (can_steal_fallback(order, migratetype))
2007 if (!only_stealable)
2018 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2019 * there are no empty page blocks that contain a page with a suitable order
2021 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2022 unsigned int alloc_order)
2025 unsigned long max_managed, flags;
2028 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2029 * Check is race-prone but harmless.
2031 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2032 if (zone->nr_reserved_highatomic >= max_managed)
2035 spin_lock_irqsave(&zone->lock, flags);
2037 /* Recheck the nr_reserved_highatomic limit under the lock */
2038 if (zone->nr_reserved_highatomic >= max_managed)
2042 mt = get_pageblock_migratetype(page);
2043 if (mt != MIGRATE_HIGHATOMIC &&
2044 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2045 zone->nr_reserved_highatomic += pageblock_nr_pages;
2046 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2047 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2051 spin_unlock_irqrestore(&zone->lock, flags);
2055 * Used when an allocation is about to fail under memory pressure. This
2056 * potentially hurts the reliability of high-order allocations when under
2057 * intense memory pressure but failed atomic allocations should be easier
2058 * to recover from than an OOM.
2060 * If @force is true, try to unreserve a pageblock even though highatomic
2061 * pageblock is exhausted.
2063 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2066 struct zonelist *zonelist = ac->zonelist;
2067 unsigned long flags;
2074 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2077 * Preserve at least one pageblock unless memory pressure
2080 if (!force && zone->nr_reserved_highatomic <=
2084 spin_lock_irqsave(&zone->lock, flags);
2085 for (order = 0; order < MAX_ORDER; order++) {
2086 struct free_area *area = &(zone->free_area[order]);
2088 page = list_first_entry_or_null(
2089 &area->free_list[MIGRATE_HIGHATOMIC],
2095 * In page freeing path, migratetype change is racy so
2096 * we can counter several free pages in a pageblock
2097 * in this loop althoug we changed the pageblock type
2098 * from highatomic to ac->migratetype. So we should
2099 * adjust the count once.
2101 if (get_pageblock_migratetype(page) ==
2102 MIGRATE_HIGHATOMIC) {
2104 * It should never happen but changes to
2105 * locking could inadvertently allow a per-cpu
2106 * drain to add pages to MIGRATE_HIGHATOMIC
2107 * while unreserving so be safe and watch for
2110 zone->nr_reserved_highatomic -= min(
2112 zone->nr_reserved_highatomic);
2116 * Convert to ac->migratetype and avoid the normal
2117 * pageblock stealing heuristics. Minimally, the caller
2118 * is doing the work and needs the pages. More
2119 * importantly, if the block was always converted to
2120 * MIGRATE_UNMOVABLE or another type then the number
2121 * of pageblocks that cannot be completely freed
2124 set_pageblock_migratetype(page, ac->migratetype);
2125 ret = move_freepages_block(zone, page, ac->migratetype);
2127 spin_unlock_irqrestore(&zone->lock, flags);
2131 spin_unlock_irqrestore(&zone->lock, flags);
2137 /* Remove an element from the buddy allocator from the fallback list */
2138 static inline struct page *
2139 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2141 struct free_area *area;
2142 unsigned int current_order;
2147 /* Find the largest possible block of pages in the other list */
2148 for (current_order = MAX_ORDER-1;
2149 current_order >= order && current_order <= MAX_ORDER-1;
2151 area = &(zone->free_area[current_order]);
2152 fallback_mt = find_suitable_fallback(area, current_order,
2153 start_migratetype, false, &can_steal);
2154 if (fallback_mt == -1)
2157 page = list_first_entry(&area->free_list[fallback_mt],
2160 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC)
2161 steal_suitable_fallback(zone, page, start_migratetype);
2163 /* Remove the page from the freelists */
2165 list_del(&page->lru);
2166 rmv_page_order(page);
2168 expand(zone, page, order, current_order, area,
2171 * The pcppage_migratetype may differ from pageblock's
2172 * migratetype depending on the decisions in
2173 * find_suitable_fallback(). This is OK as long as it does not
2174 * differ for MIGRATE_CMA pageblocks. Those can be used as
2175 * fallback only via special __rmqueue_cma_fallback() function
2177 set_pcppage_migratetype(page, start_migratetype);
2179 trace_mm_page_alloc_extfrag(page, order, current_order,
2180 start_migratetype, fallback_mt);
2189 * Do the hard work of removing an element from the buddy allocator.
2190 * Call me with the zone->lock already held.
2192 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2197 page = __rmqueue_smallest(zone, order, migratetype);
2198 if (unlikely(!page)) {
2199 if (migratetype == MIGRATE_MOVABLE)
2200 page = __rmqueue_cma_fallback(zone, order);
2203 page = __rmqueue_fallback(zone, order, migratetype);
2206 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2211 * Obtain a specified number of elements from the buddy allocator, all under
2212 * a single hold of the lock, for efficiency. Add them to the supplied list.
2213 * Returns the number of new pages which were placed at *list.
2215 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2216 unsigned long count, struct list_head *list,
2217 int migratetype, bool cold)
2220 unsigned long flags;
2222 spin_lock_irqsave(&zone->lock, flags);
2223 for (i = 0; i < count; ++i) {
2224 struct page *page = __rmqueue(zone, order, migratetype);
2225 if (unlikely(page == NULL))
2228 if (unlikely(check_pcp_refill(page)))
2232 * Split buddy pages returned by expand() are received here
2233 * in physical page order. The page is added to the callers and
2234 * list and the list head then moves forward. From the callers
2235 * perspective, the linked list is ordered by page number in
2236 * some conditions. This is useful for IO devices that can
2237 * merge IO requests if the physical pages are ordered
2241 list_add(&page->lru, list);
2243 list_add_tail(&page->lru, list);
2246 if (is_migrate_cma(get_pcppage_migratetype(page)))
2247 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2252 * i pages were removed from the buddy list even if some leak due
2253 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2254 * on i. Do not confuse with 'alloced' which is the number of
2255 * pages added to the pcp list.
2257 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2258 spin_unlock_irqrestore(&zone->lock, flags);
2264 * Called from the vmstat counter updater to drain pagesets of this
2265 * currently executing processor on remote nodes after they have
2268 * Note that this function must be called with the thread pinned to
2269 * a single processor.
2271 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2273 unsigned long flags;
2274 int to_drain, batch;
2276 local_irq_save(flags);
2277 batch = READ_ONCE(pcp->batch);
2278 to_drain = min(pcp->count, batch);
2280 free_pcppages_bulk(zone, to_drain, pcp);
2281 pcp->count -= to_drain;
2283 local_irq_restore(flags);
2288 * Drain pcplists of the indicated processor and zone.
2290 * The processor must either be the current processor and the
2291 * thread pinned to the current processor or a processor that
2294 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2296 unsigned long flags;
2297 struct per_cpu_pageset *pset;
2298 struct per_cpu_pages *pcp;
2300 local_irq_save(flags);
2301 pset = per_cpu_ptr(zone->pageset, cpu);
2305 free_pcppages_bulk(zone, pcp->count, pcp);
2308 local_irq_restore(flags);
2312 * Drain pcplists of all zones on the indicated processor.
2314 * The processor must either be the current processor and the
2315 * thread pinned to the current processor or a processor that
2318 static void drain_pages(unsigned int cpu)
2322 for_each_populated_zone(zone) {
2323 drain_pages_zone(cpu, zone);
2328 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2330 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2331 * the single zone's pages.
2333 void drain_local_pages(struct zone *zone)
2335 int cpu = smp_processor_id();
2338 drain_pages_zone(cpu, zone);
2343 static void drain_local_pages_wq(struct work_struct *work)
2346 * drain_all_pages doesn't use proper cpu hotplug protection so
2347 * we can race with cpu offline when the WQ can move this from
2348 * a cpu pinned worker to an unbound one. We can operate on a different
2349 * cpu which is allright but we also have to make sure to not move to
2353 drain_local_pages(NULL);
2358 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2360 * When zone parameter is non-NULL, spill just the single zone's pages.
2362 * Note that this can be extremely slow as the draining happens in a workqueue.
2364 void drain_all_pages(struct zone *zone)
2369 * Allocate in the BSS so we wont require allocation in
2370 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2372 static cpumask_t cpus_with_pcps;
2374 /* Workqueues cannot recurse */
2375 if (current->flags & PF_WQ_WORKER)
2379 * Do not drain if one is already in progress unless it's specific to
2380 * a zone. Such callers are primarily CMA and memory hotplug and need
2381 * the drain to be complete when the call returns.
2383 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2386 mutex_lock(&pcpu_drain_mutex);
2390 * We don't care about racing with CPU hotplug event
2391 * as offline notification will cause the notified
2392 * cpu to drain that CPU pcps and on_each_cpu_mask
2393 * disables preemption as part of its processing
2395 for_each_online_cpu(cpu) {
2396 struct per_cpu_pageset *pcp;
2398 bool has_pcps = false;
2401 pcp = per_cpu_ptr(zone->pageset, cpu);
2405 for_each_populated_zone(z) {
2406 pcp = per_cpu_ptr(z->pageset, cpu);
2407 if (pcp->pcp.count) {
2415 cpumask_set_cpu(cpu, &cpus_with_pcps);
2417 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2420 for_each_cpu(cpu, &cpus_with_pcps) {
2421 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2422 INIT_WORK(work, drain_local_pages_wq);
2423 schedule_work_on(cpu, work);
2425 for_each_cpu(cpu, &cpus_with_pcps)
2426 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2428 mutex_unlock(&pcpu_drain_mutex);
2431 #ifdef CONFIG_HIBERNATION
2433 void mark_free_pages(struct zone *zone)
2435 unsigned long pfn, max_zone_pfn;
2436 unsigned long flags;
2437 unsigned int order, t;
2440 if (zone_is_empty(zone))
2443 spin_lock_irqsave(&zone->lock, flags);
2445 max_zone_pfn = zone_end_pfn(zone);
2446 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2447 if (pfn_valid(pfn)) {
2448 page = pfn_to_page(pfn);
2450 if (page_zone(page) != zone)
2453 if (!swsusp_page_is_forbidden(page))
2454 swsusp_unset_page_free(page);
2457 for_each_migratetype_order(order, t) {
2458 list_for_each_entry(page,
2459 &zone->free_area[order].free_list[t], lru) {
2462 pfn = page_to_pfn(page);
2463 for (i = 0; i < (1UL << order); i++)
2464 swsusp_set_page_free(pfn_to_page(pfn + i));
2467 spin_unlock_irqrestore(&zone->lock, flags);
2469 #endif /* CONFIG_PM */
2472 * Free a 0-order page
2473 * cold == true ? free a cold page : free a hot page
2475 void free_hot_cold_page(struct page *page, bool cold)
2477 struct zone *zone = page_zone(page);
2478 struct per_cpu_pages *pcp;
2479 unsigned long pfn = page_to_pfn(page);
2482 if (in_interrupt()) {
2483 __free_pages_ok(page, 0);
2487 if (!free_pcp_prepare(page))
2490 migratetype = get_pfnblock_migratetype(page, pfn);
2491 set_pcppage_migratetype(page, migratetype);
2495 * We only track unmovable, reclaimable and movable on pcp lists.
2496 * Free ISOLATE pages back to the allocator because they are being
2497 * offlined but treat RESERVE as movable pages so we can get those
2498 * areas back if necessary. Otherwise, we may have to free
2499 * excessively into the page allocator
2501 if (migratetype >= MIGRATE_PCPTYPES) {
2502 if (unlikely(is_migrate_isolate(migratetype))) {
2503 free_one_page(zone, page, pfn, 0, migratetype);
2506 migratetype = MIGRATE_MOVABLE;
2509 __count_vm_event(PGFREE);
2510 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2512 list_add(&page->lru, &pcp->lists[migratetype]);
2514 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2516 if (pcp->count >= pcp->high) {
2517 unsigned long batch = READ_ONCE(pcp->batch);
2518 free_pcppages_bulk(zone, batch, pcp);
2519 pcp->count -= batch;
2527 * Free a list of 0-order pages
2529 void free_hot_cold_page_list(struct list_head *list, bool cold)
2531 struct page *page, *next;
2533 list_for_each_entry_safe(page, next, list, lru) {
2534 trace_mm_page_free_batched(page, cold);
2535 free_hot_cold_page(page, cold);
2540 * split_page takes a non-compound higher-order page, and splits it into
2541 * n (1<<order) sub-pages: page[0..n]
2542 * Each sub-page must be freed individually.
2544 * Note: this is probably too low level an operation for use in drivers.
2545 * Please consult with lkml before using this in your driver.
2547 void split_page(struct page *page, unsigned int order)
2551 VM_BUG_ON_PAGE(PageCompound(page), page);
2552 VM_BUG_ON_PAGE(!page_count(page), page);
2554 #ifdef CONFIG_KMEMCHECK
2556 * Split shadow pages too, because free(page[0]) would
2557 * otherwise free the whole shadow.
2559 if (kmemcheck_page_is_tracked(page))
2560 split_page(virt_to_page(page[0].shadow), order);
2563 for (i = 1; i < (1 << order); i++)
2564 set_page_refcounted(page + i);
2565 split_page_owner(page, order);
2567 EXPORT_SYMBOL_GPL(split_page);
2569 int __isolate_free_page(struct page *page, unsigned int order)
2571 unsigned long watermark;
2575 BUG_ON(!PageBuddy(page));
2577 zone = page_zone(page);
2578 mt = get_pageblock_migratetype(page);
2580 if (!is_migrate_isolate(mt)) {
2582 * Obey watermarks as if the page was being allocated. We can
2583 * emulate a high-order watermark check with a raised order-0
2584 * watermark, because we already know our high-order page
2587 watermark = min_wmark_pages(zone) + (1UL << order);
2588 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2591 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2594 /* Remove page from free list */
2595 list_del(&page->lru);
2596 zone->free_area[order].nr_free--;
2597 rmv_page_order(page);
2600 * Set the pageblock if the isolated page is at least half of a
2603 if (order >= pageblock_order - 1) {
2604 struct page *endpage = page + (1 << order) - 1;
2605 for (; page < endpage; page += pageblock_nr_pages) {
2606 int mt = get_pageblock_migratetype(page);
2607 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2608 && mt != MIGRATE_HIGHATOMIC)
2609 set_pageblock_migratetype(page,
2615 return 1UL << order;
2619 * Update NUMA hit/miss statistics
2621 * Must be called with interrupts disabled.
2623 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2626 enum zone_stat_item local_stat = NUMA_LOCAL;
2628 if (z->node != numa_node_id())
2629 local_stat = NUMA_OTHER;
2631 if (z->node == preferred_zone->node)
2632 __inc_zone_state(z, NUMA_HIT);
2634 __inc_zone_state(z, NUMA_MISS);
2635 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2637 __inc_zone_state(z, local_stat);
2641 /* Remove page from the per-cpu list, caller must protect the list */
2642 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2643 bool cold, struct per_cpu_pages *pcp,
2644 struct list_head *list)
2648 VM_BUG_ON(in_interrupt());
2651 if (list_empty(list)) {
2652 pcp->count += rmqueue_bulk(zone, 0,
2655 if (unlikely(list_empty(list)))
2660 page = list_last_entry(list, struct page, lru);
2662 page = list_first_entry(list, struct page, lru);
2664 list_del(&page->lru);
2666 } while (check_new_pcp(page));
2671 /* Lock and remove page from the per-cpu list */
2672 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2673 struct zone *zone, unsigned int order,
2674 gfp_t gfp_flags, int migratetype)
2676 struct per_cpu_pages *pcp;
2677 struct list_head *list;
2678 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2682 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2683 list = &pcp->lists[migratetype];
2684 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2686 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2687 zone_statistics(preferred_zone, zone);
2694 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2697 struct page *rmqueue(struct zone *preferred_zone,
2698 struct zone *zone, unsigned int order,
2699 gfp_t gfp_flags, unsigned int alloc_flags,
2702 unsigned long flags;
2705 if (likely(order == 0) && !in_interrupt()) {
2706 page = rmqueue_pcplist(preferred_zone, zone, order,
2707 gfp_flags, migratetype);
2712 * We most definitely don't want callers attempting to
2713 * allocate greater than order-1 page units with __GFP_NOFAIL.
2715 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2716 spin_lock_irqsave(&zone->lock, flags);
2720 if (alloc_flags & ALLOC_HARDER) {
2721 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2723 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2726 page = __rmqueue(zone, order, migratetype);
2727 } while (page && check_new_pages(page, order));
2728 spin_unlock(&zone->lock);
2731 __mod_zone_freepage_state(zone, -(1 << order),
2732 get_pcppage_migratetype(page));
2734 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2735 zone_statistics(preferred_zone, zone);
2736 local_irq_restore(flags);
2739 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2743 local_irq_restore(flags);
2747 #ifdef CONFIG_FAIL_PAGE_ALLOC
2750 struct fault_attr attr;
2752 bool ignore_gfp_highmem;
2753 bool ignore_gfp_reclaim;
2755 } fail_page_alloc = {
2756 .attr = FAULT_ATTR_INITIALIZER,
2757 .ignore_gfp_reclaim = true,
2758 .ignore_gfp_highmem = true,
2762 static int __init setup_fail_page_alloc(char *str)
2764 return setup_fault_attr(&fail_page_alloc.attr, str);
2766 __setup("fail_page_alloc=", setup_fail_page_alloc);
2768 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2770 if (order < fail_page_alloc.min_order)
2772 if (gfp_mask & __GFP_NOFAIL)
2774 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2776 if (fail_page_alloc.ignore_gfp_reclaim &&
2777 (gfp_mask & __GFP_DIRECT_RECLAIM))
2780 return should_fail(&fail_page_alloc.attr, 1 << order);
2783 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2785 static int __init fail_page_alloc_debugfs(void)
2787 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2790 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2791 &fail_page_alloc.attr);
2793 return PTR_ERR(dir);
2795 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2796 &fail_page_alloc.ignore_gfp_reclaim))
2798 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2799 &fail_page_alloc.ignore_gfp_highmem))
2801 if (!debugfs_create_u32("min-order", mode, dir,
2802 &fail_page_alloc.min_order))
2807 debugfs_remove_recursive(dir);
2812 late_initcall(fail_page_alloc_debugfs);
2814 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2816 #else /* CONFIG_FAIL_PAGE_ALLOC */
2818 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2823 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2826 * Return true if free base pages are above 'mark'. For high-order checks it
2827 * will return true of the order-0 watermark is reached and there is at least
2828 * one free page of a suitable size. Checking now avoids taking the zone lock
2829 * to check in the allocation paths if no pages are free.
2831 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2832 int classzone_idx, unsigned int alloc_flags,
2837 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2839 /* free_pages may go negative - that's OK */
2840 free_pages -= (1 << order) - 1;
2842 if (alloc_flags & ALLOC_HIGH)
2846 * If the caller does not have rights to ALLOC_HARDER then subtract
2847 * the high-atomic reserves. This will over-estimate the size of the
2848 * atomic reserve but it avoids a search.
2850 if (likely(!alloc_harder))
2851 free_pages -= z->nr_reserved_highatomic;
2856 /* If allocation can't use CMA areas don't use free CMA pages */
2857 if (!(alloc_flags & ALLOC_CMA))
2858 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2862 * Check watermarks for an order-0 allocation request. If these
2863 * are not met, then a high-order request also cannot go ahead
2864 * even if a suitable page happened to be free.
2866 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2869 /* If this is an order-0 request then the watermark is fine */
2873 /* For a high-order request, check at least one suitable page is free */
2874 for (o = order; o < MAX_ORDER; o++) {
2875 struct free_area *area = &z->free_area[o];
2884 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2885 if (!list_empty(&area->free_list[mt]))
2890 if ((alloc_flags & ALLOC_CMA) &&
2891 !list_empty(&area->free_list[MIGRATE_CMA])) {
2899 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2900 int classzone_idx, unsigned int alloc_flags)
2902 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2903 zone_page_state(z, NR_FREE_PAGES));
2906 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2907 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2909 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2913 /* If allocation can't use CMA areas don't use free CMA pages */
2914 if (!(alloc_flags & ALLOC_CMA))
2915 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2919 * Fast check for order-0 only. If this fails then the reserves
2920 * need to be calculated. There is a corner case where the check
2921 * passes but only the high-order atomic reserve are free. If
2922 * the caller is !atomic then it'll uselessly search the free
2923 * list. That corner case is then slower but it is harmless.
2925 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2928 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2932 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2933 unsigned long mark, int classzone_idx)
2935 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2937 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2938 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2940 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2945 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2947 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2950 #else /* CONFIG_NUMA */
2951 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2955 #endif /* CONFIG_NUMA */
2958 * get_page_from_freelist goes through the zonelist trying to allocate
2961 static struct page *
2962 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2963 const struct alloc_context *ac)
2965 struct zoneref *z = ac->preferred_zoneref;
2967 struct pglist_data *last_pgdat_dirty_limit = NULL;
2970 * Scan zonelist, looking for a zone with enough free.
2971 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2973 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2978 if (cpusets_enabled() &&
2979 (alloc_flags & ALLOC_CPUSET) &&
2980 !__cpuset_zone_allowed(zone, gfp_mask))
2983 * When allocating a page cache page for writing, we
2984 * want to get it from a node that is within its dirty
2985 * limit, such that no single node holds more than its
2986 * proportional share of globally allowed dirty pages.
2987 * The dirty limits take into account the node's
2988 * lowmem reserves and high watermark so that kswapd
2989 * should be able to balance it without having to
2990 * write pages from its LRU list.
2992 * XXX: For now, allow allocations to potentially
2993 * exceed the per-node dirty limit in the slowpath
2994 * (spread_dirty_pages unset) before going into reclaim,
2995 * which is important when on a NUMA setup the allowed
2996 * nodes are together not big enough to reach the
2997 * global limit. The proper fix for these situations
2998 * will require awareness of nodes in the
2999 * dirty-throttling and the flusher threads.
3001 if (ac->spread_dirty_pages) {
3002 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3005 if (!node_dirty_ok(zone->zone_pgdat)) {
3006 last_pgdat_dirty_limit = zone->zone_pgdat;
3011 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3012 if (!zone_watermark_fast(zone, order, mark,
3013 ac_classzone_idx(ac), alloc_flags)) {
3016 /* Checked here to keep the fast path fast */
3017 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3018 if (alloc_flags & ALLOC_NO_WATERMARKS)
3021 if (node_reclaim_mode == 0 ||
3022 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3025 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3027 case NODE_RECLAIM_NOSCAN:
3030 case NODE_RECLAIM_FULL:
3031 /* scanned but unreclaimable */
3034 /* did we reclaim enough */
3035 if (zone_watermark_ok(zone, order, mark,
3036 ac_classzone_idx(ac), alloc_flags))
3044 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3045 gfp_mask, alloc_flags, ac->migratetype);
3047 prep_new_page(page, order, gfp_mask, alloc_flags);
3050 * If this is a high-order atomic allocation then check
3051 * if the pageblock should be reserved for the future
3053 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3054 reserve_highatomic_pageblock(page, zone, order);
3064 * Large machines with many possible nodes should not always dump per-node
3065 * meminfo in irq context.
3067 static inline bool should_suppress_show_mem(void)
3072 ret = in_interrupt();
3077 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3079 unsigned int filter = SHOW_MEM_FILTER_NODES;
3080 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3082 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3086 * This documents exceptions given to allocations in certain
3087 * contexts that are allowed to allocate outside current's set
3090 if (!(gfp_mask & __GFP_NOMEMALLOC))
3091 if (test_thread_flag(TIF_MEMDIE) ||
3092 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3093 filter &= ~SHOW_MEM_FILTER_NODES;
3094 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3095 filter &= ~SHOW_MEM_FILTER_NODES;
3097 show_mem(filter, nodemask);
3100 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3102 struct va_format vaf;
3104 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3105 DEFAULT_RATELIMIT_BURST);
3107 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3108 debug_guardpage_minorder() > 0)
3111 pr_warn("%s: ", current->comm);
3113 va_start(args, fmt);
3116 pr_cont("%pV", &vaf);
3119 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3121 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3123 pr_cont("(null)\n");
3125 cpuset_print_current_mems_allowed();
3128 warn_alloc_show_mem(gfp_mask, nodemask);
3131 static inline struct page *
3132 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3133 unsigned int alloc_flags,
3134 const struct alloc_context *ac)
3138 page = get_page_from_freelist(gfp_mask, order,
3139 alloc_flags|ALLOC_CPUSET, ac);
3141 * fallback to ignore cpuset restriction if our nodes
3145 page = get_page_from_freelist(gfp_mask, order,
3151 static inline struct page *
3152 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3153 const struct alloc_context *ac, unsigned long *did_some_progress)
3155 struct oom_control oc = {
3156 .zonelist = ac->zonelist,
3157 .nodemask = ac->nodemask,
3159 .gfp_mask = gfp_mask,
3164 *did_some_progress = 0;
3167 * Acquire the oom lock. If that fails, somebody else is
3168 * making progress for us.
3170 if (!mutex_trylock(&oom_lock)) {
3171 *did_some_progress = 1;
3172 schedule_timeout_uninterruptible(1);
3177 * Go through the zonelist yet one more time, keep very high watermark
3178 * here, this is only to catch a parallel oom killing, we must fail if
3179 * we're still under heavy pressure.
3181 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3182 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3186 /* Coredumps can quickly deplete all memory reserves */
3187 if (current->flags & PF_DUMPCORE)
3189 /* The OOM killer will not help higher order allocs */
3190 if (order > PAGE_ALLOC_COSTLY_ORDER)
3192 /* The OOM killer does not needlessly kill tasks for lowmem */
3193 if (ac->high_zoneidx < ZONE_NORMAL)
3195 if (pm_suspended_storage())
3198 * XXX: GFP_NOFS allocations should rather fail than rely on
3199 * other request to make a forward progress.
3200 * We are in an unfortunate situation where out_of_memory cannot
3201 * do much for this context but let's try it to at least get
3202 * access to memory reserved if the current task is killed (see
3203 * out_of_memory). Once filesystems are ready to handle allocation
3204 * failures more gracefully we should just bail out here.
3207 /* The OOM killer may not free memory on a specific node */
3208 if (gfp_mask & __GFP_THISNODE)
3211 /* Exhausted what can be done so it's blamo time */
3212 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3213 *did_some_progress = 1;
3216 * Help non-failing allocations by giving them access to memory
3219 if (gfp_mask & __GFP_NOFAIL)
3220 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3221 ALLOC_NO_WATERMARKS, ac);
3224 mutex_unlock(&oom_lock);
3229 * Maximum number of compaction retries wit a progress before OOM
3230 * killer is consider as the only way to move forward.
3232 #define MAX_COMPACT_RETRIES 16
3234 #ifdef CONFIG_COMPACTION
3235 /* Try memory compaction for high-order allocations before reclaim */
3236 static struct page *
3237 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3238 unsigned int alloc_flags, const struct alloc_context *ac,
3239 enum compact_priority prio, enum compact_result *compact_result)
3246 current->flags |= PF_MEMALLOC;
3247 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3249 current->flags &= ~PF_MEMALLOC;
3251 if (*compact_result <= COMPACT_INACTIVE)
3255 * At least in one zone compaction wasn't deferred or skipped, so let's
3256 * count a compaction stall
3258 count_vm_event(COMPACTSTALL);
3260 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3263 struct zone *zone = page_zone(page);
3265 zone->compact_blockskip_flush = false;
3266 compaction_defer_reset(zone, order, true);
3267 count_vm_event(COMPACTSUCCESS);
3272 * It's bad if compaction run occurs and fails. The most likely reason
3273 * is that pages exist, but not enough to satisfy watermarks.
3275 count_vm_event(COMPACTFAIL);
3283 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3284 enum compact_result compact_result,
3285 enum compact_priority *compact_priority,
3286 int *compaction_retries)
3288 int max_retries = MAX_COMPACT_RETRIES;
3291 int retries = *compaction_retries;
3292 enum compact_priority priority = *compact_priority;
3297 if (compaction_made_progress(compact_result))
3298 (*compaction_retries)++;
3301 * compaction considers all the zone as desperately out of memory
3302 * so it doesn't really make much sense to retry except when the
3303 * failure could be caused by insufficient priority
3305 if (compaction_failed(compact_result))
3306 goto check_priority;
3309 * make sure the compaction wasn't deferred or didn't bail out early
3310 * due to locks contention before we declare that we should give up.
3311 * But do not retry if the given zonelist is not suitable for
3314 if (compaction_withdrawn(compact_result)) {
3315 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3320 * !costly requests are much more important than __GFP_REPEAT
3321 * costly ones because they are de facto nofail and invoke OOM
3322 * killer to move on while costly can fail and users are ready
3323 * to cope with that. 1/4 retries is rather arbitrary but we
3324 * would need much more detailed feedback from compaction to
3325 * make a better decision.
3327 if (order > PAGE_ALLOC_COSTLY_ORDER)
3329 if (*compaction_retries <= max_retries) {
3335 * Make sure there are attempts at the highest priority if we exhausted
3336 * all retries or failed at the lower priorities.
3339 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3340 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3342 if (*compact_priority > min_priority) {
3343 (*compact_priority)--;
3344 *compaction_retries = 0;
3348 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3352 static inline struct page *
3353 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3354 unsigned int alloc_flags, const struct alloc_context *ac,
3355 enum compact_priority prio, enum compact_result *compact_result)
3357 *compact_result = COMPACT_SKIPPED;
3362 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3363 enum compact_result compact_result,
3364 enum compact_priority *compact_priority,
3365 int *compaction_retries)
3370 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3374 * There are setups with compaction disabled which would prefer to loop
3375 * inside the allocator rather than hit the oom killer prematurely.
3376 * Let's give them a good hope and keep retrying while the order-0
3377 * watermarks are OK.
3379 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3381 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3382 ac_classzone_idx(ac), alloc_flags))
3387 #endif /* CONFIG_COMPACTION */
3389 /* Perform direct synchronous page reclaim */
3391 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3392 const struct alloc_context *ac)
3394 struct reclaim_state reclaim_state;
3399 /* We now go into synchronous reclaim */
3400 cpuset_memory_pressure_bump();
3401 current->flags |= PF_MEMALLOC;
3402 lockdep_set_current_reclaim_state(gfp_mask);
3403 reclaim_state.reclaimed_slab = 0;
3404 current->reclaim_state = &reclaim_state;
3406 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3409 current->reclaim_state = NULL;
3410 lockdep_clear_current_reclaim_state();
3411 current->flags &= ~PF_MEMALLOC;
3418 /* The really slow allocator path where we enter direct reclaim */
3419 static inline struct page *
3420 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3421 unsigned int alloc_flags, const struct alloc_context *ac,
3422 unsigned long *did_some_progress)
3424 struct page *page = NULL;
3425 bool drained = false;
3427 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3428 if (unlikely(!(*did_some_progress)))
3432 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3435 * If an allocation failed after direct reclaim, it could be because
3436 * pages are pinned on the per-cpu lists or in high alloc reserves.
3437 * Shrink them them and try again
3439 if (!page && !drained) {
3440 unreserve_highatomic_pageblock(ac, false);
3441 drain_all_pages(NULL);
3449 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3453 pg_data_t *last_pgdat = NULL;
3455 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3456 ac->high_zoneidx, ac->nodemask) {
3457 if (last_pgdat != zone->zone_pgdat)
3458 wakeup_kswapd(zone, order, ac->high_zoneidx);
3459 last_pgdat = zone->zone_pgdat;
3463 static inline unsigned int
3464 gfp_to_alloc_flags(gfp_t gfp_mask)
3466 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3468 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3469 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3472 * The caller may dip into page reserves a bit more if the caller
3473 * cannot run direct reclaim, or if the caller has realtime scheduling
3474 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3475 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3477 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3479 if (gfp_mask & __GFP_ATOMIC) {
3481 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3482 * if it can't schedule.
3484 if (!(gfp_mask & __GFP_NOMEMALLOC))
3485 alloc_flags |= ALLOC_HARDER;
3487 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3488 * comment for __cpuset_node_allowed().
3490 alloc_flags &= ~ALLOC_CPUSET;
3491 } else if (unlikely(rt_task(current)) && !in_interrupt())
3492 alloc_flags |= ALLOC_HARDER;
3495 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3496 alloc_flags |= ALLOC_CMA;
3501 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3503 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3506 if (gfp_mask & __GFP_MEMALLOC)
3508 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3510 if (!in_interrupt() &&
3511 ((current->flags & PF_MEMALLOC) ||
3512 unlikely(test_thread_flag(TIF_MEMDIE))))
3519 * Maximum number of reclaim retries without any progress before OOM killer
3520 * is consider as the only way to move forward.
3522 #define MAX_RECLAIM_RETRIES 16
3525 * Checks whether it makes sense to retry the reclaim to make a forward progress
3526 * for the given allocation request.
3527 * The reclaim feedback represented by did_some_progress (any progress during
3528 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3529 * any progress in a row) is considered as well as the reclaimable pages on the
3530 * applicable zone list (with a backoff mechanism which is a function of
3531 * no_progress_loops).
3533 * Returns true if a retry is viable or false to enter the oom path.
3536 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3537 struct alloc_context *ac, int alloc_flags,
3538 bool did_some_progress, int *no_progress_loops)
3544 * Costly allocations might have made a progress but this doesn't mean
3545 * their order will become available due to high fragmentation so
3546 * always increment the no progress counter for them
3548 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3549 *no_progress_loops = 0;
3551 (*no_progress_loops)++;
3554 * Make sure we converge to OOM if we cannot make any progress
3555 * several times in the row.
3557 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3558 /* Before OOM, exhaust highatomic_reserve */
3559 return unreserve_highatomic_pageblock(ac, true);
3563 * Keep reclaiming pages while there is a chance this will lead
3564 * somewhere. If none of the target zones can satisfy our allocation
3565 * request even if all reclaimable pages are considered then we are
3566 * screwed and have to go OOM.
3568 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3570 unsigned long available;
3571 unsigned long reclaimable;
3572 unsigned long min_wmark = min_wmark_pages(zone);
3575 available = reclaimable = zone_reclaimable_pages(zone);
3576 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3577 MAX_RECLAIM_RETRIES);
3578 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3581 * Would the allocation succeed if we reclaimed the whole
3584 wmark = __zone_watermark_ok(zone, order, min_wmark,
3585 ac_classzone_idx(ac), alloc_flags, available);
3586 trace_reclaim_retry_zone(z, order, reclaimable,
3587 available, min_wmark, *no_progress_loops, wmark);
3590 * If we didn't make any progress and have a lot of
3591 * dirty + writeback pages then we should wait for
3592 * an IO to complete to slow down the reclaim and
3593 * prevent from pre mature OOM
3595 if (!did_some_progress) {
3596 unsigned long write_pending;
3598 write_pending = zone_page_state_snapshot(zone,
3599 NR_ZONE_WRITE_PENDING);
3601 if (2 * write_pending > reclaimable) {
3602 congestion_wait(BLK_RW_ASYNC, HZ/10);
3608 * Memory allocation/reclaim might be called from a WQ
3609 * context and the current implementation of the WQ
3610 * concurrency control doesn't recognize that
3611 * a particular WQ is congested if the worker thread is
3612 * looping without ever sleeping. Therefore we have to
3613 * do a short sleep here rather than calling
3616 if (current->flags & PF_WQ_WORKER)
3617 schedule_timeout_uninterruptible(1);
3628 static inline struct page *
3629 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3630 struct alloc_context *ac)
3632 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3633 struct page *page = NULL;
3634 unsigned int alloc_flags;
3635 unsigned long did_some_progress;
3636 enum compact_priority compact_priority;
3637 enum compact_result compact_result;
3638 int compaction_retries;
3639 int no_progress_loops;
3640 unsigned long alloc_start = jiffies;
3641 unsigned int stall_timeout = 10 * HZ;
3642 unsigned int cpuset_mems_cookie;
3645 * In the slowpath, we sanity check order to avoid ever trying to
3646 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3647 * be using allocators in order of preference for an area that is
3650 if (order >= MAX_ORDER) {
3651 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3656 * We also sanity check to catch abuse of atomic reserves being used by
3657 * callers that are not in atomic context.
3659 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3660 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3661 gfp_mask &= ~__GFP_ATOMIC;
3664 compaction_retries = 0;
3665 no_progress_loops = 0;
3666 compact_priority = DEF_COMPACT_PRIORITY;
3667 cpuset_mems_cookie = read_mems_allowed_begin();
3670 * The fast path uses conservative alloc_flags to succeed only until
3671 * kswapd needs to be woken up, and to avoid the cost of setting up
3672 * alloc_flags precisely. So we do that now.
3674 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3677 * We need to recalculate the starting point for the zonelist iterator
3678 * because we might have used different nodemask in the fast path, or
3679 * there was a cpuset modification and we are retrying - otherwise we
3680 * could end up iterating over non-eligible zones endlessly.
3682 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3683 ac->high_zoneidx, ac->nodemask);
3684 if (!ac->preferred_zoneref->zone)
3687 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3688 wake_all_kswapds(order, ac);
3691 * The adjusted alloc_flags might result in immediate success, so try
3694 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3699 * For costly allocations, try direct compaction first, as it's likely
3700 * that we have enough base pages and don't need to reclaim. Don't try
3701 * that for allocations that are allowed to ignore watermarks, as the
3702 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3704 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3705 !gfp_pfmemalloc_allowed(gfp_mask)) {
3706 page = __alloc_pages_direct_compact(gfp_mask, order,
3708 INIT_COMPACT_PRIORITY,
3714 * Checks for costly allocations with __GFP_NORETRY, which
3715 * includes THP page fault allocations
3717 if (gfp_mask & __GFP_NORETRY) {
3719 * If compaction is deferred for high-order allocations,
3720 * it is because sync compaction recently failed. If
3721 * this is the case and the caller requested a THP
3722 * allocation, we do not want to heavily disrupt the
3723 * system, so we fail the allocation instead of entering
3726 if (compact_result == COMPACT_DEFERRED)
3730 * Looks like reclaim/compaction is worth trying, but
3731 * sync compaction could be very expensive, so keep
3732 * using async compaction.
3734 compact_priority = INIT_COMPACT_PRIORITY;
3739 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3740 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3741 wake_all_kswapds(order, ac);
3743 if (gfp_pfmemalloc_allowed(gfp_mask))
3744 alloc_flags = ALLOC_NO_WATERMARKS;
3747 * Reset the zonelist iterators if memory policies can be ignored.
3748 * These allocations are high priority and system rather than user
3751 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3752 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3753 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3754 ac->high_zoneidx, ac->nodemask);
3757 /* Attempt with potentially adjusted zonelist and alloc_flags */
3758 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3762 /* Caller is not willing to reclaim, we can't balance anything */
3763 if (!can_direct_reclaim)
3766 /* Make sure we know about allocations which stall for too long */
3767 if (time_after(jiffies, alloc_start + stall_timeout)) {
3768 warn_alloc(gfp_mask, ac->nodemask,
3769 "page allocation stalls for %ums, order:%u",
3770 jiffies_to_msecs(jiffies-alloc_start), order);
3771 stall_timeout += 10 * HZ;
3774 /* Avoid recursion of direct reclaim */
3775 if (current->flags & PF_MEMALLOC)
3778 /* Try direct reclaim and then allocating */
3779 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3780 &did_some_progress);
3784 /* Try direct compaction and then allocating */
3785 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3786 compact_priority, &compact_result);
3790 /* Do not loop if specifically requested */
3791 if (gfp_mask & __GFP_NORETRY)
3795 * Do not retry costly high order allocations unless they are
3798 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3801 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3802 did_some_progress > 0, &no_progress_loops))
3806 * It doesn't make any sense to retry for the compaction if the order-0
3807 * reclaim is not able to make any progress because the current
3808 * implementation of the compaction depends on the sufficient amount
3809 * of free memory (see __compaction_suitable)
3811 if (did_some_progress > 0 &&
3812 should_compact_retry(ac, order, alloc_flags,
3813 compact_result, &compact_priority,
3814 &compaction_retries))
3818 * It's possible we raced with cpuset update so the OOM would be
3819 * premature (see below the nopage: label for full explanation).
3821 if (read_mems_allowed_retry(cpuset_mems_cookie))
3824 /* Reclaim has failed us, start killing things */
3825 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3829 /* Avoid allocations with no watermarks from looping endlessly */
3830 if (test_thread_flag(TIF_MEMDIE))
3833 /* Retry as long as the OOM killer is making progress */
3834 if (did_some_progress) {
3835 no_progress_loops = 0;
3841 * When updating a task's mems_allowed or mempolicy nodemask, it is
3842 * possible to race with parallel threads in such a way that our
3843 * allocation can fail while the mask is being updated. If we are about
3844 * to fail, check if the cpuset changed during allocation and if so,
3847 if (read_mems_allowed_retry(cpuset_mems_cookie))
3851 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3854 if (gfp_mask & __GFP_NOFAIL) {
3856 * All existing users of the __GFP_NOFAIL are blockable, so warn
3857 * of any new users that actually require GFP_NOWAIT
3859 if (WARN_ON_ONCE(!can_direct_reclaim))
3863 * PF_MEMALLOC request from this context is rather bizarre
3864 * because we cannot reclaim anything and only can loop waiting
3865 * for somebody to do a work for us
3867 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
3870 * non failing costly orders are a hard requirement which we
3871 * are not prepared for much so let's warn about these users
3872 * so that we can identify them and convert them to something
3875 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
3878 * Help non-failing allocations by giving them access to memory
3879 * reserves but do not use ALLOC_NO_WATERMARKS because this
3880 * could deplete whole memory reserves which would just make
3881 * the situation worse
3883 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
3891 warn_alloc(gfp_mask, ac->nodemask,
3892 "page allocation failure: order:%u", order);
3897 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
3898 struct zonelist *zonelist, nodemask_t *nodemask,
3899 struct alloc_context *ac, gfp_t *alloc_mask,
3900 unsigned int *alloc_flags)
3902 ac->high_zoneidx = gfp_zone(gfp_mask);
3903 ac->zonelist = zonelist;
3904 ac->nodemask = nodemask;
3905 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
3907 if (cpusets_enabled()) {
3908 *alloc_mask |= __GFP_HARDWALL;
3910 ac->nodemask = &cpuset_current_mems_allowed;
3912 *alloc_flags |= ALLOC_CPUSET;
3915 lockdep_trace_alloc(gfp_mask);
3917 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3919 if (should_fail_alloc_page(gfp_mask, order))
3922 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
3923 *alloc_flags |= ALLOC_CMA;
3928 /* Determine whether to spread dirty pages and what the first usable zone */
3929 static inline void finalise_ac(gfp_t gfp_mask,
3930 unsigned int order, struct alloc_context *ac)
3932 /* Dirty zone balancing only done in the fast path */
3933 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3936 * The preferred zone is used for statistics but crucially it is
3937 * also used as the starting point for the zonelist iterator. It
3938 * may get reset for allocations that ignore memory policies.
3940 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3941 ac->high_zoneidx, ac->nodemask);
3945 * This is the 'heart' of the zoned buddy allocator.
3948 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3949 struct zonelist *zonelist, nodemask_t *nodemask)
3952 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3953 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3954 struct alloc_context ac = { };
3956 gfp_mask &= gfp_allowed_mask;
3957 if (!prepare_alloc_pages(gfp_mask, order, zonelist, nodemask, &ac, &alloc_mask, &alloc_flags))
3960 finalise_ac(gfp_mask, order, &ac);
3962 /* First allocation attempt */
3963 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3968 * Runtime PM, block IO and its error handling path can deadlock
3969 * because I/O on the device might not complete.
3971 alloc_mask = memalloc_noio_flags(gfp_mask);
3972 ac.spread_dirty_pages = false;
3975 * Restore the original nodemask if it was potentially replaced with
3976 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3978 if (unlikely(ac.nodemask != nodemask))
3979 ac.nodemask = nodemask;
3981 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3984 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3985 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3986 __free_pages(page, order);
3990 if (kmemcheck_enabled && page)
3991 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3993 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3997 EXPORT_SYMBOL(__alloc_pages_nodemask);
4000 * Common helper functions.
4002 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4007 * __get_free_pages() returns a 32-bit address, which cannot represent
4010 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4012 page = alloc_pages(gfp_mask, order);
4015 return (unsigned long) page_address(page);
4017 EXPORT_SYMBOL(__get_free_pages);
4019 unsigned long get_zeroed_page(gfp_t gfp_mask)
4021 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4023 EXPORT_SYMBOL(get_zeroed_page);
4025 void __free_pages(struct page *page, unsigned int order)
4027 if (put_page_testzero(page)) {
4029 free_hot_cold_page(page, false);
4031 __free_pages_ok(page, order);
4035 EXPORT_SYMBOL(__free_pages);
4037 void free_pages(unsigned long addr, unsigned int order)
4040 VM_BUG_ON(!virt_addr_valid((void *)addr));
4041 __free_pages(virt_to_page((void *)addr), order);
4045 EXPORT_SYMBOL(free_pages);
4049 * An arbitrary-length arbitrary-offset area of memory which resides
4050 * within a 0 or higher order page. Multiple fragments within that page
4051 * are individually refcounted, in the page's reference counter.
4053 * The page_frag functions below provide a simple allocation framework for
4054 * page fragments. This is used by the network stack and network device
4055 * drivers to provide a backing region of memory for use as either an
4056 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4058 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4061 struct page *page = NULL;
4062 gfp_t gfp = gfp_mask;
4064 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4065 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4067 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4068 PAGE_FRAG_CACHE_MAX_ORDER);
4069 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4071 if (unlikely(!page))
4072 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4074 nc->va = page ? page_address(page) : NULL;
4079 void __page_frag_cache_drain(struct page *page, unsigned int count)
4081 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4083 if (page_ref_sub_and_test(page, count)) {
4084 unsigned int order = compound_order(page);
4087 free_hot_cold_page(page, false);
4089 __free_pages_ok(page, order);
4092 EXPORT_SYMBOL(__page_frag_cache_drain);
4094 void *page_frag_alloc(struct page_frag_cache *nc,
4095 unsigned int fragsz, gfp_t gfp_mask)
4097 unsigned int size = PAGE_SIZE;
4101 if (unlikely(!nc->va)) {
4103 page = __page_frag_cache_refill(nc, gfp_mask);
4107 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4108 /* if size can vary use size else just use PAGE_SIZE */
4111 /* Even if we own the page, we do not use atomic_set().
4112 * This would break get_page_unless_zero() users.
4114 page_ref_add(page, size - 1);
4116 /* reset page count bias and offset to start of new frag */
4117 nc->pfmemalloc = page_is_pfmemalloc(page);
4118 nc->pagecnt_bias = size;
4122 offset = nc->offset - fragsz;
4123 if (unlikely(offset < 0)) {
4124 page = virt_to_page(nc->va);
4126 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4129 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4130 /* if size can vary use size else just use PAGE_SIZE */
4133 /* OK, page count is 0, we can safely set it */
4134 set_page_count(page, size);
4136 /* reset page count bias and offset to start of new frag */
4137 nc->pagecnt_bias = size;
4138 offset = size - fragsz;
4142 nc->offset = offset;
4144 return nc->va + offset;
4146 EXPORT_SYMBOL(page_frag_alloc);
4149 * Frees a page fragment allocated out of either a compound or order 0 page.
4151 void page_frag_free(void *addr)
4153 struct page *page = virt_to_head_page(addr);
4155 if (unlikely(put_page_testzero(page)))
4156 __free_pages_ok(page, compound_order(page));
4158 EXPORT_SYMBOL(page_frag_free);
4160 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4164 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4165 unsigned long used = addr + PAGE_ALIGN(size);
4167 split_page(virt_to_page((void *)addr), order);
4168 while (used < alloc_end) {
4173 return (void *)addr;
4177 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4178 * @size: the number of bytes to allocate
4179 * @gfp_mask: GFP flags for the allocation
4181 * This function is similar to alloc_pages(), except that it allocates the
4182 * minimum number of pages to satisfy the request. alloc_pages() can only
4183 * allocate memory in power-of-two pages.
4185 * This function is also limited by MAX_ORDER.
4187 * Memory allocated by this function must be released by free_pages_exact().
4189 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4191 unsigned int order = get_order(size);
4194 addr = __get_free_pages(gfp_mask, order);
4195 return make_alloc_exact(addr, order, size);
4197 EXPORT_SYMBOL(alloc_pages_exact);
4200 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4202 * @nid: the preferred node ID where memory should be allocated
4203 * @size: the number of bytes to allocate
4204 * @gfp_mask: GFP flags for the allocation
4206 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4209 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4211 unsigned int order = get_order(size);
4212 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4215 return make_alloc_exact((unsigned long)page_address(p), order, size);
4219 * free_pages_exact - release memory allocated via alloc_pages_exact()
4220 * @virt: the value returned by alloc_pages_exact.
4221 * @size: size of allocation, same value as passed to alloc_pages_exact().
4223 * Release the memory allocated by a previous call to alloc_pages_exact.
4225 void free_pages_exact(void *virt, size_t size)
4227 unsigned long addr = (unsigned long)virt;
4228 unsigned long end = addr + PAGE_ALIGN(size);
4230 while (addr < end) {
4235 EXPORT_SYMBOL(free_pages_exact);
4238 * nr_free_zone_pages - count number of pages beyond high watermark
4239 * @offset: The zone index of the highest zone
4241 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4242 * high watermark within all zones at or below a given zone index. For each
4243 * zone, the number of pages is calculated as:
4244 * managed_pages - high_pages
4246 static unsigned long nr_free_zone_pages(int offset)
4251 /* Just pick one node, since fallback list is circular */
4252 unsigned long sum = 0;
4254 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4256 for_each_zone_zonelist(zone, z, zonelist, offset) {
4257 unsigned long size = zone->managed_pages;
4258 unsigned long high = high_wmark_pages(zone);
4267 * nr_free_buffer_pages - count number of pages beyond high watermark
4269 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4270 * watermark within ZONE_DMA and ZONE_NORMAL.
4272 unsigned long nr_free_buffer_pages(void)
4274 return nr_free_zone_pages(gfp_zone(GFP_USER));
4276 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4279 * nr_free_pagecache_pages - count number of pages beyond high watermark
4281 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4282 * high watermark within all zones.
4284 unsigned long nr_free_pagecache_pages(void)
4286 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4289 static inline void show_node(struct zone *zone)
4291 if (IS_ENABLED(CONFIG_NUMA))
4292 printk("Node %d ", zone_to_nid(zone));
4295 long si_mem_available(void)
4298 unsigned long pagecache;
4299 unsigned long wmark_low = 0;
4300 unsigned long pages[NR_LRU_LISTS];
4304 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4305 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4308 wmark_low += zone->watermark[WMARK_LOW];
4311 * Estimate the amount of memory available for userspace allocations,
4312 * without causing swapping.
4314 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4317 * Not all the page cache can be freed, otherwise the system will
4318 * start swapping. Assume at least half of the page cache, or the
4319 * low watermark worth of cache, needs to stay.
4321 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4322 pagecache -= min(pagecache / 2, wmark_low);
4323 available += pagecache;
4326 * Part of the reclaimable slab consists of items that are in use,
4327 * and cannot be freed. Cap this estimate at the low watermark.
4329 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4330 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4336 EXPORT_SYMBOL_GPL(si_mem_available);
4338 void si_meminfo(struct sysinfo *val)
4340 val->totalram = totalram_pages;
4341 val->sharedram = global_node_page_state(NR_SHMEM);
4342 val->freeram = global_page_state(NR_FREE_PAGES);
4343 val->bufferram = nr_blockdev_pages();
4344 val->totalhigh = totalhigh_pages;
4345 val->freehigh = nr_free_highpages();
4346 val->mem_unit = PAGE_SIZE;
4349 EXPORT_SYMBOL(si_meminfo);
4352 void si_meminfo_node(struct sysinfo *val, int nid)
4354 int zone_type; /* needs to be signed */
4355 unsigned long managed_pages = 0;
4356 unsigned long managed_highpages = 0;
4357 unsigned long free_highpages = 0;
4358 pg_data_t *pgdat = NODE_DATA(nid);
4360 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4361 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4362 val->totalram = managed_pages;
4363 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4364 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4365 #ifdef CONFIG_HIGHMEM
4366 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4367 struct zone *zone = &pgdat->node_zones[zone_type];
4369 if (is_highmem(zone)) {
4370 managed_highpages += zone->managed_pages;
4371 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4374 val->totalhigh = managed_highpages;
4375 val->freehigh = free_highpages;
4377 val->totalhigh = managed_highpages;
4378 val->freehigh = free_highpages;
4380 val->mem_unit = PAGE_SIZE;
4385 * Determine whether the node should be displayed or not, depending on whether
4386 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4388 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4390 if (!(flags & SHOW_MEM_FILTER_NODES))
4394 * no node mask - aka implicit memory numa policy. Do not bother with
4395 * the synchronization - read_mems_allowed_begin - because we do not
4396 * have to be precise here.
4399 nodemask = &cpuset_current_mems_allowed;
4401 return !node_isset(nid, *nodemask);
4404 #define K(x) ((x) << (PAGE_SHIFT-10))
4406 static void show_migration_types(unsigned char type)
4408 static const char types[MIGRATE_TYPES] = {
4409 [MIGRATE_UNMOVABLE] = 'U',
4410 [MIGRATE_MOVABLE] = 'M',
4411 [MIGRATE_RECLAIMABLE] = 'E',
4412 [MIGRATE_HIGHATOMIC] = 'H',
4414 [MIGRATE_CMA] = 'C',
4416 #ifdef CONFIG_MEMORY_ISOLATION
4417 [MIGRATE_ISOLATE] = 'I',
4420 char tmp[MIGRATE_TYPES + 1];
4424 for (i = 0; i < MIGRATE_TYPES; i++) {
4425 if (type & (1 << i))
4430 printk(KERN_CONT "(%s) ", tmp);
4434 * Show free area list (used inside shift_scroll-lock stuff)
4435 * We also calculate the percentage fragmentation. We do this by counting the
4436 * memory on each free list with the exception of the first item on the list.
4439 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4442 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4444 unsigned long free_pcp = 0;
4449 for_each_populated_zone(zone) {
4450 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4453 for_each_online_cpu(cpu)
4454 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4457 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4458 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4459 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4460 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4461 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4462 " free:%lu free_pcp:%lu free_cma:%lu\n",
4463 global_node_page_state(NR_ACTIVE_ANON),
4464 global_node_page_state(NR_INACTIVE_ANON),
4465 global_node_page_state(NR_ISOLATED_ANON),
4466 global_node_page_state(NR_ACTIVE_FILE),
4467 global_node_page_state(NR_INACTIVE_FILE),
4468 global_node_page_state(NR_ISOLATED_FILE),
4469 global_node_page_state(NR_UNEVICTABLE),
4470 global_node_page_state(NR_FILE_DIRTY),
4471 global_node_page_state(NR_WRITEBACK),
4472 global_node_page_state(NR_UNSTABLE_NFS),
4473 global_page_state(NR_SLAB_RECLAIMABLE),
4474 global_page_state(NR_SLAB_UNRECLAIMABLE),
4475 global_node_page_state(NR_FILE_MAPPED),
4476 global_node_page_state(NR_SHMEM),
4477 global_page_state(NR_PAGETABLE),
4478 global_page_state(NR_BOUNCE),
4479 global_page_state(NR_FREE_PAGES),
4481 global_page_state(NR_FREE_CMA_PAGES));
4483 for_each_online_pgdat(pgdat) {
4484 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4488 " active_anon:%lukB"
4489 " inactive_anon:%lukB"
4490 " active_file:%lukB"
4491 " inactive_file:%lukB"
4492 " unevictable:%lukB"
4493 " isolated(anon):%lukB"
4494 " isolated(file):%lukB"
4499 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4501 " shmem_pmdmapped: %lukB"
4504 " writeback_tmp:%lukB"
4506 " pages_scanned:%lu"
4507 " all_unreclaimable? %s"
4510 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4511 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4512 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4513 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4514 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4515 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4516 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4517 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4518 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4519 K(node_page_state(pgdat, NR_WRITEBACK)),
4520 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4521 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4522 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4524 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4526 K(node_page_state(pgdat, NR_SHMEM)),
4527 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4528 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4529 node_page_state(pgdat, NR_PAGES_SCANNED),
4530 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4533 for_each_populated_zone(zone) {
4536 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4540 for_each_online_cpu(cpu)
4541 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4550 " active_anon:%lukB"
4551 " inactive_anon:%lukB"
4552 " active_file:%lukB"
4553 " inactive_file:%lukB"
4554 " unevictable:%lukB"
4555 " writepending:%lukB"
4559 " slab_reclaimable:%lukB"
4560 " slab_unreclaimable:%lukB"
4561 " kernel_stack:%lukB"
4569 K(zone_page_state(zone, NR_FREE_PAGES)),
4570 K(min_wmark_pages(zone)),
4571 K(low_wmark_pages(zone)),
4572 K(high_wmark_pages(zone)),
4573 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4574 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4575 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4576 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4577 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4578 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4579 K(zone->present_pages),
4580 K(zone->managed_pages),
4581 K(zone_page_state(zone, NR_MLOCK)),
4582 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4583 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4584 zone_page_state(zone, NR_KERNEL_STACK_KB),
4585 K(zone_page_state(zone, NR_PAGETABLE)),
4586 K(zone_page_state(zone, NR_BOUNCE)),
4588 K(this_cpu_read(zone->pageset->pcp.count)),
4589 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4590 printk("lowmem_reserve[]:");
4591 for (i = 0; i < MAX_NR_ZONES; i++)
4592 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4593 printk(KERN_CONT "\n");
4596 for_each_populated_zone(zone) {
4598 unsigned long nr[MAX_ORDER], flags, total = 0;
4599 unsigned char types[MAX_ORDER];
4601 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4604 printk(KERN_CONT "%s: ", zone->name);
4606 spin_lock_irqsave(&zone->lock, flags);
4607 for (order = 0; order < MAX_ORDER; order++) {
4608 struct free_area *area = &zone->free_area[order];
4611 nr[order] = area->nr_free;
4612 total += nr[order] << order;
4615 for (type = 0; type < MIGRATE_TYPES; type++) {
4616 if (!list_empty(&area->free_list[type]))
4617 types[order] |= 1 << type;
4620 spin_unlock_irqrestore(&zone->lock, flags);
4621 for (order = 0; order < MAX_ORDER; order++) {
4622 printk(KERN_CONT "%lu*%lukB ",
4623 nr[order], K(1UL) << order);
4625 show_migration_types(types[order]);
4627 printk(KERN_CONT "= %lukB\n", K(total));
4630 hugetlb_show_meminfo();
4632 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4634 show_swap_cache_info();
4637 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4639 zoneref->zone = zone;
4640 zoneref->zone_idx = zone_idx(zone);
4644 * Builds allocation fallback zone lists.
4646 * Add all populated zones of a node to the zonelist.
4648 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4652 enum zone_type zone_type = MAX_NR_ZONES;
4656 zone = pgdat->node_zones + zone_type;
4657 if (managed_zone(zone)) {
4658 zoneref_set_zone(zone,
4659 &zonelist->_zonerefs[nr_zones++]);
4660 check_highest_zone(zone_type);
4662 } while (zone_type);
4670 * 0 = automatic detection of better ordering.
4671 * 1 = order by ([node] distance, -zonetype)
4672 * 2 = order by (-zonetype, [node] distance)
4674 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4675 * the same zonelist. So only NUMA can configure this param.
4677 #define ZONELIST_ORDER_DEFAULT 0
4678 #define ZONELIST_ORDER_NODE 1
4679 #define ZONELIST_ORDER_ZONE 2
4681 /* zonelist order in the kernel.
4682 * set_zonelist_order() will set this to NODE or ZONE.
4684 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4685 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4689 /* The value user specified ....changed by config */
4690 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4691 /* string for sysctl */
4692 #define NUMA_ZONELIST_ORDER_LEN 16
4693 char numa_zonelist_order[16] = "default";
4696 * interface for configure zonelist ordering.
4697 * command line option "numa_zonelist_order"
4698 * = "[dD]efault - default, automatic configuration.
4699 * = "[nN]ode - order by node locality, then by zone within node
4700 * = "[zZ]one - order by zone, then by locality within zone
4703 static int __parse_numa_zonelist_order(char *s)
4705 if (*s == 'd' || *s == 'D') {
4706 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4707 } else if (*s == 'n' || *s == 'N') {
4708 user_zonelist_order = ZONELIST_ORDER_NODE;
4709 } else if (*s == 'z' || *s == 'Z') {
4710 user_zonelist_order = ZONELIST_ORDER_ZONE;
4712 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4718 static __init int setup_numa_zonelist_order(char *s)
4725 ret = __parse_numa_zonelist_order(s);
4727 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4731 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4734 * sysctl handler for numa_zonelist_order
4736 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4737 void __user *buffer, size_t *length,
4740 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4742 static DEFINE_MUTEX(zl_order_mutex);
4744 mutex_lock(&zl_order_mutex);
4746 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4750 strcpy(saved_string, (char *)table->data);
4752 ret = proc_dostring(table, write, buffer, length, ppos);
4756 int oldval = user_zonelist_order;
4758 ret = __parse_numa_zonelist_order((char *)table->data);
4761 * bogus value. restore saved string
4763 strncpy((char *)table->data, saved_string,
4764 NUMA_ZONELIST_ORDER_LEN);
4765 user_zonelist_order = oldval;
4766 } else if (oldval != user_zonelist_order) {
4767 mutex_lock(&zonelists_mutex);
4768 build_all_zonelists(NULL, NULL);
4769 mutex_unlock(&zonelists_mutex);
4773 mutex_unlock(&zl_order_mutex);
4778 #define MAX_NODE_LOAD (nr_online_nodes)
4779 static int node_load[MAX_NUMNODES];
4782 * find_next_best_node - find the next node that should appear in a given node's fallback list
4783 * @node: node whose fallback list we're appending
4784 * @used_node_mask: nodemask_t of already used nodes
4786 * We use a number of factors to determine which is the next node that should
4787 * appear on a given node's fallback list. The node should not have appeared
4788 * already in @node's fallback list, and it should be the next closest node
4789 * according to the distance array (which contains arbitrary distance values
4790 * from each node to each node in the system), and should also prefer nodes
4791 * with no CPUs, since presumably they'll have very little allocation pressure
4792 * on them otherwise.
4793 * It returns -1 if no node is found.
4795 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4798 int min_val = INT_MAX;
4799 int best_node = NUMA_NO_NODE;
4800 const struct cpumask *tmp = cpumask_of_node(0);
4802 /* Use the local node if we haven't already */
4803 if (!node_isset(node, *used_node_mask)) {
4804 node_set(node, *used_node_mask);
4808 for_each_node_state(n, N_MEMORY) {
4810 /* Don't want a node to appear more than once */
4811 if (node_isset(n, *used_node_mask))
4814 /* Use the distance array to find the distance */
4815 val = node_distance(node, n);
4817 /* Penalize nodes under us ("prefer the next node") */
4820 /* Give preference to headless and unused nodes */
4821 tmp = cpumask_of_node(n);
4822 if (!cpumask_empty(tmp))
4823 val += PENALTY_FOR_NODE_WITH_CPUS;
4825 /* Slight preference for less loaded node */
4826 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4827 val += node_load[n];
4829 if (val < min_val) {
4836 node_set(best_node, *used_node_mask);
4843 * Build zonelists ordered by node and zones within node.
4844 * This results in maximum locality--normal zone overflows into local
4845 * DMA zone, if any--but risks exhausting DMA zone.
4847 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4850 struct zonelist *zonelist;
4852 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4853 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4855 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4856 zonelist->_zonerefs[j].zone = NULL;
4857 zonelist->_zonerefs[j].zone_idx = 0;
4861 * Build gfp_thisnode zonelists
4863 static void build_thisnode_zonelists(pg_data_t *pgdat)
4866 struct zonelist *zonelist;
4868 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4869 j = build_zonelists_node(pgdat, zonelist, 0);
4870 zonelist->_zonerefs[j].zone = NULL;
4871 zonelist->_zonerefs[j].zone_idx = 0;
4875 * Build zonelists ordered by zone and nodes within zones.
4876 * This results in conserving DMA zone[s] until all Normal memory is
4877 * exhausted, but results in overflowing to remote node while memory
4878 * may still exist in local DMA zone.
4880 static int node_order[MAX_NUMNODES];
4882 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4885 int zone_type; /* needs to be signed */
4887 struct zonelist *zonelist;
4889 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4891 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4892 for (j = 0; j < nr_nodes; j++) {
4893 node = node_order[j];
4894 z = &NODE_DATA(node)->node_zones[zone_type];
4895 if (managed_zone(z)) {
4897 &zonelist->_zonerefs[pos++]);
4898 check_highest_zone(zone_type);
4902 zonelist->_zonerefs[pos].zone = NULL;
4903 zonelist->_zonerefs[pos].zone_idx = 0;
4906 #if defined(CONFIG_64BIT)
4908 * Devices that require DMA32/DMA are relatively rare and do not justify a
4909 * penalty to every machine in case the specialised case applies. Default
4910 * to Node-ordering on 64-bit NUMA machines
4912 static int default_zonelist_order(void)
4914 return ZONELIST_ORDER_NODE;
4918 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4919 * by the kernel. If processes running on node 0 deplete the low memory zone
4920 * then reclaim will occur more frequency increasing stalls and potentially
4921 * be easier to OOM if a large percentage of the zone is under writeback or
4922 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4923 * Hence, default to zone ordering on 32-bit.
4925 static int default_zonelist_order(void)
4927 return ZONELIST_ORDER_ZONE;
4929 #endif /* CONFIG_64BIT */
4931 static void set_zonelist_order(void)
4933 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4934 current_zonelist_order = default_zonelist_order();
4936 current_zonelist_order = user_zonelist_order;
4939 static void build_zonelists(pg_data_t *pgdat)
4942 nodemask_t used_mask;
4943 int local_node, prev_node;
4944 struct zonelist *zonelist;
4945 unsigned int order = current_zonelist_order;
4947 /* initialize zonelists */
4948 for (i = 0; i < MAX_ZONELISTS; i++) {
4949 zonelist = pgdat->node_zonelists + i;
4950 zonelist->_zonerefs[0].zone = NULL;
4951 zonelist->_zonerefs[0].zone_idx = 0;
4954 /* NUMA-aware ordering of nodes */
4955 local_node = pgdat->node_id;
4956 load = nr_online_nodes;
4957 prev_node = local_node;
4958 nodes_clear(used_mask);
4960 memset(node_order, 0, sizeof(node_order));
4963 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4965 * We don't want to pressure a particular node.
4966 * So adding penalty to the first node in same
4967 * distance group to make it round-robin.
4969 if (node_distance(local_node, node) !=
4970 node_distance(local_node, prev_node))
4971 node_load[node] = load;
4975 if (order == ZONELIST_ORDER_NODE)
4976 build_zonelists_in_node_order(pgdat, node);
4978 node_order[i++] = node; /* remember order */
4981 if (order == ZONELIST_ORDER_ZONE) {
4982 /* calculate node order -- i.e., DMA last! */
4983 build_zonelists_in_zone_order(pgdat, i);
4986 build_thisnode_zonelists(pgdat);
4989 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4991 * Return node id of node used for "local" allocations.
4992 * I.e., first node id of first zone in arg node's generic zonelist.
4993 * Used for initializing percpu 'numa_mem', which is used primarily
4994 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4996 int local_memory_node(int node)
5000 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5001 gfp_zone(GFP_KERNEL),
5003 return z->zone->node;
5007 static void setup_min_unmapped_ratio(void);
5008 static void setup_min_slab_ratio(void);
5009 #else /* CONFIG_NUMA */
5011 static void set_zonelist_order(void)
5013 current_zonelist_order = ZONELIST_ORDER_ZONE;
5016 static void build_zonelists(pg_data_t *pgdat)
5018 int node, local_node;
5020 struct zonelist *zonelist;
5022 local_node = pgdat->node_id;
5024 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5025 j = build_zonelists_node(pgdat, zonelist, 0);
5028 * Now we build the zonelist so that it contains the zones
5029 * of all the other nodes.
5030 * We don't want to pressure a particular node, so when
5031 * building the zones for node N, we make sure that the
5032 * zones coming right after the local ones are those from
5033 * node N+1 (modulo N)
5035 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5036 if (!node_online(node))
5038 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5040 for (node = 0; node < local_node; node++) {
5041 if (!node_online(node))
5043 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5046 zonelist->_zonerefs[j].zone = NULL;
5047 zonelist->_zonerefs[j].zone_idx = 0;
5050 #endif /* CONFIG_NUMA */
5053 * Boot pageset table. One per cpu which is going to be used for all
5054 * zones and all nodes. The parameters will be set in such a way
5055 * that an item put on a list will immediately be handed over to
5056 * the buddy list. This is safe since pageset manipulation is done
5057 * with interrupts disabled.
5059 * The boot_pagesets must be kept even after bootup is complete for
5060 * unused processors and/or zones. They do play a role for bootstrapping
5061 * hotplugged processors.
5063 * zoneinfo_show() and maybe other functions do
5064 * not check if the processor is online before following the pageset pointer.
5065 * Other parts of the kernel may not check if the zone is available.
5067 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5068 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5069 static void setup_zone_pageset(struct zone *zone);
5072 * Global mutex to protect against size modification of zonelists
5073 * as well as to serialize pageset setup for the new populated zone.
5075 DEFINE_MUTEX(zonelists_mutex);
5077 /* return values int ....just for stop_machine() */
5078 static int __build_all_zonelists(void *data)
5082 pg_data_t *self = data;
5085 memset(node_load, 0, sizeof(node_load));
5088 if (self && !node_online(self->node_id)) {
5089 build_zonelists(self);
5092 for_each_online_node(nid) {
5093 pg_data_t *pgdat = NODE_DATA(nid);
5095 build_zonelists(pgdat);
5099 * Initialize the boot_pagesets that are going to be used
5100 * for bootstrapping processors. The real pagesets for
5101 * each zone will be allocated later when the per cpu
5102 * allocator is available.
5104 * boot_pagesets are used also for bootstrapping offline
5105 * cpus if the system is already booted because the pagesets
5106 * are needed to initialize allocators on a specific cpu too.
5107 * F.e. the percpu allocator needs the page allocator which
5108 * needs the percpu allocator in order to allocate its pagesets
5109 * (a chicken-egg dilemma).
5111 for_each_possible_cpu(cpu) {
5112 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5114 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5116 * We now know the "local memory node" for each node--
5117 * i.e., the node of the first zone in the generic zonelist.
5118 * Set up numa_mem percpu variable for on-line cpus. During
5119 * boot, only the boot cpu should be on-line; we'll init the
5120 * secondary cpus' numa_mem as they come on-line. During
5121 * node/memory hotplug, we'll fixup all on-line cpus.
5123 if (cpu_online(cpu))
5124 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5131 static noinline void __init
5132 build_all_zonelists_init(void)
5134 __build_all_zonelists(NULL);
5135 mminit_verify_zonelist();
5136 cpuset_init_current_mems_allowed();
5140 * Called with zonelists_mutex held always
5141 * unless system_state == SYSTEM_BOOTING.
5143 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5144 * [we're only called with non-NULL zone through __meminit paths] and
5145 * (2) call of __init annotated helper build_all_zonelists_init
5146 * [protected by SYSTEM_BOOTING].
5148 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5150 set_zonelist_order();
5152 if (system_state == SYSTEM_BOOTING) {
5153 build_all_zonelists_init();
5155 #ifdef CONFIG_MEMORY_HOTPLUG
5157 setup_zone_pageset(zone);
5159 /* we have to stop all cpus to guarantee there is no user
5161 stop_machine(__build_all_zonelists, pgdat, NULL);
5162 /* cpuset refresh routine should be here */
5164 vm_total_pages = nr_free_pagecache_pages();
5166 * Disable grouping by mobility if the number of pages in the
5167 * system is too low to allow the mechanism to work. It would be
5168 * more accurate, but expensive to check per-zone. This check is
5169 * made on memory-hotadd so a system can start with mobility
5170 * disabled and enable it later
5172 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5173 page_group_by_mobility_disabled = 1;
5175 page_group_by_mobility_disabled = 0;
5177 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5179 zonelist_order_name[current_zonelist_order],
5180 page_group_by_mobility_disabled ? "off" : "on",
5183 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5188 * Initially all pages are reserved - free ones are freed
5189 * up by free_all_bootmem() once the early boot process is
5190 * done. Non-atomic initialization, single-pass.
5192 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5193 unsigned long start_pfn, enum memmap_context context)
5195 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5196 unsigned long end_pfn = start_pfn + size;
5197 pg_data_t *pgdat = NODE_DATA(nid);
5199 unsigned long nr_initialised = 0;
5200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5201 struct memblock_region *r = NULL, *tmp;
5204 if (highest_memmap_pfn < end_pfn - 1)
5205 highest_memmap_pfn = end_pfn - 1;
5208 * Honor reservation requested by the driver for this ZONE_DEVICE
5211 if (altmap && start_pfn == altmap->base_pfn)
5212 start_pfn += altmap->reserve;
5214 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5216 * There can be holes in boot-time mem_map[]s handed to this
5217 * function. They do not exist on hotplugged memory.
5219 if (context != MEMMAP_EARLY)
5222 if (!early_pfn_valid(pfn)) {
5223 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5225 * Skip to the pfn preceding the next valid one (or
5226 * end_pfn), such that we hit a valid pfn (or end_pfn)
5227 * on our next iteration of the loop.
5229 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5233 if (!early_pfn_in_nid(pfn, nid))
5235 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5238 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5240 * Check given memblock attribute by firmware which can affect
5241 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5242 * mirrored, it's an overlapped memmap init. skip it.
5244 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5245 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5246 for_each_memblock(memory, tmp)
5247 if (pfn < memblock_region_memory_end_pfn(tmp))
5251 if (pfn >= memblock_region_memory_base_pfn(r) &&
5252 memblock_is_mirror(r)) {
5253 /* already initialized as NORMAL */
5254 pfn = memblock_region_memory_end_pfn(r);
5262 * Mark the block movable so that blocks are reserved for
5263 * movable at startup. This will force kernel allocations
5264 * to reserve their blocks rather than leaking throughout
5265 * the address space during boot when many long-lived
5266 * kernel allocations are made.
5268 * bitmap is created for zone's valid pfn range. but memmap
5269 * can be created for invalid pages (for alignment)
5270 * check here not to call set_pageblock_migratetype() against
5273 if (!(pfn & (pageblock_nr_pages - 1))) {
5274 struct page *page = pfn_to_page(pfn);
5276 __init_single_page(page, pfn, zone, nid);
5277 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5279 __init_single_pfn(pfn, zone, nid);
5284 static void __meminit zone_init_free_lists(struct zone *zone)
5286 unsigned int order, t;
5287 for_each_migratetype_order(order, t) {
5288 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5289 zone->free_area[order].nr_free = 0;
5293 #ifndef __HAVE_ARCH_MEMMAP_INIT
5294 #define memmap_init(size, nid, zone, start_pfn) \
5295 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5298 static int zone_batchsize(struct zone *zone)
5304 * The per-cpu-pages pools are set to around 1000th of the
5305 * size of the zone. But no more than 1/2 of a meg.
5307 * OK, so we don't know how big the cache is. So guess.
5309 batch = zone->managed_pages / 1024;
5310 if (batch * PAGE_SIZE > 512 * 1024)
5311 batch = (512 * 1024) / PAGE_SIZE;
5312 batch /= 4; /* We effectively *= 4 below */
5317 * Clamp the batch to a 2^n - 1 value. Having a power
5318 * of 2 value was found to be more likely to have
5319 * suboptimal cache aliasing properties in some cases.
5321 * For example if 2 tasks are alternately allocating
5322 * batches of pages, one task can end up with a lot
5323 * of pages of one half of the possible page colors
5324 * and the other with pages of the other colors.
5326 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5331 /* The deferral and batching of frees should be suppressed under NOMMU
5334 * The problem is that NOMMU needs to be able to allocate large chunks
5335 * of contiguous memory as there's no hardware page translation to
5336 * assemble apparent contiguous memory from discontiguous pages.
5338 * Queueing large contiguous runs of pages for batching, however,
5339 * causes the pages to actually be freed in smaller chunks. As there
5340 * can be a significant delay between the individual batches being
5341 * recycled, this leads to the once large chunks of space being
5342 * fragmented and becoming unavailable for high-order allocations.
5349 * pcp->high and pcp->batch values are related and dependent on one another:
5350 * ->batch must never be higher then ->high.
5351 * The following function updates them in a safe manner without read side
5354 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5355 * those fields changing asynchronously (acording the the above rule).
5357 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5358 * outside of boot time (or some other assurance that no concurrent updaters
5361 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5362 unsigned long batch)
5364 /* start with a fail safe value for batch */
5368 /* Update high, then batch, in order */
5375 /* a companion to pageset_set_high() */
5376 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5378 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5381 static void pageset_init(struct per_cpu_pageset *p)
5383 struct per_cpu_pages *pcp;
5386 memset(p, 0, sizeof(*p));
5390 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5391 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5394 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5397 pageset_set_batch(p, batch);
5401 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5402 * to the value high for the pageset p.
5404 static void pageset_set_high(struct per_cpu_pageset *p,
5407 unsigned long batch = max(1UL, high / 4);
5408 if ((high / 4) > (PAGE_SHIFT * 8))
5409 batch = PAGE_SHIFT * 8;
5411 pageset_update(&p->pcp, high, batch);
5414 static void pageset_set_high_and_batch(struct zone *zone,
5415 struct per_cpu_pageset *pcp)
5417 if (percpu_pagelist_fraction)
5418 pageset_set_high(pcp,
5419 (zone->managed_pages /
5420 percpu_pagelist_fraction));
5422 pageset_set_batch(pcp, zone_batchsize(zone));
5425 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5427 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5430 pageset_set_high_and_batch(zone, pcp);
5433 static void __meminit setup_zone_pageset(struct zone *zone)
5436 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5437 for_each_possible_cpu(cpu)
5438 zone_pageset_init(zone, cpu);
5442 * Allocate per cpu pagesets and initialize them.
5443 * Before this call only boot pagesets were available.
5445 void __init setup_per_cpu_pageset(void)
5447 struct pglist_data *pgdat;
5450 for_each_populated_zone(zone)
5451 setup_zone_pageset(zone);
5453 for_each_online_pgdat(pgdat)
5454 pgdat->per_cpu_nodestats =
5455 alloc_percpu(struct per_cpu_nodestat);
5458 static __meminit void zone_pcp_init(struct zone *zone)
5461 * per cpu subsystem is not up at this point. The following code
5462 * relies on the ability of the linker to provide the
5463 * offset of a (static) per cpu variable into the per cpu area.
5465 zone->pageset = &boot_pageset;
5467 if (populated_zone(zone))
5468 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5469 zone->name, zone->present_pages,
5470 zone_batchsize(zone));
5473 int __meminit init_currently_empty_zone(struct zone *zone,
5474 unsigned long zone_start_pfn,
5477 struct pglist_data *pgdat = zone->zone_pgdat;
5479 pgdat->nr_zones = zone_idx(zone) + 1;
5481 zone->zone_start_pfn = zone_start_pfn;
5483 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5484 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5486 (unsigned long)zone_idx(zone),
5487 zone_start_pfn, (zone_start_pfn + size));
5489 zone_init_free_lists(zone);
5490 zone->initialized = 1;
5495 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5496 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5499 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5501 int __meminit __early_pfn_to_nid(unsigned long pfn,
5502 struct mminit_pfnnid_cache *state)
5504 unsigned long start_pfn, end_pfn;
5507 if (state->last_start <= pfn && pfn < state->last_end)
5508 return state->last_nid;
5510 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5512 state->last_start = start_pfn;
5513 state->last_end = end_pfn;
5514 state->last_nid = nid;
5519 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5522 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5523 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5524 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5526 * If an architecture guarantees that all ranges registered contain no holes
5527 * and may be freed, this this function may be used instead of calling
5528 * memblock_free_early_nid() manually.
5530 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5532 unsigned long start_pfn, end_pfn;
5535 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5536 start_pfn = min(start_pfn, max_low_pfn);
5537 end_pfn = min(end_pfn, max_low_pfn);
5539 if (start_pfn < end_pfn)
5540 memblock_free_early_nid(PFN_PHYS(start_pfn),
5541 (end_pfn - start_pfn) << PAGE_SHIFT,
5547 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5548 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5550 * If an architecture guarantees that all ranges registered contain no holes and may
5551 * be freed, this function may be used instead of calling memory_present() manually.
5553 void __init sparse_memory_present_with_active_regions(int nid)
5555 unsigned long start_pfn, end_pfn;
5558 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5559 memory_present(this_nid, start_pfn, end_pfn);
5563 * get_pfn_range_for_nid - Return the start and end page frames for a node
5564 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5565 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5566 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5568 * It returns the start and end page frame of a node based on information
5569 * provided by memblock_set_node(). If called for a node
5570 * with no available memory, a warning is printed and the start and end
5573 void __meminit get_pfn_range_for_nid(unsigned int nid,
5574 unsigned long *start_pfn, unsigned long *end_pfn)
5576 unsigned long this_start_pfn, this_end_pfn;
5582 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5583 *start_pfn = min(*start_pfn, this_start_pfn);
5584 *end_pfn = max(*end_pfn, this_end_pfn);
5587 if (*start_pfn == -1UL)
5592 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5593 * assumption is made that zones within a node are ordered in monotonic
5594 * increasing memory addresses so that the "highest" populated zone is used
5596 static void __init find_usable_zone_for_movable(void)
5599 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5600 if (zone_index == ZONE_MOVABLE)
5603 if (arch_zone_highest_possible_pfn[zone_index] >
5604 arch_zone_lowest_possible_pfn[zone_index])
5608 VM_BUG_ON(zone_index == -1);
5609 movable_zone = zone_index;
5613 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5614 * because it is sized independent of architecture. Unlike the other zones,
5615 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5616 * in each node depending on the size of each node and how evenly kernelcore
5617 * is distributed. This helper function adjusts the zone ranges
5618 * provided by the architecture for a given node by using the end of the
5619 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5620 * zones within a node are in order of monotonic increases memory addresses
5622 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5623 unsigned long zone_type,
5624 unsigned long node_start_pfn,
5625 unsigned long node_end_pfn,
5626 unsigned long *zone_start_pfn,
5627 unsigned long *zone_end_pfn)
5629 /* Only adjust if ZONE_MOVABLE is on this node */
5630 if (zone_movable_pfn[nid]) {
5631 /* Size ZONE_MOVABLE */
5632 if (zone_type == ZONE_MOVABLE) {
5633 *zone_start_pfn = zone_movable_pfn[nid];
5634 *zone_end_pfn = min(node_end_pfn,
5635 arch_zone_highest_possible_pfn[movable_zone]);
5637 /* Adjust for ZONE_MOVABLE starting within this range */
5638 } else if (!mirrored_kernelcore &&
5639 *zone_start_pfn < zone_movable_pfn[nid] &&
5640 *zone_end_pfn > zone_movable_pfn[nid]) {
5641 *zone_end_pfn = zone_movable_pfn[nid];
5643 /* Check if this whole range is within ZONE_MOVABLE */
5644 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5645 *zone_start_pfn = *zone_end_pfn;
5650 * Return the number of pages a zone spans in a node, including holes
5651 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5653 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5654 unsigned long zone_type,
5655 unsigned long node_start_pfn,
5656 unsigned long node_end_pfn,
5657 unsigned long *zone_start_pfn,
5658 unsigned long *zone_end_pfn,
5659 unsigned long *ignored)
5661 /* When hotadd a new node from cpu_up(), the node should be empty */
5662 if (!node_start_pfn && !node_end_pfn)
5665 /* Get the start and end of the zone */
5666 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5667 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5668 adjust_zone_range_for_zone_movable(nid, zone_type,
5669 node_start_pfn, node_end_pfn,
5670 zone_start_pfn, zone_end_pfn);
5672 /* Check that this node has pages within the zone's required range */
5673 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5676 /* Move the zone boundaries inside the node if necessary */
5677 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5678 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5680 /* Return the spanned pages */
5681 return *zone_end_pfn - *zone_start_pfn;
5685 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5686 * then all holes in the requested range will be accounted for.
5688 unsigned long __meminit __absent_pages_in_range(int nid,
5689 unsigned long range_start_pfn,
5690 unsigned long range_end_pfn)
5692 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5693 unsigned long start_pfn, end_pfn;
5696 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5697 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5698 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5699 nr_absent -= end_pfn - start_pfn;
5705 * absent_pages_in_range - Return number of page frames in holes within a range
5706 * @start_pfn: The start PFN to start searching for holes
5707 * @end_pfn: The end PFN to stop searching for holes
5709 * It returns the number of pages frames in memory holes within a range.
5711 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5712 unsigned long end_pfn)
5714 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5717 /* Return the number of page frames in holes in a zone on a node */
5718 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5719 unsigned long zone_type,
5720 unsigned long node_start_pfn,
5721 unsigned long node_end_pfn,
5722 unsigned long *ignored)
5724 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5725 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5726 unsigned long zone_start_pfn, zone_end_pfn;
5727 unsigned long nr_absent;
5729 /* When hotadd a new node from cpu_up(), the node should be empty */
5730 if (!node_start_pfn && !node_end_pfn)
5733 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5734 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5736 adjust_zone_range_for_zone_movable(nid, zone_type,
5737 node_start_pfn, node_end_pfn,
5738 &zone_start_pfn, &zone_end_pfn);
5739 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5742 * ZONE_MOVABLE handling.
5743 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5746 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5747 unsigned long start_pfn, end_pfn;
5748 struct memblock_region *r;
5750 for_each_memblock(memory, r) {
5751 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5752 zone_start_pfn, zone_end_pfn);
5753 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5754 zone_start_pfn, zone_end_pfn);
5756 if (zone_type == ZONE_MOVABLE &&
5757 memblock_is_mirror(r))
5758 nr_absent += end_pfn - start_pfn;
5760 if (zone_type == ZONE_NORMAL &&
5761 !memblock_is_mirror(r))
5762 nr_absent += end_pfn - start_pfn;
5769 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5770 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5771 unsigned long zone_type,
5772 unsigned long node_start_pfn,
5773 unsigned long node_end_pfn,
5774 unsigned long *zone_start_pfn,
5775 unsigned long *zone_end_pfn,
5776 unsigned long *zones_size)
5780 *zone_start_pfn = node_start_pfn;
5781 for (zone = 0; zone < zone_type; zone++)
5782 *zone_start_pfn += zones_size[zone];
5784 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5786 return zones_size[zone_type];
5789 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5790 unsigned long zone_type,
5791 unsigned long node_start_pfn,
5792 unsigned long node_end_pfn,
5793 unsigned long *zholes_size)
5798 return zholes_size[zone_type];
5801 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5803 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5804 unsigned long node_start_pfn,
5805 unsigned long node_end_pfn,
5806 unsigned long *zones_size,
5807 unsigned long *zholes_size)
5809 unsigned long realtotalpages = 0, totalpages = 0;
5812 for (i = 0; i < MAX_NR_ZONES; i++) {
5813 struct zone *zone = pgdat->node_zones + i;
5814 unsigned long zone_start_pfn, zone_end_pfn;
5815 unsigned long size, real_size;
5817 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5823 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5824 node_start_pfn, node_end_pfn,
5827 zone->zone_start_pfn = zone_start_pfn;
5829 zone->zone_start_pfn = 0;
5830 zone->spanned_pages = size;
5831 zone->present_pages = real_size;
5834 realtotalpages += real_size;
5837 pgdat->node_spanned_pages = totalpages;
5838 pgdat->node_present_pages = realtotalpages;
5839 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5843 #ifndef CONFIG_SPARSEMEM
5845 * Calculate the size of the zone->blockflags rounded to an unsigned long
5846 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5847 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5848 * round what is now in bits to nearest long in bits, then return it in
5851 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5853 unsigned long usemapsize;
5855 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5856 usemapsize = roundup(zonesize, pageblock_nr_pages);
5857 usemapsize = usemapsize >> pageblock_order;
5858 usemapsize *= NR_PAGEBLOCK_BITS;
5859 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5861 return usemapsize / 8;
5864 static void __init setup_usemap(struct pglist_data *pgdat,
5866 unsigned long zone_start_pfn,
5867 unsigned long zonesize)
5869 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5870 zone->pageblock_flags = NULL;
5872 zone->pageblock_flags =
5873 memblock_virt_alloc_node_nopanic(usemapsize,
5877 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5878 unsigned long zone_start_pfn, unsigned long zonesize) {}
5879 #endif /* CONFIG_SPARSEMEM */
5881 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5883 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5884 void __paginginit set_pageblock_order(void)
5888 /* Check that pageblock_nr_pages has not already been setup */
5889 if (pageblock_order)
5892 if (HPAGE_SHIFT > PAGE_SHIFT)
5893 order = HUGETLB_PAGE_ORDER;
5895 order = MAX_ORDER - 1;
5898 * Assume the largest contiguous order of interest is a huge page.
5899 * This value may be variable depending on boot parameters on IA64 and
5902 pageblock_order = order;
5904 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5907 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5908 * is unused as pageblock_order is set at compile-time. See
5909 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5912 void __paginginit set_pageblock_order(void)
5916 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5918 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5919 unsigned long present_pages)
5921 unsigned long pages = spanned_pages;
5924 * Provide a more accurate estimation if there are holes within
5925 * the zone and SPARSEMEM is in use. If there are holes within the
5926 * zone, each populated memory region may cost us one or two extra
5927 * memmap pages due to alignment because memmap pages for each
5928 * populated regions may not be naturally aligned on page boundary.
5929 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5931 if (spanned_pages > present_pages + (present_pages >> 4) &&
5932 IS_ENABLED(CONFIG_SPARSEMEM))
5933 pages = present_pages;
5935 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5939 * Set up the zone data structures:
5940 * - mark all pages reserved
5941 * - mark all memory queues empty
5942 * - clear the memory bitmaps
5944 * NOTE: pgdat should get zeroed by caller.
5946 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5949 int nid = pgdat->node_id;
5952 pgdat_resize_init(pgdat);
5953 #ifdef CONFIG_NUMA_BALANCING
5954 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5955 pgdat->numabalancing_migrate_nr_pages = 0;
5956 pgdat->numabalancing_migrate_next_window = jiffies;
5958 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5959 spin_lock_init(&pgdat->split_queue_lock);
5960 INIT_LIST_HEAD(&pgdat->split_queue);
5961 pgdat->split_queue_len = 0;
5963 init_waitqueue_head(&pgdat->kswapd_wait);
5964 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5965 #ifdef CONFIG_COMPACTION
5966 init_waitqueue_head(&pgdat->kcompactd_wait);
5968 pgdat_page_ext_init(pgdat);
5969 spin_lock_init(&pgdat->lru_lock);
5970 lruvec_init(node_lruvec(pgdat));
5972 for (j = 0; j < MAX_NR_ZONES; j++) {
5973 struct zone *zone = pgdat->node_zones + j;
5974 unsigned long size, realsize, freesize, memmap_pages;
5975 unsigned long zone_start_pfn = zone->zone_start_pfn;
5977 size = zone->spanned_pages;
5978 realsize = freesize = zone->present_pages;
5981 * Adjust freesize so that it accounts for how much memory
5982 * is used by this zone for memmap. This affects the watermark
5983 * and per-cpu initialisations
5985 memmap_pages = calc_memmap_size(size, realsize);
5986 if (!is_highmem_idx(j)) {
5987 if (freesize >= memmap_pages) {
5988 freesize -= memmap_pages;
5991 " %s zone: %lu pages used for memmap\n",
5992 zone_names[j], memmap_pages);
5994 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5995 zone_names[j], memmap_pages, freesize);
5998 /* Account for reserved pages */
5999 if (j == 0 && freesize > dma_reserve) {
6000 freesize -= dma_reserve;
6001 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6002 zone_names[0], dma_reserve);
6005 if (!is_highmem_idx(j))
6006 nr_kernel_pages += freesize;
6007 /* Charge for highmem memmap if there are enough kernel pages */
6008 else if (nr_kernel_pages > memmap_pages * 2)
6009 nr_kernel_pages -= memmap_pages;
6010 nr_all_pages += freesize;
6013 * Set an approximate value for lowmem here, it will be adjusted
6014 * when the bootmem allocator frees pages into the buddy system.
6015 * And all highmem pages will be managed by the buddy system.
6017 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6021 zone->name = zone_names[j];
6022 zone->zone_pgdat = pgdat;
6023 spin_lock_init(&zone->lock);
6024 zone_seqlock_init(zone);
6025 zone_pcp_init(zone);
6030 set_pageblock_order();
6031 setup_usemap(pgdat, zone, zone_start_pfn, size);
6032 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6034 memmap_init(size, nid, j, zone_start_pfn);
6038 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6040 unsigned long __maybe_unused start = 0;
6041 unsigned long __maybe_unused offset = 0;
6043 /* Skip empty nodes */
6044 if (!pgdat->node_spanned_pages)
6047 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6048 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6049 offset = pgdat->node_start_pfn - start;
6050 /* ia64 gets its own node_mem_map, before this, without bootmem */
6051 if (!pgdat->node_mem_map) {
6052 unsigned long size, end;
6056 * The zone's endpoints aren't required to be MAX_ORDER
6057 * aligned but the node_mem_map endpoints must be in order
6058 * for the buddy allocator to function correctly.
6060 end = pgdat_end_pfn(pgdat);
6061 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6062 size = (end - start) * sizeof(struct page);
6063 map = alloc_remap(pgdat->node_id, size);
6065 map = memblock_virt_alloc_node_nopanic(size,
6067 pgdat->node_mem_map = map + offset;
6069 #ifndef CONFIG_NEED_MULTIPLE_NODES
6071 * With no DISCONTIG, the global mem_map is just set as node 0's
6073 if (pgdat == NODE_DATA(0)) {
6074 mem_map = NODE_DATA(0)->node_mem_map;
6075 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6076 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6078 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6081 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6084 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6085 unsigned long node_start_pfn, unsigned long *zholes_size)
6087 pg_data_t *pgdat = NODE_DATA(nid);
6088 unsigned long start_pfn = 0;
6089 unsigned long end_pfn = 0;
6091 /* pg_data_t should be reset to zero when it's allocated */
6092 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6094 reset_deferred_meminit(pgdat);
6095 pgdat->node_id = nid;
6096 pgdat->node_start_pfn = node_start_pfn;
6097 pgdat->per_cpu_nodestats = NULL;
6098 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6099 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6100 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6101 (u64)start_pfn << PAGE_SHIFT,
6102 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6104 start_pfn = node_start_pfn;
6106 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6107 zones_size, zholes_size);
6109 alloc_node_mem_map(pgdat);
6110 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6111 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6112 nid, (unsigned long)pgdat,
6113 (unsigned long)pgdat->node_mem_map);
6116 free_area_init_core(pgdat);
6119 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6121 #if MAX_NUMNODES > 1
6123 * Figure out the number of possible node ids.
6125 void __init setup_nr_node_ids(void)
6127 unsigned int highest;
6129 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6130 nr_node_ids = highest + 1;
6135 * node_map_pfn_alignment - determine the maximum internode alignment
6137 * This function should be called after node map is populated and sorted.
6138 * It calculates the maximum power of two alignment which can distinguish
6141 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6142 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6143 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6144 * shifted, 1GiB is enough and this function will indicate so.
6146 * This is used to test whether pfn -> nid mapping of the chosen memory
6147 * model has fine enough granularity to avoid incorrect mapping for the
6148 * populated node map.
6150 * Returns the determined alignment in pfn's. 0 if there is no alignment
6151 * requirement (single node).
6153 unsigned long __init node_map_pfn_alignment(void)
6155 unsigned long accl_mask = 0, last_end = 0;
6156 unsigned long start, end, mask;
6160 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6161 if (!start || last_nid < 0 || last_nid == nid) {
6168 * Start with a mask granular enough to pin-point to the
6169 * start pfn and tick off bits one-by-one until it becomes
6170 * too coarse to separate the current node from the last.
6172 mask = ~((1 << __ffs(start)) - 1);
6173 while (mask && last_end <= (start & (mask << 1)))
6176 /* accumulate all internode masks */
6180 /* convert mask to number of pages */
6181 return ~accl_mask + 1;
6184 /* Find the lowest pfn for a node */
6185 static unsigned long __init find_min_pfn_for_node(int nid)
6187 unsigned long min_pfn = ULONG_MAX;
6188 unsigned long start_pfn;
6191 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6192 min_pfn = min(min_pfn, start_pfn);
6194 if (min_pfn == ULONG_MAX) {
6195 pr_warn("Could not find start_pfn for node %d\n", nid);
6203 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6205 * It returns the minimum PFN based on information provided via
6206 * memblock_set_node().
6208 unsigned long __init find_min_pfn_with_active_regions(void)
6210 return find_min_pfn_for_node(MAX_NUMNODES);
6214 * early_calculate_totalpages()
6215 * Sum pages in active regions for movable zone.
6216 * Populate N_MEMORY for calculating usable_nodes.
6218 static unsigned long __init early_calculate_totalpages(void)
6220 unsigned long totalpages = 0;
6221 unsigned long start_pfn, end_pfn;
6224 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6225 unsigned long pages = end_pfn - start_pfn;
6227 totalpages += pages;
6229 node_set_state(nid, N_MEMORY);
6235 * Find the PFN the Movable zone begins in each node. Kernel memory
6236 * is spread evenly between nodes as long as the nodes have enough
6237 * memory. When they don't, some nodes will have more kernelcore than
6240 static void __init find_zone_movable_pfns_for_nodes(void)
6243 unsigned long usable_startpfn;
6244 unsigned long kernelcore_node, kernelcore_remaining;
6245 /* save the state before borrow the nodemask */
6246 nodemask_t saved_node_state = node_states[N_MEMORY];
6247 unsigned long totalpages = early_calculate_totalpages();
6248 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6249 struct memblock_region *r;
6251 /* Need to find movable_zone earlier when movable_node is specified. */
6252 find_usable_zone_for_movable();
6255 * If movable_node is specified, ignore kernelcore and movablecore
6258 if (movable_node_is_enabled()) {
6259 for_each_memblock(memory, r) {
6260 if (!memblock_is_hotpluggable(r))
6265 usable_startpfn = PFN_DOWN(r->base);
6266 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6267 min(usable_startpfn, zone_movable_pfn[nid]) :
6275 * If kernelcore=mirror is specified, ignore movablecore option
6277 if (mirrored_kernelcore) {
6278 bool mem_below_4gb_not_mirrored = false;
6280 for_each_memblock(memory, r) {
6281 if (memblock_is_mirror(r))
6286 usable_startpfn = memblock_region_memory_base_pfn(r);
6288 if (usable_startpfn < 0x100000) {
6289 mem_below_4gb_not_mirrored = true;
6293 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6294 min(usable_startpfn, zone_movable_pfn[nid]) :
6298 if (mem_below_4gb_not_mirrored)
6299 pr_warn("This configuration results in unmirrored kernel memory.");
6305 * If movablecore=nn[KMG] was specified, calculate what size of
6306 * kernelcore that corresponds so that memory usable for
6307 * any allocation type is evenly spread. If both kernelcore
6308 * and movablecore are specified, then the value of kernelcore
6309 * will be used for required_kernelcore if it's greater than
6310 * what movablecore would have allowed.
6312 if (required_movablecore) {
6313 unsigned long corepages;
6316 * Round-up so that ZONE_MOVABLE is at least as large as what
6317 * was requested by the user
6319 required_movablecore =
6320 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6321 required_movablecore = min(totalpages, required_movablecore);
6322 corepages = totalpages - required_movablecore;
6324 required_kernelcore = max(required_kernelcore, corepages);
6328 * If kernelcore was not specified or kernelcore size is larger
6329 * than totalpages, there is no ZONE_MOVABLE.
6331 if (!required_kernelcore || required_kernelcore >= totalpages)
6334 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6335 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6338 /* Spread kernelcore memory as evenly as possible throughout nodes */
6339 kernelcore_node = required_kernelcore / usable_nodes;
6340 for_each_node_state(nid, N_MEMORY) {
6341 unsigned long start_pfn, end_pfn;
6344 * Recalculate kernelcore_node if the division per node
6345 * now exceeds what is necessary to satisfy the requested
6346 * amount of memory for the kernel
6348 if (required_kernelcore < kernelcore_node)
6349 kernelcore_node = required_kernelcore / usable_nodes;
6352 * As the map is walked, we track how much memory is usable
6353 * by the kernel using kernelcore_remaining. When it is
6354 * 0, the rest of the node is usable by ZONE_MOVABLE
6356 kernelcore_remaining = kernelcore_node;
6358 /* Go through each range of PFNs within this node */
6359 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6360 unsigned long size_pages;
6362 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6363 if (start_pfn >= end_pfn)
6366 /* Account for what is only usable for kernelcore */
6367 if (start_pfn < usable_startpfn) {
6368 unsigned long kernel_pages;
6369 kernel_pages = min(end_pfn, usable_startpfn)
6372 kernelcore_remaining -= min(kernel_pages,
6373 kernelcore_remaining);
6374 required_kernelcore -= min(kernel_pages,
6375 required_kernelcore);
6377 /* Continue if range is now fully accounted */
6378 if (end_pfn <= usable_startpfn) {
6381 * Push zone_movable_pfn to the end so
6382 * that if we have to rebalance
6383 * kernelcore across nodes, we will
6384 * not double account here
6386 zone_movable_pfn[nid] = end_pfn;
6389 start_pfn = usable_startpfn;
6393 * The usable PFN range for ZONE_MOVABLE is from
6394 * start_pfn->end_pfn. Calculate size_pages as the
6395 * number of pages used as kernelcore
6397 size_pages = end_pfn - start_pfn;
6398 if (size_pages > kernelcore_remaining)
6399 size_pages = kernelcore_remaining;
6400 zone_movable_pfn[nid] = start_pfn + size_pages;
6403 * Some kernelcore has been met, update counts and
6404 * break if the kernelcore for this node has been
6407 required_kernelcore -= min(required_kernelcore,
6409 kernelcore_remaining -= size_pages;
6410 if (!kernelcore_remaining)
6416 * If there is still required_kernelcore, we do another pass with one
6417 * less node in the count. This will push zone_movable_pfn[nid] further
6418 * along on the nodes that still have memory until kernelcore is
6422 if (usable_nodes && required_kernelcore > usable_nodes)
6426 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6427 for (nid = 0; nid < MAX_NUMNODES; nid++)
6428 zone_movable_pfn[nid] =
6429 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6432 /* restore the node_state */
6433 node_states[N_MEMORY] = saved_node_state;
6436 /* Any regular or high memory on that node ? */
6437 static void check_for_memory(pg_data_t *pgdat, int nid)
6439 enum zone_type zone_type;
6441 if (N_MEMORY == N_NORMAL_MEMORY)
6444 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6445 struct zone *zone = &pgdat->node_zones[zone_type];
6446 if (populated_zone(zone)) {
6447 node_set_state(nid, N_HIGH_MEMORY);
6448 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6449 zone_type <= ZONE_NORMAL)
6450 node_set_state(nid, N_NORMAL_MEMORY);
6457 * free_area_init_nodes - Initialise all pg_data_t and zone data
6458 * @max_zone_pfn: an array of max PFNs for each zone
6460 * This will call free_area_init_node() for each active node in the system.
6461 * Using the page ranges provided by memblock_set_node(), the size of each
6462 * zone in each node and their holes is calculated. If the maximum PFN
6463 * between two adjacent zones match, it is assumed that the zone is empty.
6464 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6465 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6466 * starts where the previous one ended. For example, ZONE_DMA32 starts
6467 * at arch_max_dma_pfn.
6469 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6471 unsigned long start_pfn, end_pfn;
6474 /* Record where the zone boundaries are */
6475 memset(arch_zone_lowest_possible_pfn, 0,
6476 sizeof(arch_zone_lowest_possible_pfn));
6477 memset(arch_zone_highest_possible_pfn, 0,
6478 sizeof(arch_zone_highest_possible_pfn));
6480 start_pfn = find_min_pfn_with_active_regions();
6482 for (i = 0; i < MAX_NR_ZONES; i++) {
6483 if (i == ZONE_MOVABLE)
6486 end_pfn = max(max_zone_pfn[i], start_pfn);
6487 arch_zone_lowest_possible_pfn[i] = start_pfn;
6488 arch_zone_highest_possible_pfn[i] = end_pfn;
6490 start_pfn = end_pfn;
6493 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6494 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6495 find_zone_movable_pfns_for_nodes();
6497 /* Print out the zone ranges */
6498 pr_info("Zone ranges:\n");
6499 for (i = 0; i < MAX_NR_ZONES; i++) {
6500 if (i == ZONE_MOVABLE)
6502 pr_info(" %-8s ", zone_names[i]);
6503 if (arch_zone_lowest_possible_pfn[i] ==
6504 arch_zone_highest_possible_pfn[i])
6507 pr_cont("[mem %#018Lx-%#018Lx]\n",
6508 (u64)arch_zone_lowest_possible_pfn[i]
6510 ((u64)arch_zone_highest_possible_pfn[i]
6511 << PAGE_SHIFT) - 1);
6514 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6515 pr_info("Movable zone start for each node\n");
6516 for (i = 0; i < MAX_NUMNODES; i++) {
6517 if (zone_movable_pfn[i])
6518 pr_info(" Node %d: %#018Lx\n", i,
6519 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6522 /* Print out the early node map */
6523 pr_info("Early memory node ranges\n");
6524 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6525 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6526 (u64)start_pfn << PAGE_SHIFT,
6527 ((u64)end_pfn << PAGE_SHIFT) - 1);
6529 /* Initialise every node */
6530 mminit_verify_pageflags_layout();
6531 setup_nr_node_ids();
6532 for_each_online_node(nid) {
6533 pg_data_t *pgdat = NODE_DATA(nid);
6534 free_area_init_node(nid, NULL,
6535 find_min_pfn_for_node(nid), NULL);
6537 /* Any memory on that node */
6538 if (pgdat->node_present_pages)
6539 node_set_state(nid, N_MEMORY);
6540 check_for_memory(pgdat, nid);
6544 static int __init cmdline_parse_core(char *p, unsigned long *core)
6546 unsigned long long coremem;
6550 coremem = memparse(p, &p);
6551 *core = coremem >> PAGE_SHIFT;
6553 /* Paranoid check that UL is enough for the coremem value */
6554 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6560 * kernelcore=size sets the amount of memory for use for allocations that
6561 * cannot be reclaimed or migrated.
6563 static int __init cmdline_parse_kernelcore(char *p)
6565 /* parse kernelcore=mirror */
6566 if (parse_option_str(p, "mirror")) {
6567 mirrored_kernelcore = true;
6571 return cmdline_parse_core(p, &required_kernelcore);
6575 * movablecore=size sets the amount of memory for use for allocations that
6576 * can be reclaimed or migrated.
6578 static int __init cmdline_parse_movablecore(char *p)
6580 return cmdline_parse_core(p, &required_movablecore);
6583 early_param("kernelcore", cmdline_parse_kernelcore);
6584 early_param("movablecore", cmdline_parse_movablecore);
6586 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6588 void adjust_managed_page_count(struct page *page, long count)
6590 spin_lock(&managed_page_count_lock);
6591 page_zone(page)->managed_pages += count;
6592 totalram_pages += count;
6593 #ifdef CONFIG_HIGHMEM
6594 if (PageHighMem(page))
6595 totalhigh_pages += count;
6597 spin_unlock(&managed_page_count_lock);
6599 EXPORT_SYMBOL(adjust_managed_page_count);
6601 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6604 unsigned long pages = 0;
6606 start = (void *)PAGE_ALIGN((unsigned long)start);
6607 end = (void *)((unsigned long)end & PAGE_MASK);
6608 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6609 if ((unsigned int)poison <= 0xFF)
6610 memset(pos, poison, PAGE_SIZE);
6611 free_reserved_page(virt_to_page(pos));
6615 pr_info("Freeing %s memory: %ldK\n",
6616 s, pages << (PAGE_SHIFT - 10));
6620 EXPORT_SYMBOL(free_reserved_area);
6622 #ifdef CONFIG_HIGHMEM
6623 void free_highmem_page(struct page *page)
6625 __free_reserved_page(page);
6627 page_zone(page)->managed_pages++;
6633 void __init mem_init_print_info(const char *str)
6635 unsigned long physpages, codesize, datasize, rosize, bss_size;
6636 unsigned long init_code_size, init_data_size;
6638 physpages = get_num_physpages();
6639 codesize = _etext - _stext;
6640 datasize = _edata - _sdata;
6641 rosize = __end_rodata - __start_rodata;
6642 bss_size = __bss_stop - __bss_start;
6643 init_data_size = __init_end - __init_begin;
6644 init_code_size = _einittext - _sinittext;
6647 * Detect special cases and adjust section sizes accordingly:
6648 * 1) .init.* may be embedded into .data sections
6649 * 2) .init.text.* may be out of [__init_begin, __init_end],
6650 * please refer to arch/tile/kernel/vmlinux.lds.S.
6651 * 3) .rodata.* may be embedded into .text or .data sections.
6653 #define adj_init_size(start, end, size, pos, adj) \
6655 if (start <= pos && pos < end && size > adj) \
6659 adj_init_size(__init_begin, __init_end, init_data_size,
6660 _sinittext, init_code_size);
6661 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6662 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6663 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6664 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6666 #undef adj_init_size
6668 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6669 #ifdef CONFIG_HIGHMEM
6673 nr_free_pages() << (PAGE_SHIFT - 10),
6674 physpages << (PAGE_SHIFT - 10),
6675 codesize >> 10, datasize >> 10, rosize >> 10,
6676 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6677 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6678 totalcma_pages << (PAGE_SHIFT - 10),
6679 #ifdef CONFIG_HIGHMEM
6680 totalhigh_pages << (PAGE_SHIFT - 10),
6682 str ? ", " : "", str ? str : "");
6686 * set_dma_reserve - set the specified number of pages reserved in the first zone
6687 * @new_dma_reserve: The number of pages to mark reserved
6689 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6690 * In the DMA zone, a significant percentage may be consumed by kernel image
6691 * and other unfreeable allocations which can skew the watermarks badly. This
6692 * function may optionally be used to account for unfreeable pages in the
6693 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6694 * smaller per-cpu batchsize.
6696 void __init set_dma_reserve(unsigned long new_dma_reserve)
6698 dma_reserve = new_dma_reserve;
6701 void __init free_area_init(unsigned long *zones_size)
6703 free_area_init_node(0, zones_size,
6704 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6707 static int page_alloc_cpu_dead(unsigned int cpu)
6710 lru_add_drain_cpu(cpu);
6714 * Spill the event counters of the dead processor
6715 * into the current processors event counters.
6716 * This artificially elevates the count of the current
6719 vm_events_fold_cpu(cpu);
6722 * Zero the differential counters of the dead processor
6723 * so that the vm statistics are consistent.
6725 * This is only okay since the processor is dead and cannot
6726 * race with what we are doing.
6728 cpu_vm_stats_fold(cpu);
6732 void __init page_alloc_init(void)
6736 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6737 "mm/page_alloc:dead", NULL,
6738 page_alloc_cpu_dead);
6743 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6744 * or min_free_kbytes changes.
6746 static void calculate_totalreserve_pages(void)
6748 struct pglist_data *pgdat;
6749 unsigned long reserve_pages = 0;
6750 enum zone_type i, j;
6752 for_each_online_pgdat(pgdat) {
6754 pgdat->totalreserve_pages = 0;
6756 for (i = 0; i < MAX_NR_ZONES; i++) {
6757 struct zone *zone = pgdat->node_zones + i;
6760 /* Find valid and maximum lowmem_reserve in the zone */
6761 for (j = i; j < MAX_NR_ZONES; j++) {
6762 if (zone->lowmem_reserve[j] > max)
6763 max = zone->lowmem_reserve[j];
6766 /* we treat the high watermark as reserved pages. */
6767 max += high_wmark_pages(zone);
6769 if (max > zone->managed_pages)
6770 max = zone->managed_pages;
6772 pgdat->totalreserve_pages += max;
6774 reserve_pages += max;
6777 totalreserve_pages = reserve_pages;
6781 * setup_per_zone_lowmem_reserve - called whenever
6782 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6783 * has a correct pages reserved value, so an adequate number of
6784 * pages are left in the zone after a successful __alloc_pages().
6786 static void setup_per_zone_lowmem_reserve(void)
6788 struct pglist_data *pgdat;
6789 enum zone_type j, idx;
6791 for_each_online_pgdat(pgdat) {
6792 for (j = 0; j < MAX_NR_ZONES; j++) {
6793 struct zone *zone = pgdat->node_zones + j;
6794 unsigned long managed_pages = zone->managed_pages;
6796 zone->lowmem_reserve[j] = 0;
6800 struct zone *lower_zone;
6804 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6805 sysctl_lowmem_reserve_ratio[idx] = 1;
6807 lower_zone = pgdat->node_zones + idx;
6808 lower_zone->lowmem_reserve[j] = managed_pages /
6809 sysctl_lowmem_reserve_ratio[idx];
6810 managed_pages += lower_zone->managed_pages;
6815 /* update totalreserve_pages */
6816 calculate_totalreserve_pages();
6819 static void __setup_per_zone_wmarks(void)
6821 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6822 unsigned long lowmem_pages = 0;
6824 unsigned long flags;
6826 /* Calculate total number of !ZONE_HIGHMEM pages */
6827 for_each_zone(zone) {
6828 if (!is_highmem(zone))
6829 lowmem_pages += zone->managed_pages;
6832 for_each_zone(zone) {
6835 spin_lock_irqsave(&zone->lock, flags);
6836 tmp = (u64)pages_min * zone->managed_pages;
6837 do_div(tmp, lowmem_pages);
6838 if (is_highmem(zone)) {
6840 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6841 * need highmem pages, so cap pages_min to a small
6844 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6845 * deltas control asynch page reclaim, and so should
6846 * not be capped for highmem.
6848 unsigned long min_pages;
6850 min_pages = zone->managed_pages / 1024;
6851 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6852 zone->watermark[WMARK_MIN] = min_pages;
6855 * If it's a lowmem zone, reserve a number of pages
6856 * proportionate to the zone's size.
6858 zone->watermark[WMARK_MIN] = tmp;
6862 * Set the kswapd watermarks distance according to the
6863 * scale factor in proportion to available memory, but
6864 * ensure a minimum size on small systems.
6866 tmp = max_t(u64, tmp >> 2,
6867 mult_frac(zone->managed_pages,
6868 watermark_scale_factor, 10000));
6870 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6871 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6873 spin_unlock_irqrestore(&zone->lock, flags);
6876 /* update totalreserve_pages */
6877 calculate_totalreserve_pages();
6881 * setup_per_zone_wmarks - called when min_free_kbytes changes
6882 * or when memory is hot-{added|removed}
6884 * Ensures that the watermark[min,low,high] values for each zone are set
6885 * correctly with respect to min_free_kbytes.
6887 void setup_per_zone_wmarks(void)
6889 mutex_lock(&zonelists_mutex);
6890 __setup_per_zone_wmarks();
6891 mutex_unlock(&zonelists_mutex);
6895 * Initialise min_free_kbytes.
6897 * For small machines we want it small (128k min). For large machines
6898 * we want it large (64MB max). But it is not linear, because network
6899 * bandwidth does not increase linearly with machine size. We use
6901 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6902 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6918 int __meminit init_per_zone_wmark_min(void)
6920 unsigned long lowmem_kbytes;
6921 int new_min_free_kbytes;
6923 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6924 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6926 if (new_min_free_kbytes > user_min_free_kbytes) {
6927 min_free_kbytes = new_min_free_kbytes;
6928 if (min_free_kbytes < 128)
6929 min_free_kbytes = 128;
6930 if (min_free_kbytes > 65536)
6931 min_free_kbytes = 65536;
6933 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6934 new_min_free_kbytes, user_min_free_kbytes);
6936 setup_per_zone_wmarks();
6937 refresh_zone_stat_thresholds();
6938 setup_per_zone_lowmem_reserve();
6941 setup_min_unmapped_ratio();
6942 setup_min_slab_ratio();
6947 core_initcall(init_per_zone_wmark_min)
6950 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6951 * that we can call two helper functions whenever min_free_kbytes
6954 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6955 void __user *buffer, size_t *length, loff_t *ppos)
6959 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6964 user_min_free_kbytes = min_free_kbytes;
6965 setup_per_zone_wmarks();
6970 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6971 void __user *buffer, size_t *length, loff_t *ppos)
6975 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6980 setup_per_zone_wmarks();
6986 static void setup_min_unmapped_ratio(void)
6991 for_each_online_pgdat(pgdat)
6992 pgdat->min_unmapped_pages = 0;
6995 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6996 sysctl_min_unmapped_ratio) / 100;
7000 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7001 void __user *buffer, size_t *length, loff_t *ppos)
7005 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7009 setup_min_unmapped_ratio();
7014 static void setup_min_slab_ratio(void)
7019 for_each_online_pgdat(pgdat)
7020 pgdat->min_slab_pages = 0;
7023 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7024 sysctl_min_slab_ratio) / 100;
7027 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7028 void __user *buffer, size_t *length, loff_t *ppos)
7032 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7036 setup_min_slab_ratio();
7043 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7044 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7045 * whenever sysctl_lowmem_reserve_ratio changes.
7047 * The reserve ratio obviously has absolutely no relation with the
7048 * minimum watermarks. The lowmem reserve ratio can only make sense
7049 * if in function of the boot time zone sizes.
7051 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7052 void __user *buffer, size_t *length, loff_t *ppos)
7054 proc_dointvec_minmax(table, write, buffer, length, ppos);
7055 setup_per_zone_lowmem_reserve();
7060 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7061 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7062 * pagelist can have before it gets flushed back to buddy allocator.
7064 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7065 void __user *buffer, size_t *length, loff_t *ppos)
7068 int old_percpu_pagelist_fraction;
7071 mutex_lock(&pcp_batch_high_lock);
7072 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7074 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7075 if (!write || ret < 0)
7078 /* Sanity checking to avoid pcp imbalance */
7079 if (percpu_pagelist_fraction &&
7080 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7081 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7087 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7090 for_each_populated_zone(zone) {
7093 for_each_possible_cpu(cpu)
7094 pageset_set_high_and_batch(zone,
7095 per_cpu_ptr(zone->pageset, cpu));
7098 mutex_unlock(&pcp_batch_high_lock);
7103 int hashdist = HASHDIST_DEFAULT;
7105 static int __init set_hashdist(char *str)
7109 hashdist = simple_strtoul(str, &str, 0);
7112 __setup("hashdist=", set_hashdist);
7115 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7117 * Returns the number of pages that arch has reserved but
7118 * is not known to alloc_large_system_hash().
7120 static unsigned long __init arch_reserved_kernel_pages(void)
7127 * allocate a large system hash table from bootmem
7128 * - it is assumed that the hash table must contain an exact power-of-2
7129 * quantity of entries
7130 * - limit is the number of hash buckets, not the total allocation size
7132 void *__init alloc_large_system_hash(const char *tablename,
7133 unsigned long bucketsize,
7134 unsigned long numentries,
7137 unsigned int *_hash_shift,
7138 unsigned int *_hash_mask,
7139 unsigned long low_limit,
7140 unsigned long high_limit)
7142 unsigned long long max = high_limit;
7143 unsigned long log2qty, size;
7146 /* allow the kernel cmdline to have a say */
7148 /* round applicable memory size up to nearest megabyte */
7149 numentries = nr_kernel_pages;
7150 numentries -= arch_reserved_kernel_pages();
7152 /* It isn't necessary when PAGE_SIZE >= 1MB */
7153 if (PAGE_SHIFT < 20)
7154 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7156 /* limit to 1 bucket per 2^scale bytes of low memory */
7157 if (scale > PAGE_SHIFT)
7158 numentries >>= (scale - PAGE_SHIFT);
7160 numentries <<= (PAGE_SHIFT - scale);
7162 /* Make sure we've got at least a 0-order allocation.. */
7163 if (unlikely(flags & HASH_SMALL)) {
7164 /* Makes no sense without HASH_EARLY */
7165 WARN_ON(!(flags & HASH_EARLY));
7166 if (!(numentries >> *_hash_shift)) {
7167 numentries = 1UL << *_hash_shift;
7168 BUG_ON(!numentries);
7170 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7171 numentries = PAGE_SIZE / bucketsize;
7173 numentries = roundup_pow_of_two(numentries);
7175 /* limit allocation size to 1/16 total memory by default */
7177 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7178 do_div(max, bucketsize);
7180 max = min(max, 0x80000000ULL);
7182 if (numentries < low_limit)
7183 numentries = low_limit;
7184 if (numentries > max)
7187 log2qty = ilog2(numentries);
7190 size = bucketsize << log2qty;
7191 if (flags & HASH_EARLY)
7192 table = memblock_virt_alloc_nopanic(size, 0);
7194 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7197 * If bucketsize is not a power-of-two, we may free
7198 * some pages at the end of hash table which
7199 * alloc_pages_exact() automatically does
7201 if (get_order(size) < MAX_ORDER) {
7202 table = alloc_pages_exact(size, GFP_ATOMIC);
7203 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7206 } while (!table && size > PAGE_SIZE && --log2qty);
7209 panic("Failed to allocate %s hash table\n", tablename);
7211 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7212 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7215 *_hash_shift = log2qty;
7217 *_hash_mask = (1 << log2qty) - 1;
7223 * This function checks whether pageblock includes unmovable pages or not.
7224 * If @count is not zero, it is okay to include less @count unmovable pages
7226 * PageLRU check without isolation or lru_lock could race so that
7227 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7228 * check without lock_page also may miss some movable non-lru pages at
7229 * race condition. So you can't expect this function should be exact.
7231 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7232 bool skip_hwpoisoned_pages)
7234 unsigned long pfn, iter, found;
7238 * For avoiding noise data, lru_add_drain_all() should be called
7239 * If ZONE_MOVABLE, the zone never contains unmovable pages
7241 if (zone_idx(zone) == ZONE_MOVABLE)
7243 mt = get_pageblock_migratetype(page);
7244 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7247 pfn = page_to_pfn(page);
7248 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7249 unsigned long check = pfn + iter;
7251 if (!pfn_valid_within(check))
7254 page = pfn_to_page(check);
7257 * Hugepages are not in LRU lists, but they're movable.
7258 * We need not scan over tail pages bacause we don't
7259 * handle each tail page individually in migration.
7261 if (PageHuge(page)) {
7262 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7267 * We can't use page_count without pin a page
7268 * because another CPU can free compound page.
7269 * This check already skips compound tails of THP
7270 * because their page->_refcount is zero at all time.
7272 if (!page_ref_count(page)) {
7273 if (PageBuddy(page))
7274 iter += (1 << page_order(page)) - 1;
7279 * The HWPoisoned page may be not in buddy system, and
7280 * page_count() is not 0.
7282 if (skip_hwpoisoned_pages && PageHWPoison(page))
7285 if (__PageMovable(page))
7291 * If there are RECLAIMABLE pages, we need to check
7292 * it. But now, memory offline itself doesn't call
7293 * shrink_node_slabs() and it still to be fixed.
7296 * If the page is not RAM, page_count()should be 0.
7297 * we don't need more check. This is an _used_ not-movable page.
7299 * The problematic thing here is PG_reserved pages. PG_reserved
7300 * is set to both of a memory hole page and a _used_ kernel
7309 bool is_pageblock_removable_nolock(struct page *page)
7315 * We have to be careful here because we are iterating over memory
7316 * sections which are not zone aware so we might end up outside of
7317 * the zone but still within the section.
7318 * We have to take care about the node as well. If the node is offline
7319 * its NODE_DATA will be NULL - see page_zone.
7321 if (!node_online(page_to_nid(page)))
7324 zone = page_zone(page);
7325 pfn = page_to_pfn(page);
7326 if (!zone_spans_pfn(zone, pfn))
7329 return !has_unmovable_pages(zone, page, 0, true);
7332 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7334 static unsigned long pfn_max_align_down(unsigned long pfn)
7336 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7337 pageblock_nr_pages) - 1);
7340 static unsigned long pfn_max_align_up(unsigned long pfn)
7342 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7343 pageblock_nr_pages));
7346 /* [start, end) must belong to a single zone. */
7347 static int __alloc_contig_migrate_range(struct compact_control *cc,
7348 unsigned long start, unsigned long end)
7350 /* This function is based on compact_zone() from compaction.c. */
7351 unsigned long nr_reclaimed;
7352 unsigned long pfn = start;
7353 unsigned int tries = 0;
7358 while (pfn < end || !list_empty(&cc->migratepages)) {
7359 if (fatal_signal_pending(current)) {
7364 if (list_empty(&cc->migratepages)) {
7365 cc->nr_migratepages = 0;
7366 pfn = isolate_migratepages_range(cc, pfn, end);
7372 } else if (++tries == 5) {
7373 ret = ret < 0 ? ret : -EBUSY;
7377 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7379 cc->nr_migratepages -= nr_reclaimed;
7381 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7382 NULL, 0, cc->mode, MR_CMA);
7385 putback_movable_pages(&cc->migratepages);
7392 * alloc_contig_range() -- tries to allocate given range of pages
7393 * @start: start PFN to allocate
7394 * @end: one-past-the-last PFN to allocate
7395 * @migratetype: migratetype of the underlaying pageblocks (either
7396 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7397 * in range must have the same migratetype and it must
7398 * be either of the two.
7399 * @gfp_mask: GFP mask to use during compaction
7401 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7402 * aligned, however it's the caller's responsibility to guarantee that
7403 * we are the only thread that changes migrate type of pageblocks the
7406 * The PFN range must belong to a single zone.
7408 * Returns zero on success or negative error code. On success all
7409 * pages which PFN is in [start, end) are allocated for the caller and
7410 * need to be freed with free_contig_range().
7412 int alloc_contig_range(unsigned long start, unsigned long end,
7413 unsigned migratetype, gfp_t gfp_mask)
7415 unsigned long outer_start, outer_end;
7419 struct compact_control cc = {
7420 .nr_migratepages = 0,
7422 .zone = page_zone(pfn_to_page(start)),
7423 .mode = MIGRATE_SYNC,
7424 .ignore_skip_hint = true,
7425 .gfp_mask = memalloc_noio_flags(gfp_mask),
7427 INIT_LIST_HEAD(&cc.migratepages);
7430 * What we do here is we mark all pageblocks in range as
7431 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7432 * have different sizes, and due to the way page allocator
7433 * work, we align the range to biggest of the two pages so
7434 * that page allocator won't try to merge buddies from
7435 * different pageblocks and change MIGRATE_ISOLATE to some
7436 * other migration type.
7438 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7439 * migrate the pages from an unaligned range (ie. pages that
7440 * we are interested in). This will put all the pages in
7441 * range back to page allocator as MIGRATE_ISOLATE.
7443 * When this is done, we take the pages in range from page
7444 * allocator removing them from the buddy system. This way
7445 * page allocator will never consider using them.
7447 * This lets us mark the pageblocks back as
7448 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7449 * aligned range but not in the unaligned, original range are
7450 * put back to page allocator so that buddy can use them.
7453 ret = start_isolate_page_range(pfn_max_align_down(start),
7454 pfn_max_align_up(end), migratetype,
7460 * In case of -EBUSY, we'd like to know which page causes problem.
7461 * So, just fall through. We will check it in test_pages_isolated().
7463 ret = __alloc_contig_migrate_range(&cc, start, end);
7464 if (ret && ret != -EBUSY)
7468 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7469 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7470 * more, all pages in [start, end) are free in page allocator.
7471 * What we are going to do is to allocate all pages from
7472 * [start, end) (that is remove them from page allocator).
7474 * The only problem is that pages at the beginning and at the
7475 * end of interesting range may be not aligned with pages that
7476 * page allocator holds, ie. they can be part of higher order
7477 * pages. Because of this, we reserve the bigger range and
7478 * once this is done free the pages we are not interested in.
7480 * We don't have to hold zone->lock here because the pages are
7481 * isolated thus they won't get removed from buddy.
7484 lru_add_drain_all();
7485 drain_all_pages(cc.zone);
7488 outer_start = start;
7489 while (!PageBuddy(pfn_to_page(outer_start))) {
7490 if (++order >= MAX_ORDER) {
7491 outer_start = start;
7494 outer_start &= ~0UL << order;
7497 if (outer_start != start) {
7498 order = page_order(pfn_to_page(outer_start));
7501 * outer_start page could be small order buddy page and
7502 * it doesn't include start page. Adjust outer_start
7503 * in this case to report failed page properly
7504 * on tracepoint in test_pages_isolated()
7506 if (outer_start + (1UL << order) <= start)
7507 outer_start = start;
7510 /* Make sure the range is really isolated. */
7511 if (test_pages_isolated(outer_start, end, false)) {
7512 pr_info("%s: [%lx, %lx) PFNs busy\n",
7513 __func__, outer_start, end);
7518 /* Grab isolated pages from freelists. */
7519 outer_end = isolate_freepages_range(&cc, outer_start, end);
7525 /* Free head and tail (if any) */
7526 if (start != outer_start)
7527 free_contig_range(outer_start, start - outer_start);
7528 if (end != outer_end)
7529 free_contig_range(end, outer_end - end);
7532 undo_isolate_page_range(pfn_max_align_down(start),
7533 pfn_max_align_up(end), migratetype);
7537 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7539 unsigned int count = 0;
7541 for (; nr_pages--; pfn++) {
7542 struct page *page = pfn_to_page(pfn);
7544 count += page_count(page) != 1;
7547 WARN(count != 0, "%d pages are still in use!\n", count);
7551 #ifdef CONFIG_MEMORY_HOTPLUG
7553 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7554 * page high values need to be recalulated.
7556 void __meminit zone_pcp_update(struct zone *zone)
7559 mutex_lock(&pcp_batch_high_lock);
7560 for_each_possible_cpu(cpu)
7561 pageset_set_high_and_batch(zone,
7562 per_cpu_ptr(zone->pageset, cpu));
7563 mutex_unlock(&pcp_batch_high_lock);
7567 void zone_pcp_reset(struct zone *zone)
7569 unsigned long flags;
7571 struct per_cpu_pageset *pset;
7573 /* avoid races with drain_pages() */
7574 local_irq_save(flags);
7575 if (zone->pageset != &boot_pageset) {
7576 for_each_online_cpu(cpu) {
7577 pset = per_cpu_ptr(zone->pageset, cpu);
7578 drain_zonestat(zone, pset);
7580 free_percpu(zone->pageset);
7581 zone->pageset = &boot_pageset;
7583 local_irq_restore(flags);
7586 #ifdef CONFIG_MEMORY_HOTREMOVE
7588 * All pages in the range must be in a single zone and isolated
7589 * before calling this.
7592 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7596 unsigned int order, i;
7598 unsigned long flags;
7599 /* find the first valid pfn */
7600 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7605 zone = page_zone(pfn_to_page(pfn));
7606 spin_lock_irqsave(&zone->lock, flags);
7608 while (pfn < end_pfn) {
7609 if (!pfn_valid(pfn)) {
7613 page = pfn_to_page(pfn);
7615 * The HWPoisoned page may be not in buddy system, and
7616 * page_count() is not 0.
7618 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7620 SetPageReserved(page);
7624 BUG_ON(page_count(page));
7625 BUG_ON(!PageBuddy(page));
7626 order = page_order(page);
7627 #ifdef CONFIG_DEBUG_VM
7628 pr_info("remove from free list %lx %d %lx\n",
7629 pfn, 1 << order, end_pfn);
7631 list_del(&page->lru);
7632 rmv_page_order(page);
7633 zone->free_area[order].nr_free--;
7634 for (i = 0; i < (1 << order); i++)
7635 SetPageReserved((page+i));
7636 pfn += (1 << order);
7638 spin_unlock_irqrestore(&zone->lock, flags);
7642 bool is_free_buddy_page(struct page *page)
7644 struct zone *zone = page_zone(page);
7645 unsigned long pfn = page_to_pfn(page);
7646 unsigned long flags;
7649 spin_lock_irqsave(&zone->lock, flags);
7650 for (order = 0; order < MAX_ORDER; order++) {
7651 struct page *page_head = page - (pfn & ((1 << order) - 1));
7653 if (PageBuddy(page_head) && page_order(page_head) >= order)
7656 spin_unlock_irqrestore(&zone->lock, flags);
7658 return order < MAX_ORDER;