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 <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node);
78 EXPORT_PER_CPU_SYMBOL(numa_node);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 int _node_numa_mem_[MAX_NUMNODES];
94 * Array of node states.
96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
107 [N_CPU] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page *page)
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
137 page->index = migratetype;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
181 static void __free_pages_ok(struct page *page, unsigned int order);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages);
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names[MIGRATE_TYPES] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor * const compound_page_dtors[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
252 int watermark_scale_factor = 10;
254 static unsigned long __meminitdata nr_kernel_pages;
255 static unsigned long __meminitdata nr_all_pages;
256 static unsigned long __meminitdata dma_reserve;
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
261 static unsigned long __initdata required_kernelcore;
262 static unsigned long __initdata required_movablecore;
263 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
264 static bool mirrored_kernelcore;
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
268 EXPORT_SYMBOL(movable_zone);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
272 int nr_node_ids __read_mostly = MAX_NUMNODES;
273 int nr_online_nodes __read_mostly = 1;
274 EXPORT_SYMBOL(nr_node_ids);
275 EXPORT_SYMBOL(nr_online_nodes);
278 int page_group_by_mobility_disabled __read_mostly;
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t *pgdat)
283 pgdat->first_deferred_pfn = ULONG_MAX;
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
289 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
297 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
311 unsigned long max_initialise;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
333 static inline void reset_deferred_meminit(pg_data_t *pgdat)
337 static inline bool early_page_uninitialised(unsigned long pfn)
342 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
347 static inline bool update_defer_init(pg_data_t *pgdat,
348 unsigned long pfn, unsigned long zone_end,
349 unsigned long *nr_initialised)
355 /* Return a pointer to the bitmap storing bits affecting a block of pages */
356 static inline unsigned long *get_pageblock_bitmap(struct page *page,
359 #ifdef CONFIG_SPARSEMEM
360 return __pfn_to_section(pfn)->pageblock_flags;
362 return page_zone(page)->pageblock_flags;
363 #endif /* CONFIG_SPARSEMEM */
366 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
368 #ifdef CONFIG_SPARSEMEM
369 pfn &= (PAGES_PER_SECTION-1);
370 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
372 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
373 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
374 #endif /* CONFIG_SPARSEMEM */
378 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
379 * @page: The page within the block of interest
380 * @pfn: The target page frame number
381 * @end_bitidx: The last bit of interest to retrieve
382 * @mask: mask of bits that the caller is interested in
384 * Return: pageblock_bits flags
386 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
388 unsigned long end_bitidx,
391 unsigned long *bitmap;
392 unsigned long bitidx, word_bitidx;
395 bitmap = get_pageblock_bitmap(page, pfn);
396 bitidx = pfn_to_bitidx(page, pfn);
397 word_bitidx = bitidx / BITS_PER_LONG;
398 bitidx &= (BITS_PER_LONG-1);
400 word = bitmap[word_bitidx];
401 bitidx += end_bitidx;
402 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
405 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
406 unsigned long end_bitidx,
409 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
412 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
414 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
418 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
419 * @page: The page within the block of interest
420 * @flags: The flags to set
421 * @pfn: The target page frame number
422 * @end_bitidx: The last bit of interest
423 * @mask: mask of bits that the caller is interested in
425 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
427 unsigned long end_bitidx,
430 unsigned long *bitmap;
431 unsigned long bitidx, word_bitidx;
432 unsigned long old_word, word;
434 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
436 bitmap = get_pageblock_bitmap(page, pfn);
437 bitidx = pfn_to_bitidx(page, pfn);
438 word_bitidx = bitidx / BITS_PER_LONG;
439 bitidx &= (BITS_PER_LONG-1);
441 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
443 bitidx += end_bitidx;
444 mask <<= (BITS_PER_LONG - bitidx - 1);
445 flags <<= (BITS_PER_LONG - bitidx - 1);
447 word = READ_ONCE(bitmap[word_bitidx]);
449 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
450 if (word == old_word)
456 void set_pageblock_migratetype(struct page *page, int migratetype)
458 if (unlikely(page_group_by_mobility_disabled &&
459 migratetype < MIGRATE_PCPTYPES))
460 migratetype = MIGRATE_UNMOVABLE;
462 set_pageblock_flags_group(page, (unsigned long)migratetype,
463 PB_migrate, PB_migrate_end);
466 #ifdef CONFIG_DEBUG_VM
467 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
471 unsigned long pfn = page_to_pfn(page);
472 unsigned long sp, start_pfn;
475 seq = zone_span_seqbegin(zone);
476 start_pfn = zone->zone_start_pfn;
477 sp = zone->spanned_pages;
478 if (!zone_spans_pfn(zone, pfn))
480 } while (zone_span_seqretry(zone, seq));
483 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
484 pfn, zone_to_nid(zone), zone->name,
485 start_pfn, start_pfn + sp);
490 static int page_is_consistent(struct zone *zone, struct page *page)
492 if (!pfn_valid_within(page_to_pfn(page)))
494 if (zone != page_zone(page))
500 * Temporary debugging check for pages not lying within a given zone.
502 static int bad_range(struct zone *zone, struct page *page)
504 if (page_outside_zone_boundaries(zone, page))
506 if (!page_is_consistent(zone, page))
512 static inline int bad_range(struct zone *zone, struct page *page)
518 static void bad_page(struct page *page, const char *reason,
519 unsigned long bad_flags)
521 static unsigned long resume;
522 static unsigned long nr_shown;
523 static unsigned long nr_unshown;
525 /* Don't complain about poisoned pages */
526 if (PageHWPoison(page)) {
527 page_mapcount_reset(page); /* remove PageBuddy */
532 * Allow a burst of 60 reports, then keep quiet for that minute;
533 * or allow a steady drip of one report per second.
535 if (nr_shown == 60) {
536 if (time_before(jiffies, resume)) {
542 "BUG: Bad page state: %lu messages suppressed\n",
549 resume = jiffies + 60 * HZ;
551 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
552 current->comm, page_to_pfn(page));
553 __dump_page(page, reason);
554 bad_flags &= page->flags;
556 pr_alert("bad because of flags: %#lx(%pGp)\n",
557 bad_flags, &bad_flags);
558 dump_page_owner(page);
563 /* Leave bad fields for debug, except PageBuddy could make trouble */
564 page_mapcount_reset(page); /* remove PageBuddy */
565 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
569 * Higher-order pages are called "compound pages". They are structured thusly:
571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
576 * The first tail page's ->compound_dtor holds the offset in array of compound
577 * page destructors. See compound_page_dtors.
579 * The first tail page's ->compound_order holds the order of allocation.
580 * This usage means that zero-order pages may not be compound.
583 void free_compound_page(struct page *page)
585 __free_pages_ok(page, compound_order(page));
588 void prep_compound_page(struct page *page, unsigned int order)
591 int nr_pages = 1 << order;
593 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
594 set_compound_order(page, order);
596 for (i = 1; i < nr_pages; i++) {
597 struct page *p = page + i;
598 set_page_count(p, 0);
599 p->mapping = TAIL_MAPPING;
600 set_compound_head(p, page);
602 atomic_set(compound_mapcount_ptr(page), -1);
605 #ifdef CONFIG_DEBUG_PAGEALLOC
606 unsigned int _debug_guardpage_minorder;
607 bool _debug_pagealloc_enabled __read_mostly
608 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
609 EXPORT_SYMBOL(_debug_pagealloc_enabled);
610 bool _debug_guardpage_enabled __read_mostly;
612 static int __init early_debug_pagealloc(char *buf)
617 if (strcmp(buf, "on") == 0)
618 _debug_pagealloc_enabled = true;
620 if (strcmp(buf, "off") == 0)
621 _debug_pagealloc_enabled = false;
625 early_param("debug_pagealloc", early_debug_pagealloc);
627 static bool need_debug_guardpage(void)
629 /* If we don't use debug_pagealloc, we don't need guard page */
630 if (!debug_pagealloc_enabled())
636 static void init_debug_guardpage(void)
638 if (!debug_pagealloc_enabled())
641 _debug_guardpage_enabled = true;
644 struct page_ext_operations debug_guardpage_ops = {
645 .need = need_debug_guardpage,
646 .init = init_debug_guardpage,
649 static int __init debug_guardpage_minorder_setup(char *buf)
653 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
654 pr_err("Bad debug_guardpage_minorder value\n");
657 _debug_guardpage_minorder = res;
658 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
661 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
663 static inline void set_page_guard(struct zone *zone, struct page *page,
664 unsigned int order, int migratetype)
666 struct page_ext *page_ext;
668 if (!debug_guardpage_enabled())
671 page_ext = lookup_page_ext(page);
672 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
674 INIT_LIST_HEAD(&page->lru);
675 set_page_private(page, order);
676 /* Guard pages are not available for any usage */
677 __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 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
691 set_page_private(page, 0);
692 if (!is_migrate_isolate(migratetype))
693 __mod_zone_freepage_state(zone, (1 << order), migratetype);
696 struct page_ext_operations debug_guardpage_ops = { NULL, };
697 static inline void set_page_guard(struct zone *zone, struct page *page,
698 unsigned int order, int migratetype) {}
699 static inline void clear_page_guard(struct zone *zone, struct page *page,
700 unsigned int order, int migratetype) {}
703 static inline void set_page_order(struct page *page, unsigned int order)
705 set_page_private(page, order);
706 __SetPageBuddy(page);
709 static inline void rmv_page_order(struct page *page)
711 __ClearPageBuddy(page);
712 set_page_private(page, 0);
716 * This function checks whether a page is free && is the buddy
717 * we can do coalesce a page and its buddy if
718 * (a) the buddy is not in a hole &&
719 * (b) the buddy is in the buddy system &&
720 * (c) a page and its buddy have the same order &&
721 * (d) a page and its buddy are in the same zone.
723 * For recording whether a page is in the buddy system, we set ->_mapcount
724 * PAGE_BUDDY_MAPCOUNT_VALUE.
725 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
726 * serialized by zone->lock.
728 * For recording page's order, we use page_private(page).
730 static inline int page_is_buddy(struct page *page, struct page *buddy,
733 if (!pfn_valid_within(page_to_pfn(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 page_idx;
792 unsigned long combined_idx;
793 unsigned long uninitialized_var(buddy_idx);
795 unsigned int max_order;
797 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
799 VM_BUG_ON(!zone_is_initialized(zone));
800 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
802 VM_BUG_ON(migratetype == -1);
803 if (likely(!is_migrate_isolate(migratetype)))
804 __mod_zone_freepage_state(zone, 1 << order, migratetype);
806 page_idx = pfn & ((1 << MAX_ORDER) - 1);
808 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
809 VM_BUG_ON_PAGE(bad_range(zone, page), page);
812 while (order < max_order - 1) {
813 buddy_idx = __find_buddy_index(page_idx, order);
814 buddy = page + (buddy_idx - page_idx);
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_idx = buddy_idx & page_idx;
829 page = page + (combined_idx - page_idx);
830 page_idx = combined_idx;
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_idx = __find_buddy_index(page_idx, order);
846 buddy = page + (buddy_idx - page_idx);
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(page_to_pfn(buddy))) {
870 struct page *higher_page, *higher_buddy;
871 combined_idx = buddy_idx & page_idx;
872 higher_page = page + (combined_idx - page_idx);
873 buddy_idx = __find_buddy_index(combined_idx, order + 1);
874 higher_buddy = higher_page + (buddy_idx - combined_idx);
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);
1003 kasan_free_pages(page, order);
1006 * Check tail pages before head page information is cleared to
1007 * avoid checking PageCompound for order-0 pages.
1009 if (unlikely(order)) {
1010 bool compound = PageCompound(page);
1013 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1015 for (i = 1; i < (1 << order); i++) {
1017 bad += free_tail_pages_check(page, page + i);
1018 if (unlikely(free_pages_check(page + i))) {
1022 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1025 if (PageAnonHead(page))
1026 page->mapping = NULL;
1028 bad += free_pages_check(page);
1032 page_cpupid_reset_last(page);
1033 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1034 reset_page_owner(page, order);
1036 if (!PageHighMem(page)) {
1037 debug_check_no_locks_freed(page_address(page),
1038 PAGE_SIZE << order);
1039 debug_check_no_obj_freed(page_address(page),
1040 PAGE_SIZE << order);
1042 arch_free_page(page, order);
1043 kernel_poison_pages(page, 1 << order, 0);
1044 kernel_map_pages(page, 1 << order, 0);
1049 #ifdef CONFIG_DEBUG_VM
1050 static inline bool free_pcp_prepare(struct page *page)
1052 return free_pages_prepare(page, 0, true);
1055 static inline bool bulkfree_pcp_prepare(struct page *page)
1060 static bool free_pcp_prepare(struct page *page)
1062 return free_pages_prepare(page, 0, false);
1065 static bool bulkfree_pcp_prepare(struct page *page)
1067 return free_pages_check(page);
1069 #endif /* CONFIG_DEBUG_VM */
1072 * Frees a number of pages from the PCP lists
1073 * Assumes all pages on list are in same zone, and of same order.
1074 * count is the number of pages to free.
1076 * If the zone was previously in an "all pages pinned" state then look to
1077 * see if this freeing clears that state.
1079 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1080 * pinned" detection logic.
1082 static void free_pcppages_bulk(struct zone *zone, int count,
1083 struct per_cpu_pages *pcp)
1085 int migratetype = 0;
1087 unsigned long nr_scanned;
1088 bool isolated_pageblocks;
1090 spin_lock(&zone->lock);
1091 isolated_pageblocks = has_isolate_pageblock(zone);
1092 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1094 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1098 struct list_head *list;
1101 * Remove pages from lists in a round-robin fashion. A
1102 * batch_free count is maintained that is incremented when an
1103 * empty list is encountered. This is so more pages are freed
1104 * off fuller lists instead of spinning excessively around empty
1109 if (++migratetype == MIGRATE_PCPTYPES)
1111 list = &pcp->lists[migratetype];
1112 } while (list_empty(list));
1114 /* This is the only non-empty list. Free them all. */
1115 if (batch_free == MIGRATE_PCPTYPES)
1119 int mt; /* migratetype of the to-be-freed page */
1121 page = list_last_entry(list, struct page, lru);
1122 /* must delete as __free_one_page list manipulates */
1123 list_del(&page->lru);
1125 mt = get_pcppage_migratetype(page);
1126 /* MIGRATE_ISOLATE page should not go to pcplists */
1127 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1128 /* Pageblock could have been isolated meanwhile */
1129 if (unlikely(isolated_pageblocks))
1130 mt = get_pageblock_migratetype(page);
1132 if (bulkfree_pcp_prepare(page))
1135 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1136 trace_mm_page_pcpu_drain(page, 0, mt);
1137 } while (--count && --batch_free && !list_empty(list));
1139 spin_unlock(&zone->lock);
1142 static void free_one_page(struct zone *zone,
1143 struct page *page, unsigned long pfn,
1147 unsigned long nr_scanned;
1148 spin_lock(&zone->lock);
1149 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1151 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1153 if (unlikely(has_isolate_pageblock(zone) ||
1154 is_migrate_isolate(migratetype))) {
1155 migratetype = get_pfnblock_migratetype(page, pfn);
1157 __free_one_page(page, pfn, zone, order, migratetype);
1158 spin_unlock(&zone->lock);
1161 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1162 unsigned long zone, int nid)
1164 set_page_links(page, zone, nid, pfn);
1165 init_page_count(page);
1166 page_mapcount_reset(page);
1167 page_cpupid_reset_last(page);
1169 INIT_LIST_HEAD(&page->lru);
1170 #ifdef WANT_PAGE_VIRTUAL
1171 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1172 if (!is_highmem_idx(zone))
1173 set_page_address(page, __va(pfn << PAGE_SHIFT));
1177 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1180 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1183 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1184 static void init_reserved_page(unsigned long pfn)
1189 if (!early_page_uninitialised(pfn))
1192 nid = early_pfn_to_nid(pfn);
1193 pgdat = NODE_DATA(nid);
1195 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1196 struct zone *zone = &pgdat->node_zones[zid];
1198 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1201 __init_single_pfn(pfn, zid, nid);
1204 static inline void init_reserved_page(unsigned long pfn)
1207 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1210 * Initialised pages do not have PageReserved set. This function is
1211 * called for each range allocated by the bootmem allocator and
1212 * marks the pages PageReserved. The remaining valid pages are later
1213 * sent to the buddy page allocator.
1215 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
1217 unsigned long start_pfn = PFN_DOWN(start);
1218 unsigned long end_pfn = PFN_UP(end);
1220 for (; start_pfn < end_pfn; start_pfn++) {
1221 if (pfn_valid(start_pfn)) {
1222 struct page *page = pfn_to_page(start_pfn);
1224 init_reserved_page(start_pfn);
1226 /* Avoid false-positive PageTail() */
1227 INIT_LIST_HEAD(&page->lru);
1229 SetPageReserved(page);
1234 static void __free_pages_ok(struct page *page, unsigned int order)
1236 unsigned long flags;
1238 unsigned long pfn = page_to_pfn(page);
1240 if (!free_pages_prepare(page, order, true))
1243 migratetype = get_pfnblock_migratetype(page, pfn);
1244 local_irq_save(flags);
1245 __count_vm_events(PGFREE, 1 << order);
1246 free_one_page(page_zone(page), page, pfn, order, migratetype);
1247 local_irq_restore(flags);
1250 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1252 unsigned int nr_pages = 1 << order;
1253 struct page *p = page;
1257 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1259 __ClearPageReserved(p);
1260 set_page_count(p, 0);
1262 __ClearPageReserved(p);
1263 set_page_count(p, 0);
1265 page_zone(page)->managed_pages += nr_pages;
1266 set_page_refcounted(page);
1267 __free_pages(page, order);
1270 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1271 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1273 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1275 int __meminit early_pfn_to_nid(unsigned long pfn)
1277 static DEFINE_SPINLOCK(early_pfn_lock);
1280 spin_lock(&early_pfn_lock);
1281 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1284 spin_unlock(&early_pfn_lock);
1290 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1291 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1292 struct mminit_pfnnid_cache *state)
1296 nid = __early_pfn_to_nid(pfn, state);
1297 if (nid >= 0 && nid != node)
1302 /* Only safe to use early in boot when initialisation is single-threaded */
1303 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1305 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1310 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1315 struct mminit_pfnnid_cache *state)
1322 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1325 if (early_page_uninitialised(pfn))
1327 return __free_pages_boot_core(page, order);
1331 * Check that the whole (or subset of) a pageblock given by the interval of
1332 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1333 * with the migration of free compaction scanner. The scanners then need to
1334 * use only pfn_valid_within() check for arches that allow holes within
1337 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1339 * It's possible on some configurations to have a setup like node0 node1 node0
1340 * i.e. it's possible that all pages within a zones range of pages do not
1341 * belong to a single zone. We assume that a border between node0 and node1
1342 * can occur within a single pageblock, but not a node0 node1 node0
1343 * interleaving within a single pageblock. It is therefore sufficient to check
1344 * the first and last page of a pageblock and avoid checking each individual
1345 * page in a pageblock.
1347 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1348 unsigned long end_pfn, struct zone *zone)
1350 struct page *start_page;
1351 struct page *end_page;
1353 /* end_pfn is one past the range we are checking */
1356 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1359 start_page = pfn_to_page(start_pfn);
1361 if (page_zone(start_page) != zone)
1364 end_page = pfn_to_page(end_pfn);
1366 /* This gives a shorter code than deriving page_zone(end_page) */
1367 if (page_zone_id(start_page) != page_zone_id(end_page))
1373 void set_zone_contiguous(struct zone *zone)
1375 unsigned long block_start_pfn = zone->zone_start_pfn;
1376 unsigned long block_end_pfn;
1378 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1379 for (; block_start_pfn < zone_end_pfn(zone);
1380 block_start_pfn = block_end_pfn,
1381 block_end_pfn += pageblock_nr_pages) {
1383 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1385 if (!__pageblock_pfn_to_page(block_start_pfn,
1386 block_end_pfn, zone))
1390 /* We confirm that there is no hole */
1391 zone->contiguous = true;
1394 void clear_zone_contiguous(struct zone *zone)
1396 zone->contiguous = false;
1399 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1400 static void __init deferred_free_range(struct page *page,
1401 unsigned long pfn, int nr_pages)
1408 /* Free a large naturally-aligned chunk if possible */
1409 if (nr_pages == MAX_ORDER_NR_PAGES &&
1410 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1411 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1412 __free_pages_boot_core(page, MAX_ORDER-1);
1416 for (i = 0; i < nr_pages; i++, page++)
1417 __free_pages_boot_core(page, 0);
1420 /* Completion tracking for deferred_init_memmap() threads */
1421 static atomic_t pgdat_init_n_undone __initdata;
1422 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1424 static inline void __init pgdat_init_report_one_done(void)
1426 if (atomic_dec_and_test(&pgdat_init_n_undone))
1427 complete(&pgdat_init_all_done_comp);
1430 /* Initialise remaining memory on a node */
1431 static int __init deferred_init_memmap(void *data)
1433 pg_data_t *pgdat = data;
1434 int nid = pgdat->node_id;
1435 struct mminit_pfnnid_cache nid_init_state = { };
1436 unsigned long start = jiffies;
1437 unsigned long nr_pages = 0;
1438 unsigned long walk_start, walk_end;
1441 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1442 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1444 if (first_init_pfn == ULONG_MAX) {
1445 pgdat_init_report_one_done();
1449 /* Bind memory initialisation thread to a local node if possible */
1450 if (!cpumask_empty(cpumask))
1451 set_cpus_allowed_ptr(current, cpumask);
1453 /* Sanity check boundaries */
1454 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1455 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1456 pgdat->first_deferred_pfn = ULONG_MAX;
1458 /* Only the highest zone is deferred so find it */
1459 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1460 zone = pgdat->node_zones + zid;
1461 if (first_init_pfn < zone_end_pfn(zone))
1465 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1466 unsigned long pfn, end_pfn;
1467 struct page *page = NULL;
1468 struct page *free_base_page = NULL;
1469 unsigned long free_base_pfn = 0;
1472 end_pfn = min(walk_end, zone_end_pfn(zone));
1473 pfn = first_init_pfn;
1474 if (pfn < walk_start)
1476 if (pfn < zone->zone_start_pfn)
1477 pfn = zone->zone_start_pfn;
1479 for (; pfn < end_pfn; pfn++) {
1480 if (!pfn_valid_within(pfn))
1484 * Ensure pfn_valid is checked every
1485 * MAX_ORDER_NR_PAGES for memory holes
1487 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1488 if (!pfn_valid(pfn)) {
1494 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1499 /* Minimise pfn page lookups and scheduler checks */
1500 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1503 nr_pages += nr_to_free;
1504 deferred_free_range(free_base_page,
1505 free_base_pfn, nr_to_free);
1506 free_base_page = NULL;
1507 free_base_pfn = nr_to_free = 0;
1509 page = pfn_to_page(pfn);
1514 VM_BUG_ON(page_zone(page) != zone);
1518 __init_single_page(page, pfn, zid, nid);
1519 if (!free_base_page) {
1520 free_base_page = page;
1521 free_base_pfn = pfn;
1526 /* Where possible, batch up pages for a single free */
1529 /* Free the current block of pages to allocator */
1530 nr_pages += nr_to_free;
1531 deferred_free_range(free_base_page, free_base_pfn,
1533 free_base_page = NULL;
1534 free_base_pfn = nr_to_free = 0;
1537 first_init_pfn = max(end_pfn, first_init_pfn);
1540 /* Sanity check that the next zone really is unpopulated */
1541 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1543 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1544 jiffies_to_msecs(jiffies - start));
1546 pgdat_init_report_one_done();
1549 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1551 void __init page_alloc_init_late(void)
1555 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1558 /* There will be num_node_state(N_MEMORY) threads */
1559 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1560 for_each_node_state(nid, N_MEMORY) {
1561 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1564 /* Block until all are initialised */
1565 wait_for_completion(&pgdat_init_all_done_comp);
1567 /* Reinit limits that are based on free pages after the kernel is up */
1568 files_maxfiles_init();
1571 for_each_populated_zone(zone)
1572 set_zone_contiguous(zone);
1576 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1577 void __init init_cma_reserved_pageblock(struct page *page)
1579 unsigned i = pageblock_nr_pages;
1580 struct page *p = page;
1583 __ClearPageReserved(p);
1584 set_page_count(p, 0);
1587 set_pageblock_migratetype(page, MIGRATE_CMA);
1589 if (pageblock_order >= MAX_ORDER) {
1590 i = pageblock_nr_pages;
1593 set_page_refcounted(p);
1594 __free_pages(p, MAX_ORDER - 1);
1595 p += MAX_ORDER_NR_PAGES;
1596 } while (i -= MAX_ORDER_NR_PAGES);
1598 set_page_refcounted(page);
1599 __free_pages(page, pageblock_order);
1602 adjust_managed_page_count(page, pageblock_nr_pages);
1607 * The order of subdivision here is critical for the IO subsystem.
1608 * Please do not alter this order without good reasons and regression
1609 * testing. Specifically, as large blocks of memory are subdivided,
1610 * the order in which smaller blocks are delivered depends on the order
1611 * they're subdivided in this function. This is the primary factor
1612 * influencing the order in which pages are delivered to the IO
1613 * subsystem according to empirical testing, and this is also justified
1614 * by considering the behavior of a buddy system containing a single
1615 * large block of memory acted on by a series of small allocations.
1616 * This behavior is a critical factor in sglist merging's success.
1620 static inline void expand(struct zone *zone, struct page *page,
1621 int low, int high, struct free_area *area,
1624 unsigned long size = 1 << high;
1626 while (high > low) {
1630 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1632 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1633 debug_guardpage_enabled() &&
1634 high < debug_guardpage_minorder()) {
1636 * Mark as guard pages (or page), that will allow to
1637 * merge back to allocator when buddy will be freed.
1638 * Corresponding page table entries will not be touched,
1639 * pages will stay not present in virtual address space
1641 set_page_guard(zone, &page[size], high, migratetype);
1644 list_add(&page[size].lru, &area->free_list[migratetype]);
1646 set_page_order(&page[size], high);
1650 static void check_new_page_bad(struct page *page)
1652 const char *bad_reason = NULL;
1653 unsigned long bad_flags = 0;
1655 if (unlikely(atomic_read(&page->_mapcount) != -1))
1656 bad_reason = "nonzero mapcount";
1657 if (unlikely(page->mapping != NULL))
1658 bad_reason = "non-NULL mapping";
1659 if (unlikely(page_ref_count(page) != 0))
1660 bad_reason = "nonzero _count";
1661 if (unlikely(page->flags & __PG_HWPOISON)) {
1662 bad_reason = "HWPoisoned (hardware-corrupted)";
1663 bad_flags = __PG_HWPOISON;
1665 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1666 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1667 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1670 if (unlikely(page->mem_cgroup))
1671 bad_reason = "page still charged to cgroup";
1673 bad_page(page, bad_reason, bad_flags);
1677 * This page is about to be returned from the page allocator
1679 static inline int check_new_page(struct page *page)
1681 if (likely(page_expected_state(page,
1682 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1685 check_new_page_bad(page);
1689 static inline bool free_pages_prezeroed(bool poisoned)
1691 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1692 page_poisoning_enabled() && poisoned;
1695 #ifdef CONFIG_DEBUG_VM
1696 static bool check_pcp_refill(struct page *page)
1701 static bool check_new_pcp(struct page *page)
1703 return check_new_page(page);
1706 static bool check_pcp_refill(struct page *page)
1708 return check_new_page(page);
1710 static bool check_new_pcp(struct page *page)
1714 #endif /* CONFIG_DEBUG_VM */
1716 static bool check_new_pages(struct page *page, unsigned int order)
1719 for (i = 0; i < (1 << order); i++) {
1720 struct page *p = page + i;
1722 if (unlikely(check_new_page(p)))
1729 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1730 unsigned int alloc_flags)
1733 bool poisoned = true;
1735 for (i = 0; i < (1 << order); i++) {
1736 struct page *p = page + i;
1738 poisoned &= page_is_poisoned(p);
1741 set_page_private(page, 0);
1742 set_page_refcounted(page);
1744 arch_alloc_page(page, order);
1745 kernel_map_pages(page, 1 << order, 1);
1746 kernel_poison_pages(page, 1 << order, 1);
1747 kasan_alloc_pages(page, order);
1749 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1750 for (i = 0; i < (1 << order); i++)
1751 clear_highpage(page + i);
1753 if (order && (gfp_flags & __GFP_COMP))
1754 prep_compound_page(page, order);
1756 set_page_owner(page, order, gfp_flags);
1759 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1760 * allocate the page. The expectation is that the caller is taking
1761 * steps that will free more memory. The caller should avoid the page
1762 * being used for !PFMEMALLOC purposes.
1764 if (alloc_flags & ALLOC_NO_WATERMARKS)
1765 set_page_pfmemalloc(page);
1767 clear_page_pfmemalloc(page);
1771 * Go through the free lists for the given migratetype and remove
1772 * the smallest available page from the freelists
1775 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1778 unsigned int current_order;
1779 struct free_area *area;
1782 /* Find a page of the appropriate size in the preferred list */
1783 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1784 area = &(zone->free_area[current_order]);
1785 page = list_first_entry_or_null(&area->free_list[migratetype],
1789 list_del(&page->lru);
1790 rmv_page_order(page);
1792 expand(zone, page, order, current_order, area, migratetype);
1793 set_pcppage_migratetype(page, migratetype);
1802 * This array describes the order lists are fallen back to when
1803 * the free lists for the desirable migrate type are depleted
1805 static int fallbacks[MIGRATE_TYPES][4] = {
1806 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1807 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1808 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1810 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1812 #ifdef CONFIG_MEMORY_ISOLATION
1813 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1818 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1821 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1824 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1825 unsigned int order) { return NULL; }
1829 * Move the free pages in a range to the free lists of the requested type.
1830 * Note that start_page and end_pages are not aligned on a pageblock
1831 * boundary. If alignment is required, use move_freepages_block()
1833 int move_freepages(struct zone *zone,
1834 struct page *start_page, struct page *end_page,
1839 int pages_moved = 0;
1841 #ifndef CONFIG_HOLES_IN_ZONE
1843 * page_zone is not safe to call in this context when
1844 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1845 * anyway as we check zone boundaries in move_freepages_block().
1846 * Remove at a later date when no bug reports exist related to
1847 * grouping pages by mobility
1849 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1852 for (page = start_page; page <= end_page;) {
1853 /* Make sure we are not inadvertently changing nodes */
1854 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1856 if (!pfn_valid_within(page_to_pfn(page))) {
1861 if (!PageBuddy(page)) {
1866 order = page_order(page);
1867 list_move(&page->lru,
1868 &zone->free_area[order].free_list[migratetype]);
1870 pages_moved += 1 << order;
1876 int move_freepages_block(struct zone *zone, struct page *page,
1879 unsigned long start_pfn, end_pfn;
1880 struct page *start_page, *end_page;
1882 start_pfn = page_to_pfn(page);
1883 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1884 start_page = pfn_to_page(start_pfn);
1885 end_page = start_page + pageblock_nr_pages - 1;
1886 end_pfn = start_pfn + pageblock_nr_pages - 1;
1888 /* Do not cross zone boundaries */
1889 if (!zone_spans_pfn(zone, start_pfn))
1891 if (!zone_spans_pfn(zone, end_pfn))
1894 return move_freepages(zone, start_page, end_page, migratetype);
1897 static void change_pageblock_range(struct page *pageblock_page,
1898 int start_order, int migratetype)
1900 int nr_pageblocks = 1 << (start_order - pageblock_order);
1902 while (nr_pageblocks--) {
1903 set_pageblock_migratetype(pageblock_page, migratetype);
1904 pageblock_page += pageblock_nr_pages;
1909 * When we are falling back to another migratetype during allocation, try to
1910 * steal extra free pages from the same pageblocks to satisfy further
1911 * allocations, instead of polluting multiple pageblocks.
1913 * If we are stealing a relatively large buddy page, it is likely there will
1914 * be more free pages in the pageblock, so try to steal them all. For
1915 * reclaimable and unmovable allocations, we steal regardless of page size,
1916 * as fragmentation caused by those allocations polluting movable pageblocks
1917 * is worse than movable allocations stealing from unmovable and reclaimable
1920 static bool can_steal_fallback(unsigned int order, int start_mt)
1923 * Leaving this order check is intended, although there is
1924 * relaxed order check in next check. The reason is that
1925 * we can actually steal whole pageblock if this condition met,
1926 * but, below check doesn't guarantee it and that is just heuristic
1927 * so could be changed anytime.
1929 if (order >= pageblock_order)
1932 if (order >= pageblock_order / 2 ||
1933 start_mt == MIGRATE_RECLAIMABLE ||
1934 start_mt == MIGRATE_UNMOVABLE ||
1935 page_group_by_mobility_disabled)
1942 * This function implements actual steal behaviour. If order is large enough,
1943 * we can steal whole pageblock. If not, we first move freepages in this
1944 * pageblock and check whether half of pages are moved or not. If half of
1945 * pages are moved, we can change migratetype of pageblock and permanently
1946 * use it's pages as requested migratetype in the future.
1948 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1951 unsigned int current_order = page_order(page);
1954 /* Take ownership for orders >= pageblock_order */
1955 if (current_order >= pageblock_order) {
1956 change_pageblock_range(page, current_order, start_type);
1960 pages = move_freepages_block(zone, page, start_type);
1962 /* Claim the whole block if over half of it is free */
1963 if (pages >= (1 << (pageblock_order-1)) ||
1964 page_group_by_mobility_disabled)
1965 set_pageblock_migratetype(page, start_type);
1969 * Check whether there is a suitable fallback freepage with requested order.
1970 * If only_stealable is true, this function returns fallback_mt only if
1971 * we can steal other freepages all together. This would help to reduce
1972 * fragmentation due to mixed migratetype pages in one pageblock.
1974 int find_suitable_fallback(struct free_area *area, unsigned int order,
1975 int migratetype, bool only_stealable, bool *can_steal)
1980 if (area->nr_free == 0)
1985 fallback_mt = fallbacks[migratetype][i];
1986 if (fallback_mt == MIGRATE_TYPES)
1989 if (list_empty(&area->free_list[fallback_mt]))
1992 if (can_steal_fallback(order, migratetype))
1995 if (!only_stealable)
2006 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2007 * there are no empty page blocks that contain a page with a suitable order
2009 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2010 unsigned int alloc_order)
2013 unsigned long max_managed, flags;
2016 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2017 * Check is race-prone but harmless.
2019 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2020 if (zone->nr_reserved_highatomic >= max_managed)
2023 spin_lock_irqsave(&zone->lock, flags);
2025 /* Recheck the nr_reserved_highatomic limit under the lock */
2026 if (zone->nr_reserved_highatomic >= max_managed)
2030 mt = get_pageblock_migratetype(page);
2031 if (mt != MIGRATE_HIGHATOMIC &&
2032 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2033 zone->nr_reserved_highatomic += pageblock_nr_pages;
2034 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2035 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2039 spin_unlock_irqrestore(&zone->lock, flags);
2043 * Used when an allocation is about to fail under memory pressure. This
2044 * potentially hurts the reliability of high-order allocations when under
2045 * intense memory pressure but failed atomic allocations should be easier
2046 * to recover from than an OOM.
2048 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2050 struct zonelist *zonelist = ac->zonelist;
2051 unsigned long flags;
2057 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2059 /* Preserve at least one pageblock */
2060 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2063 spin_lock_irqsave(&zone->lock, flags);
2064 for (order = 0; order < MAX_ORDER; order++) {
2065 struct free_area *area = &(zone->free_area[order]);
2067 page = list_first_entry_or_null(
2068 &area->free_list[MIGRATE_HIGHATOMIC],
2074 * It should never happen but changes to locking could
2075 * inadvertently allow a per-cpu drain to add pages
2076 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2077 * and watch for underflows.
2079 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2080 zone->nr_reserved_highatomic);
2083 * Convert to ac->migratetype and avoid the normal
2084 * pageblock stealing heuristics. Minimally, the caller
2085 * is doing the work and needs the pages. More
2086 * importantly, if the block was always converted to
2087 * MIGRATE_UNMOVABLE or another type then the number
2088 * of pageblocks that cannot be completely freed
2091 set_pageblock_migratetype(page, ac->migratetype);
2092 move_freepages_block(zone, page, ac->migratetype);
2093 spin_unlock_irqrestore(&zone->lock, flags);
2096 spin_unlock_irqrestore(&zone->lock, flags);
2100 /* Remove an element from the buddy allocator from the fallback list */
2101 static inline struct page *
2102 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2104 struct free_area *area;
2105 unsigned int current_order;
2110 /* Find the largest possible block of pages in the other list */
2111 for (current_order = MAX_ORDER-1;
2112 current_order >= order && current_order <= MAX_ORDER-1;
2114 area = &(zone->free_area[current_order]);
2115 fallback_mt = find_suitable_fallback(area, current_order,
2116 start_migratetype, false, &can_steal);
2117 if (fallback_mt == -1)
2120 page = list_first_entry(&area->free_list[fallback_mt],
2123 steal_suitable_fallback(zone, page, start_migratetype);
2125 /* Remove the page from the freelists */
2127 list_del(&page->lru);
2128 rmv_page_order(page);
2130 expand(zone, page, order, current_order, area,
2133 * The pcppage_migratetype may differ from pageblock's
2134 * migratetype depending on the decisions in
2135 * find_suitable_fallback(). This is OK as long as it does not
2136 * differ for MIGRATE_CMA pageblocks. Those can be used as
2137 * fallback only via special __rmqueue_cma_fallback() function
2139 set_pcppage_migratetype(page, start_migratetype);
2141 trace_mm_page_alloc_extfrag(page, order, current_order,
2142 start_migratetype, fallback_mt);
2151 * Do the hard work of removing an element from the buddy allocator.
2152 * Call me with the zone->lock already held.
2154 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2159 page = __rmqueue_smallest(zone, order, migratetype);
2160 if (unlikely(!page)) {
2161 if (migratetype == MIGRATE_MOVABLE)
2162 page = __rmqueue_cma_fallback(zone, order);
2165 page = __rmqueue_fallback(zone, order, migratetype);
2168 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2173 * Obtain a specified number of elements from the buddy allocator, all under
2174 * a single hold of the lock, for efficiency. Add them to the supplied list.
2175 * Returns the number of new pages which were placed at *list.
2177 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2178 unsigned long count, struct list_head *list,
2179 int migratetype, bool cold)
2183 spin_lock(&zone->lock);
2184 for (i = 0; i < count; ++i) {
2185 struct page *page = __rmqueue(zone, order, migratetype);
2186 if (unlikely(page == NULL))
2189 if (unlikely(check_pcp_refill(page)))
2193 * Split buddy pages returned by expand() are received here
2194 * in physical page order. The page is added to the callers and
2195 * list and the list head then moves forward. From the callers
2196 * perspective, the linked list is ordered by page number in
2197 * some conditions. This is useful for IO devices that can
2198 * merge IO requests if the physical pages are ordered
2202 list_add(&page->lru, list);
2204 list_add_tail(&page->lru, list);
2206 if (is_migrate_cma(get_pcppage_migratetype(page)))
2207 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2210 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2211 spin_unlock(&zone->lock);
2217 * Called from the vmstat counter updater to drain pagesets of this
2218 * currently executing processor on remote nodes after they have
2221 * Note that this function must be called with the thread pinned to
2222 * a single processor.
2224 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2226 unsigned long flags;
2227 int to_drain, batch;
2229 local_irq_save(flags);
2230 batch = READ_ONCE(pcp->batch);
2231 to_drain = min(pcp->count, batch);
2233 free_pcppages_bulk(zone, to_drain, pcp);
2234 pcp->count -= to_drain;
2236 local_irq_restore(flags);
2241 * Drain pcplists of the indicated processor and zone.
2243 * The processor must either be the current processor and the
2244 * thread pinned to the current processor or a processor that
2247 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2249 unsigned long flags;
2250 struct per_cpu_pageset *pset;
2251 struct per_cpu_pages *pcp;
2253 local_irq_save(flags);
2254 pset = per_cpu_ptr(zone->pageset, cpu);
2258 free_pcppages_bulk(zone, pcp->count, pcp);
2261 local_irq_restore(flags);
2265 * Drain pcplists of all zones on the indicated processor.
2267 * The processor must either be the current processor and the
2268 * thread pinned to the current processor or a processor that
2271 static void drain_pages(unsigned int cpu)
2275 for_each_populated_zone(zone) {
2276 drain_pages_zone(cpu, zone);
2281 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2283 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2284 * the single zone's pages.
2286 void drain_local_pages(struct zone *zone)
2288 int cpu = smp_processor_id();
2291 drain_pages_zone(cpu, zone);
2297 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2299 * When zone parameter is non-NULL, spill just the single zone's pages.
2301 * Note that this code is protected against sending an IPI to an offline
2302 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2303 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2304 * nothing keeps CPUs from showing up after we populated the cpumask and
2305 * before the call to on_each_cpu_mask().
2307 void drain_all_pages(struct zone *zone)
2312 * Allocate in the BSS so we wont require allocation in
2313 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2315 static cpumask_t cpus_with_pcps;
2318 * We don't care about racing with CPU hotplug event
2319 * as offline notification will cause the notified
2320 * cpu to drain that CPU pcps and on_each_cpu_mask
2321 * disables preemption as part of its processing
2323 for_each_online_cpu(cpu) {
2324 struct per_cpu_pageset *pcp;
2326 bool has_pcps = false;
2329 pcp = per_cpu_ptr(zone->pageset, cpu);
2333 for_each_populated_zone(z) {
2334 pcp = per_cpu_ptr(z->pageset, cpu);
2335 if (pcp->pcp.count) {
2343 cpumask_set_cpu(cpu, &cpus_with_pcps);
2345 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2347 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2351 #ifdef CONFIG_HIBERNATION
2353 void mark_free_pages(struct zone *zone)
2355 unsigned long pfn, max_zone_pfn;
2356 unsigned long flags;
2357 unsigned int order, t;
2360 if (zone_is_empty(zone))
2363 spin_lock_irqsave(&zone->lock, flags);
2365 max_zone_pfn = zone_end_pfn(zone);
2366 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2367 if (pfn_valid(pfn)) {
2368 page = pfn_to_page(pfn);
2370 if (page_zone(page) != zone)
2373 if (!swsusp_page_is_forbidden(page))
2374 swsusp_unset_page_free(page);
2377 for_each_migratetype_order(order, t) {
2378 list_for_each_entry(page,
2379 &zone->free_area[order].free_list[t], lru) {
2382 pfn = page_to_pfn(page);
2383 for (i = 0; i < (1UL << order); i++)
2384 swsusp_set_page_free(pfn_to_page(pfn + i));
2387 spin_unlock_irqrestore(&zone->lock, flags);
2389 #endif /* CONFIG_PM */
2392 * Free a 0-order page
2393 * cold == true ? free a cold page : free a hot page
2395 void free_hot_cold_page(struct page *page, bool cold)
2397 struct zone *zone = page_zone(page);
2398 struct per_cpu_pages *pcp;
2399 unsigned long flags;
2400 unsigned long pfn = page_to_pfn(page);
2403 if (!free_pcp_prepare(page))
2406 migratetype = get_pfnblock_migratetype(page, pfn);
2407 set_pcppage_migratetype(page, migratetype);
2408 local_irq_save(flags);
2409 __count_vm_event(PGFREE);
2412 * We only track unmovable, reclaimable and movable on pcp lists.
2413 * Free ISOLATE pages back to the allocator because they are being
2414 * offlined but treat RESERVE as movable pages so we can get those
2415 * areas back if necessary. Otherwise, we may have to free
2416 * excessively into the page allocator
2418 if (migratetype >= MIGRATE_PCPTYPES) {
2419 if (unlikely(is_migrate_isolate(migratetype))) {
2420 free_one_page(zone, page, pfn, 0, migratetype);
2423 migratetype = MIGRATE_MOVABLE;
2426 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2428 list_add(&page->lru, &pcp->lists[migratetype]);
2430 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2432 if (pcp->count >= pcp->high) {
2433 unsigned long batch = READ_ONCE(pcp->batch);
2434 free_pcppages_bulk(zone, batch, pcp);
2435 pcp->count -= batch;
2439 local_irq_restore(flags);
2443 * Free a list of 0-order pages
2445 void free_hot_cold_page_list(struct list_head *list, bool cold)
2447 struct page *page, *next;
2449 list_for_each_entry_safe(page, next, list, lru) {
2450 trace_mm_page_free_batched(page, cold);
2451 free_hot_cold_page(page, cold);
2456 * split_page takes a non-compound higher-order page, and splits it into
2457 * n (1<<order) sub-pages: page[0..n]
2458 * Each sub-page must be freed individually.
2460 * Note: this is probably too low level an operation for use in drivers.
2461 * Please consult with lkml before using this in your driver.
2463 void split_page(struct page *page, unsigned int order)
2468 VM_BUG_ON_PAGE(PageCompound(page), page);
2469 VM_BUG_ON_PAGE(!page_count(page), page);
2471 #ifdef CONFIG_KMEMCHECK
2473 * Split shadow pages too, because free(page[0]) would
2474 * otherwise free the whole shadow.
2476 if (kmemcheck_page_is_tracked(page))
2477 split_page(virt_to_page(page[0].shadow), order);
2480 gfp_mask = get_page_owner_gfp(page);
2481 set_page_owner(page, 0, gfp_mask);
2482 for (i = 1; i < (1 << order); i++) {
2483 set_page_refcounted(page + i);
2484 set_page_owner(page + i, 0, gfp_mask);
2487 EXPORT_SYMBOL_GPL(split_page);
2489 int __isolate_free_page(struct page *page, unsigned int order)
2491 unsigned long watermark;
2495 BUG_ON(!PageBuddy(page));
2497 zone = page_zone(page);
2498 mt = get_pageblock_migratetype(page);
2500 if (!is_migrate_isolate(mt)) {
2501 /* Obey watermarks as if the page was being allocated */
2502 watermark = low_wmark_pages(zone) + (1 << order);
2503 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2506 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2509 /* Remove page from free list */
2510 list_del(&page->lru);
2511 zone->free_area[order].nr_free--;
2512 rmv_page_order(page);
2514 set_page_owner(page, order, __GFP_MOVABLE);
2516 /* Set the pageblock if the isolated page is at least a pageblock */
2517 if (order >= pageblock_order - 1) {
2518 struct page *endpage = page + (1 << order) - 1;
2519 for (; page < endpage; page += pageblock_nr_pages) {
2520 int mt = get_pageblock_migratetype(page);
2521 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2522 set_pageblock_migratetype(page,
2528 return 1UL << order;
2532 * Similar to split_page except the page is already free. As this is only
2533 * being used for migration, the migratetype of the block also changes.
2534 * As this is called with interrupts disabled, the caller is responsible
2535 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2538 * Note: this is probably too low level an operation for use in drivers.
2539 * Please consult with lkml before using this in your driver.
2541 int split_free_page(struct page *page)
2546 order = page_order(page);
2548 nr_pages = __isolate_free_page(page, order);
2552 /* Split into individual pages */
2553 set_page_refcounted(page);
2554 split_page(page, order);
2559 * Update NUMA hit/miss statistics
2561 * Must be called with interrupts disabled.
2563 * When __GFP_OTHER_NODE is set assume the node of the preferred
2564 * zone is the local node. This is useful for daemons who allocate
2565 * memory on behalf of other processes.
2567 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2571 int local_nid = numa_node_id();
2572 enum zone_stat_item local_stat = NUMA_LOCAL;
2574 if (unlikely(flags & __GFP_OTHER_NODE)) {
2575 local_stat = NUMA_OTHER;
2576 local_nid = preferred_zone->node;
2579 if (z->node == local_nid) {
2580 __inc_zone_state(z, NUMA_HIT);
2581 __inc_zone_state(z, local_stat);
2583 __inc_zone_state(z, NUMA_MISS);
2584 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2590 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2593 struct page *buffered_rmqueue(struct zone *preferred_zone,
2594 struct zone *zone, unsigned int order,
2595 gfp_t gfp_flags, unsigned int alloc_flags,
2598 unsigned long flags;
2600 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2602 if (likely(order == 0)) {
2603 struct per_cpu_pages *pcp;
2604 struct list_head *list;
2606 local_irq_save(flags);
2608 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2609 list = &pcp->lists[migratetype];
2610 if (list_empty(list)) {
2611 pcp->count += rmqueue_bulk(zone, 0,
2614 if (unlikely(list_empty(list)))
2619 page = list_last_entry(list, struct page, lru);
2621 page = list_first_entry(list, struct page, lru);
2622 } while (page && check_new_pcp(page));
2624 __dec_zone_state(zone, NR_ALLOC_BATCH);
2625 list_del(&page->lru);
2629 * We most definitely don't want callers attempting to
2630 * allocate greater than order-1 page units with __GFP_NOFAIL.
2632 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2633 spin_lock_irqsave(&zone->lock, flags);
2637 if (alloc_flags & ALLOC_HARDER) {
2638 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2640 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2643 page = __rmqueue(zone, order, migratetype);
2644 } while (page && check_new_pages(page, order));
2645 spin_unlock(&zone->lock);
2648 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2649 __mod_zone_freepage_state(zone, -(1 << order),
2650 get_pcppage_migratetype(page));
2653 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2654 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2655 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2657 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2658 zone_statistics(preferred_zone, zone, gfp_flags);
2659 local_irq_restore(flags);
2661 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2665 local_irq_restore(flags);
2669 #ifdef CONFIG_FAIL_PAGE_ALLOC
2672 struct fault_attr attr;
2674 bool ignore_gfp_highmem;
2675 bool ignore_gfp_reclaim;
2677 } fail_page_alloc = {
2678 .attr = FAULT_ATTR_INITIALIZER,
2679 .ignore_gfp_reclaim = true,
2680 .ignore_gfp_highmem = true,
2684 static int __init setup_fail_page_alloc(char *str)
2686 return setup_fault_attr(&fail_page_alloc.attr, str);
2688 __setup("fail_page_alloc=", setup_fail_page_alloc);
2690 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2692 if (order < fail_page_alloc.min_order)
2694 if (gfp_mask & __GFP_NOFAIL)
2696 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2698 if (fail_page_alloc.ignore_gfp_reclaim &&
2699 (gfp_mask & __GFP_DIRECT_RECLAIM))
2702 return should_fail(&fail_page_alloc.attr, 1 << order);
2705 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2707 static int __init fail_page_alloc_debugfs(void)
2709 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2712 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2713 &fail_page_alloc.attr);
2715 return PTR_ERR(dir);
2717 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2718 &fail_page_alloc.ignore_gfp_reclaim))
2720 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2721 &fail_page_alloc.ignore_gfp_highmem))
2723 if (!debugfs_create_u32("min-order", mode, dir,
2724 &fail_page_alloc.min_order))
2729 debugfs_remove_recursive(dir);
2734 late_initcall(fail_page_alloc_debugfs);
2736 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2738 #else /* CONFIG_FAIL_PAGE_ALLOC */
2740 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2745 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2748 * Return true if free base pages are above 'mark'. For high-order checks it
2749 * will return true of the order-0 watermark is reached and there is at least
2750 * one free page of a suitable size. Checking now avoids taking the zone lock
2751 * to check in the allocation paths if no pages are free.
2753 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2754 int classzone_idx, unsigned int alloc_flags,
2759 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2761 /* free_pages may go negative - that's OK */
2762 free_pages -= (1 << order) - 1;
2764 if (alloc_flags & ALLOC_HIGH)
2768 * If the caller does not have rights to ALLOC_HARDER then subtract
2769 * the high-atomic reserves. This will over-estimate the size of the
2770 * atomic reserve but it avoids a search.
2772 if (likely(!alloc_harder))
2773 free_pages -= z->nr_reserved_highatomic;
2778 /* If allocation can't use CMA areas don't use free CMA pages */
2779 if (!(alloc_flags & ALLOC_CMA))
2780 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2784 * Check watermarks for an order-0 allocation request. If these
2785 * are not met, then a high-order request also cannot go ahead
2786 * even if a suitable page happened to be free.
2788 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2791 /* If this is an order-0 request then the watermark is fine */
2795 /* For a high-order request, check at least one suitable page is free */
2796 for (o = order; o < MAX_ORDER; o++) {
2797 struct free_area *area = &z->free_area[o];
2806 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2807 if (!list_empty(&area->free_list[mt]))
2812 if ((alloc_flags & ALLOC_CMA) &&
2813 !list_empty(&area->free_list[MIGRATE_CMA])) {
2821 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2822 int classzone_idx, unsigned int alloc_flags)
2824 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2825 zone_page_state(z, NR_FREE_PAGES));
2828 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2829 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2831 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2835 /* If allocation can't use CMA areas don't use free CMA pages */
2836 if (!(alloc_flags & ALLOC_CMA))
2837 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2841 * Fast check for order-0 only. If this fails then the reserves
2842 * need to be calculated. There is a corner case where the check
2843 * passes but only the high-order atomic reserve are free. If
2844 * the caller is !atomic then it'll uselessly search the free
2845 * list. That corner case is then slower but it is harmless.
2847 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2850 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2854 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2855 unsigned long mark, int classzone_idx)
2857 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2859 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2860 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2862 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2867 static bool zone_local(struct zone *local_zone, struct zone *zone)
2869 return local_zone->node == zone->node;
2872 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2874 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2877 #else /* CONFIG_NUMA */
2878 static bool zone_local(struct zone *local_zone, struct zone *zone)
2883 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2887 #endif /* CONFIG_NUMA */
2889 static void reset_alloc_batches(struct zone *preferred_zone)
2891 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2894 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2895 high_wmark_pages(zone) - low_wmark_pages(zone) -
2896 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2897 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2898 } while (zone++ != preferred_zone);
2902 * get_page_from_freelist goes through the zonelist trying to allocate
2905 static struct page *
2906 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2907 const struct alloc_context *ac)
2909 struct zoneref *z = ac->preferred_zoneref;
2911 bool fair_skipped = false;
2912 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2916 * Scan zonelist, looking for a zone with enough free.
2917 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2919 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2924 if (cpusets_enabled() &&
2925 (alloc_flags & ALLOC_CPUSET) &&
2926 !__cpuset_zone_allowed(zone, gfp_mask))
2929 * Distribute pages in proportion to the individual
2930 * zone size to ensure fair page aging. The zone a
2931 * page was allocated in should have no effect on the
2932 * time the page has in memory before being reclaimed.
2935 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2936 fair_skipped = true;
2939 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2946 * When allocating a page cache page for writing, we
2947 * want to get it from a zone that is within its dirty
2948 * limit, such that no single zone holds more than its
2949 * proportional share of globally allowed dirty pages.
2950 * The dirty limits take into account the zone's
2951 * lowmem reserves and high watermark so that kswapd
2952 * should be able to balance it without having to
2953 * write pages from its LRU list.
2955 * This may look like it could increase pressure on
2956 * lower zones by failing allocations in higher zones
2957 * before they are full. But the pages that do spill
2958 * over are limited as the lower zones are protected
2959 * by this very same mechanism. It should not become
2960 * a practical burden to them.
2962 * XXX: For now, allow allocations to potentially
2963 * exceed the per-zone dirty limit in the slowpath
2964 * (spread_dirty_pages unset) before going into reclaim,
2965 * which is important when on a NUMA setup the allowed
2966 * zones are together not big enough to reach the
2967 * global limit. The proper fix for these situations
2968 * will require awareness of zones in the
2969 * dirty-throttling and the flusher threads.
2971 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2974 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2975 if (!zone_watermark_fast(zone, order, mark,
2976 ac_classzone_idx(ac), alloc_flags)) {
2979 /* Checked here to keep the fast path fast */
2980 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2981 if (alloc_flags & ALLOC_NO_WATERMARKS)
2984 if (zone_reclaim_mode == 0 ||
2985 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2988 ret = zone_reclaim(zone, gfp_mask, order);
2990 case ZONE_RECLAIM_NOSCAN:
2993 case ZONE_RECLAIM_FULL:
2994 /* scanned but unreclaimable */
2997 /* did we reclaim enough */
2998 if (zone_watermark_ok(zone, order, mark,
2999 ac_classzone_idx(ac), alloc_flags))
3007 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
3008 gfp_mask, alloc_flags, ac->migratetype);
3010 prep_new_page(page, order, gfp_mask, alloc_flags);
3013 * If this is a high-order atomic allocation then check
3014 * if the pageblock should be reserved for the future
3016 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3017 reserve_highatomic_pageblock(page, zone, order);
3024 * The first pass makes sure allocations are spread fairly within the
3025 * local node. However, the local node might have free pages left
3026 * after the fairness batches are exhausted, and remote zones haven't
3027 * even been considered yet. Try once more without fairness, and
3028 * include remote zones now, before entering the slowpath and waking
3029 * kswapd: prefer spilling to a remote zone over swapping locally.
3034 fair_skipped = false;
3035 reset_alloc_batches(ac->preferred_zoneref->zone);
3043 * Large machines with many possible nodes should not always dump per-node
3044 * meminfo in irq context.
3046 static inline bool should_suppress_show_mem(void)
3051 ret = in_interrupt();
3056 static DEFINE_RATELIMIT_STATE(nopage_rs,
3057 DEFAULT_RATELIMIT_INTERVAL,
3058 DEFAULT_RATELIMIT_BURST);
3060 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3062 unsigned int filter = SHOW_MEM_FILTER_NODES;
3064 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3065 debug_guardpage_minorder() > 0)
3069 * This documents exceptions given to allocations in certain
3070 * contexts that are allowed to allocate outside current's set
3073 if (!(gfp_mask & __GFP_NOMEMALLOC))
3074 if (test_thread_flag(TIF_MEMDIE) ||
3075 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3076 filter &= ~SHOW_MEM_FILTER_NODES;
3077 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3078 filter &= ~SHOW_MEM_FILTER_NODES;
3081 struct va_format vaf;
3084 va_start(args, fmt);
3089 pr_warn("%pV", &vaf);
3094 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3095 current->comm, order, gfp_mask, &gfp_mask);
3097 if (!should_suppress_show_mem())
3101 static inline struct page *
3102 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3103 const struct alloc_context *ac, unsigned long *did_some_progress)
3105 struct oom_control oc = {
3106 .zonelist = ac->zonelist,
3107 .nodemask = ac->nodemask,
3108 .gfp_mask = gfp_mask,
3113 *did_some_progress = 0;
3116 * Acquire the oom lock. If that fails, somebody else is
3117 * making progress for us.
3119 if (!mutex_trylock(&oom_lock)) {
3120 *did_some_progress = 1;
3121 schedule_timeout_uninterruptible(1);
3126 * Go through the zonelist yet one more time, keep very high watermark
3127 * here, this is only to catch a parallel oom killing, we must fail if
3128 * we're still under heavy pressure.
3130 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3131 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3135 if (!(gfp_mask & __GFP_NOFAIL)) {
3136 /* Coredumps can quickly deplete all memory reserves */
3137 if (current->flags & PF_DUMPCORE)
3139 /* The OOM killer will not help higher order allocs */
3140 if (order > PAGE_ALLOC_COSTLY_ORDER)
3142 /* The OOM killer does not needlessly kill tasks for lowmem */
3143 if (ac->high_zoneidx < ZONE_NORMAL)
3145 if (pm_suspended_storage())
3148 * XXX: GFP_NOFS allocations should rather fail than rely on
3149 * other request to make a forward progress.
3150 * We are in an unfortunate situation where out_of_memory cannot
3151 * do much for this context but let's try it to at least get
3152 * access to memory reserved if the current task is killed (see
3153 * out_of_memory). Once filesystems are ready to handle allocation
3154 * failures more gracefully we should just bail out here.
3157 /* The OOM killer may not free memory on a specific node */
3158 if (gfp_mask & __GFP_THISNODE)
3161 /* Exhausted what can be done so it's blamo time */
3162 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3163 *did_some_progress = 1;
3165 if (gfp_mask & __GFP_NOFAIL) {
3166 page = get_page_from_freelist(gfp_mask, order,
3167 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3169 * fallback to ignore cpuset restriction if our nodes
3173 page = get_page_from_freelist(gfp_mask, order,
3174 ALLOC_NO_WATERMARKS, ac);
3178 mutex_unlock(&oom_lock);
3184 * Maximum number of compaction retries wit a progress before OOM
3185 * killer is consider as the only way to move forward.
3187 #define MAX_COMPACT_RETRIES 16
3189 #ifdef CONFIG_COMPACTION
3190 /* Try memory compaction for high-order allocations before reclaim */
3191 static struct page *
3192 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3193 unsigned int alloc_flags, const struct alloc_context *ac,
3194 enum migrate_mode mode, enum compact_result *compact_result)
3197 int contended_compaction;
3202 current->flags |= PF_MEMALLOC;
3203 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3204 mode, &contended_compaction);
3205 current->flags &= ~PF_MEMALLOC;
3207 if (*compact_result <= COMPACT_INACTIVE)
3211 * At least in one zone compaction wasn't deferred or skipped, so let's
3212 * count a compaction stall
3214 count_vm_event(COMPACTSTALL);
3216 page = get_page_from_freelist(gfp_mask, order,
3217 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3220 struct zone *zone = page_zone(page);
3222 zone->compact_blockskip_flush = false;
3223 compaction_defer_reset(zone, order, true);
3224 count_vm_event(COMPACTSUCCESS);
3229 * It's bad if compaction run occurs and fails. The most likely reason
3230 * is that pages exist, but not enough to satisfy watermarks.
3232 count_vm_event(COMPACTFAIL);
3235 * In all zones where compaction was attempted (and not
3236 * deferred or skipped), lock contention has been detected.
3237 * For THP allocation we do not want to disrupt the others
3238 * so we fallback to base pages instead.
3240 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3241 *compact_result = COMPACT_CONTENDED;
3244 * If compaction was aborted due to need_resched(), we do not
3245 * want to further increase allocation latency, unless it is
3246 * khugepaged trying to collapse.
3248 if (contended_compaction == COMPACT_CONTENDED_SCHED
3249 && !(current->flags & PF_KTHREAD))
3250 *compact_result = COMPACT_CONTENDED;
3258 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3259 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3260 int compaction_retries)
3262 int max_retries = MAX_COMPACT_RETRIES;
3268 * compaction considers all the zone as desperately out of memory
3269 * so it doesn't really make much sense to retry except when the
3270 * failure could be caused by weak migration mode.
3272 if (compaction_failed(compact_result)) {
3273 if (*migrate_mode == MIGRATE_ASYNC) {
3274 *migrate_mode = MIGRATE_SYNC_LIGHT;
3281 * make sure the compaction wasn't deferred or didn't bail out early
3282 * due to locks contention before we declare that we should give up.
3283 * But do not retry if the given zonelist is not suitable for
3286 if (compaction_withdrawn(compact_result))
3287 return compaction_zonelist_suitable(ac, order, alloc_flags);
3290 * !costly requests are much more important than __GFP_REPEAT
3291 * costly ones because they are de facto nofail and invoke OOM
3292 * killer to move on while costly can fail and users are ready
3293 * to cope with that. 1/4 retries is rather arbitrary but we
3294 * would need much more detailed feedback from compaction to
3295 * make a better decision.
3297 if (order > PAGE_ALLOC_COSTLY_ORDER)
3299 if (compaction_retries <= max_retries)
3305 static inline struct page *
3306 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3307 unsigned int alloc_flags, const struct alloc_context *ac,
3308 enum migrate_mode mode, enum compact_result *compact_result)
3310 *compact_result = COMPACT_SKIPPED;
3315 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3316 enum compact_result compact_result,
3317 enum migrate_mode *migrate_mode,
3318 int compaction_retries)
3323 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3327 * There are setups with compaction disabled which would prefer to loop
3328 * inside the allocator rather than hit the oom killer prematurely.
3329 * Let's give them a good hope and keep retrying while the order-0
3330 * watermarks are OK.
3332 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3334 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3335 ac_classzone_idx(ac), alloc_flags))
3340 #endif /* CONFIG_COMPACTION */
3342 /* Perform direct synchronous page reclaim */
3344 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3345 const struct alloc_context *ac)
3347 struct reclaim_state reclaim_state;
3352 /* We now go into synchronous reclaim */
3353 cpuset_memory_pressure_bump();
3354 current->flags |= PF_MEMALLOC;
3355 lockdep_set_current_reclaim_state(gfp_mask);
3356 reclaim_state.reclaimed_slab = 0;
3357 current->reclaim_state = &reclaim_state;
3359 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3362 current->reclaim_state = NULL;
3363 lockdep_clear_current_reclaim_state();
3364 current->flags &= ~PF_MEMALLOC;
3371 /* The really slow allocator path where we enter direct reclaim */
3372 static inline struct page *
3373 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3374 unsigned int alloc_flags, const struct alloc_context *ac,
3375 unsigned long *did_some_progress)
3377 struct page *page = NULL;
3378 bool drained = false;
3380 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3381 if (unlikely(!(*did_some_progress)))
3385 page = get_page_from_freelist(gfp_mask, order,
3386 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3389 * If an allocation failed after direct reclaim, it could be because
3390 * pages are pinned on the per-cpu lists or in high alloc reserves.
3391 * Shrink them them and try again
3393 if (!page && !drained) {
3394 unreserve_highatomic_pageblock(ac);
3395 drain_all_pages(NULL);
3403 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3408 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3409 ac->high_zoneidx, ac->nodemask)
3410 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3413 static inline unsigned int
3414 gfp_to_alloc_flags(gfp_t gfp_mask)
3416 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3418 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3419 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3422 * The caller may dip into page reserves a bit more if the caller
3423 * cannot run direct reclaim, or if the caller has realtime scheduling
3424 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3425 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3427 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3429 if (gfp_mask & __GFP_ATOMIC) {
3431 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3432 * if it can't schedule.
3434 if (!(gfp_mask & __GFP_NOMEMALLOC))
3435 alloc_flags |= ALLOC_HARDER;
3437 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3438 * comment for __cpuset_node_allowed().
3440 alloc_flags &= ~ALLOC_CPUSET;
3441 } else if (unlikely(rt_task(current)) && !in_interrupt())
3442 alloc_flags |= ALLOC_HARDER;
3444 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3445 if (gfp_mask & __GFP_MEMALLOC)
3446 alloc_flags |= ALLOC_NO_WATERMARKS;
3447 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3448 alloc_flags |= ALLOC_NO_WATERMARKS;
3449 else if (!in_interrupt() &&
3450 ((current->flags & PF_MEMALLOC) ||
3451 unlikely(test_thread_flag(TIF_MEMDIE))))
3452 alloc_flags |= ALLOC_NO_WATERMARKS;
3455 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3456 alloc_flags |= ALLOC_CMA;
3461 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3463 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3466 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3468 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3472 * Maximum number of reclaim retries without any progress before OOM killer
3473 * is consider as the only way to move forward.
3475 #define MAX_RECLAIM_RETRIES 16
3478 * Checks whether it makes sense to retry the reclaim to make a forward progress
3479 * for the given allocation request.
3480 * The reclaim feedback represented by did_some_progress (any progress during
3481 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3482 * any progress in a row) is considered as well as the reclaimable pages on the
3483 * applicable zone list (with a backoff mechanism which is a function of
3484 * no_progress_loops).
3486 * Returns true if a retry is viable or false to enter the oom path.
3489 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3490 struct alloc_context *ac, int alloc_flags,
3491 bool did_some_progress, int no_progress_loops)
3497 * Make sure we converge to OOM if we cannot make any progress
3498 * several times in the row.
3500 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3504 * Keep reclaiming pages while there is a chance this will lead somewhere.
3505 * If none of the target zones can satisfy our allocation request even
3506 * if all reclaimable pages are considered then we are screwed and have
3509 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3511 unsigned long available;
3512 unsigned long reclaimable;
3514 available = reclaimable = zone_reclaimable_pages(zone);
3515 available -= DIV_ROUND_UP(no_progress_loops * available,
3516 MAX_RECLAIM_RETRIES);
3517 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3520 * Would the allocation succeed if we reclaimed the whole
3523 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3524 ac_classzone_idx(ac), alloc_flags, available)) {
3526 * If we didn't make any progress and have a lot of
3527 * dirty + writeback pages then we should wait for
3528 * an IO to complete to slow down the reclaim and
3529 * prevent from pre mature OOM
3531 if (!did_some_progress) {
3532 unsigned long writeback;
3533 unsigned long dirty;
3535 writeback = zone_page_state_snapshot(zone,
3537 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3539 if (2*(writeback + dirty) > reclaimable) {
3540 congestion_wait(BLK_RW_ASYNC, HZ/10);
3546 * Memory allocation/reclaim might be called from a WQ
3547 * context and the current implementation of the WQ
3548 * concurrency control doesn't recognize that
3549 * a particular WQ is congested if the worker thread is
3550 * looping without ever sleeping. Therefore we have to
3551 * do a short sleep here rather than calling
3554 if (current->flags & PF_WQ_WORKER)
3555 schedule_timeout_uninterruptible(1);
3566 static inline struct page *
3567 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3568 struct alloc_context *ac)
3570 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3571 struct page *page = NULL;
3572 unsigned int alloc_flags;
3573 unsigned long did_some_progress;
3574 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3575 enum compact_result compact_result;
3576 int compaction_retries = 0;
3577 int no_progress_loops = 0;
3580 * In the slowpath, we sanity check order to avoid ever trying to
3581 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3582 * be using allocators in order of preference for an area that is
3585 if (order >= MAX_ORDER) {
3586 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3591 * We also sanity check to catch abuse of atomic reserves being used by
3592 * callers that are not in atomic context.
3594 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3595 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3596 gfp_mask &= ~__GFP_ATOMIC;
3599 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3600 wake_all_kswapds(order, ac);
3603 * OK, we're below the kswapd watermark and have kicked background
3604 * reclaim. Now things get more complex, so set up alloc_flags according
3605 * to how we want to proceed.
3607 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3609 /* This is the last chance, in general, before the goto nopage. */
3610 page = get_page_from_freelist(gfp_mask, order,
3611 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3615 /* Allocate without watermarks if the context allows */
3616 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3618 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3619 * the allocation is high priority and these type of
3620 * allocations are system rather than user orientated
3622 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3623 page = get_page_from_freelist(gfp_mask, order,
3624 ALLOC_NO_WATERMARKS, ac);
3629 /* Caller is not willing to reclaim, we can't balance anything */
3630 if (!can_direct_reclaim) {
3632 * All existing users of the __GFP_NOFAIL are blockable, so warn
3633 * of any new users that actually allow this type of allocation
3636 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3640 /* Avoid recursion of direct reclaim */
3641 if (current->flags & PF_MEMALLOC) {
3643 * __GFP_NOFAIL request from this context is rather bizarre
3644 * because we cannot reclaim anything and only can loop waiting
3645 * for somebody to do a work for us.
3647 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3654 /* Avoid allocations with no watermarks from looping endlessly */
3655 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3659 * Try direct compaction. The first pass is asynchronous. Subsequent
3660 * attempts after direct reclaim are synchronous
3662 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3668 /* Checks for THP-specific high-order allocations */
3669 if (is_thp_gfp_mask(gfp_mask)) {
3671 * If compaction is deferred for high-order allocations, it is
3672 * because sync compaction recently failed. If this is the case
3673 * and the caller requested a THP allocation, we do not want
3674 * to heavily disrupt the system, so we fail the allocation
3675 * instead of entering direct reclaim.
3677 if (compact_result == COMPACT_DEFERRED)
3681 * Compaction is contended so rather back off than cause
3684 if(compact_result == COMPACT_CONTENDED)
3688 if (order && compaction_made_progress(compact_result))
3689 compaction_retries++;
3691 /* Try direct reclaim and then allocating */
3692 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3693 &did_some_progress);
3697 /* Do not loop if specifically requested */
3698 if (gfp_mask & __GFP_NORETRY)
3702 * Do not retry costly high order allocations unless they are
3705 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3709 * Costly allocations might have made a progress but this doesn't mean
3710 * their order will become available due to high fragmentation so
3711 * always increment the no progress counter for them
3713 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3714 no_progress_loops = 0;
3716 no_progress_loops++;
3718 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3719 did_some_progress > 0, no_progress_loops))
3723 * It doesn't make any sense to retry for the compaction if the order-0
3724 * reclaim is not able to make any progress because the current
3725 * implementation of the compaction depends on the sufficient amount
3726 * of free memory (see __compaction_suitable)
3728 if (did_some_progress > 0 &&
3729 should_compact_retry(ac, order, alloc_flags,
3730 compact_result, &migration_mode,
3731 compaction_retries))
3734 /* Reclaim has failed us, start killing things */
3735 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3739 /* Retry as long as the OOM killer is making progress */
3740 if (did_some_progress) {
3741 no_progress_loops = 0;
3747 * High-order allocations do not necessarily loop after direct reclaim
3748 * and reclaim/compaction depends on compaction being called after
3749 * reclaim so call directly if necessary.
3750 * It can become very expensive to allocate transparent hugepages at
3751 * fault, so use asynchronous memory compaction for THP unless it is
3752 * khugepaged trying to collapse. All other requests should tolerate
3753 * at least light sync migration.
3755 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3756 migration_mode = MIGRATE_ASYNC;
3758 migration_mode = MIGRATE_SYNC_LIGHT;
3759 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3765 warn_alloc_failed(gfp_mask, order, NULL);
3771 * This is the 'heart' of the zoned buddy allocator.
3774 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3775 struct zonelist *zonelist, nodemask_t *nodemask)
3778 unsigned int cpuset_mems_cookie;
3779 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3780 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3781 struct alloc_context ac = {
3782 .high_zoneidx = gfp_zone(gfp_mask),
3783 .zonelist = zonelist,
3784 .nodemask = nodemask,
3785 .migratetype = gfpflags_to_migratetype(gfp_mask),
3788 if (cpusets_enabled()) {
3789 alloc_mask |= __GFP_HARDWALL;
3790 alloc_flags |= ALLOC_CPUSET;
3792 ac.nodemask = &cpuset_current_mems_allowed;
3795 gfp_mask &= gfp_allowed_mask;
3797 lockdep_trace_alloc(gfp_mask);
3799 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3801 if (should_fail_alloc_page(gfp_mask, order))
3805 * Check the zones suitable for the gfp_mask contain at least one
3806 * valid zone. It's possible to have an empty zonelist as a result
3807 * of __GFP_THISNODE and a memoryless node
3809 if (unlikely(!zonelist->_zonerefs->zone))
3812 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3813 alloc_flags |= ALLOC_CMA;
3816 cpuset_mems_cookie = read_mems_allowed_begin();
3818 /* Dirty zone balancing only done in the fast path */
3819 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3821 /* The preferred zone is used for statistics later */
3822 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3823 ac.high_zoneidx, ac.nodemask);
3824 if (!ac.preferred_zoneref) {
3829 /* First allocation attempt */
3830 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3835 * Runtime PM, block IO and its error handling path can deadlock
3836 * because I/O on the device might not complete.
3838 alloc_mask = memalloc_noio_flags(gfp_mask);
3839 ac.spread_dirty_pages = false;
3842 * Restore the original nodemask if it was potentially replaced with
3843 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3845 if (cpusets_enabled())
3846 ac.nodemask = nodemask;
3847 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3851 * When updating a task's mems_allowed, it is possible to race with
3852 * parallel threads in such a way that an allocation can fail while
3853 * the mask is being updated. If a page allocation is about to fail,
3854 * check if the cpuset changed during allocation and if so, retry.
3856 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3857 alloc_mask = gfp_mask;
3862 if (kmemcheck_enabled && page)
3863 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3865 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3869 EXPORT_SYMBOL(__alloc_pages_nodemask);
3872 * Common helper functions.
3874 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3879 * __get_free_pages() returns a 32-bit address, which cannot represent
3882 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3884 page = alloc_pages(gfp_mask, order);
3887 return (unsigned long) page_address(page);
3889 EXPORT_SYMBOL(__get_free_pages);
3891 unsigned long get_zeroed_page(gfp_t gfp_mask)
3893 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3895 EXPORT_SYMBOL(get_zeroed_page);
3897 void __free_pages(struct page *page, unsigned int order)
3899 if (put_page_testzero(page)) {
3901 free_hot_cold_page(page, false);
3903 __free_pages_ok(page, order);
3907 EXPORT_SYMBOL(__free_pages);
3909 void free_pages(unsigned long addr, unsigned int order)
3912 VM_BUG_ON(!virt_addr_valid((void *)addr));
3913 __free_pages(virt_to_page((void *)addr), order);
3917 EXPORT_SYMBOL(free_pages);
3921 * An arbitrary-length arbitrary-offset area of memory which resides
3922 * within a 0 or higher order page. Multiple fragments within that page
3923 * are individually refcounted, in the page's reference counter.
3925 * The page_frag functions below provide a simple allocation framework for
3926 * page fragments. This is used by the network stack and network device
3927 * drivers to provide a backing region of memory for use as either an
3928 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3930 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3933 struct page *page = NULL;
3934 gfp_t gfp = gfp_mask;
3936 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3937 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3939 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3940 PAGE_FRAG_CACHE_MAX_ORDER);
3941 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3943 if (unlikely(!page))
3944 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3946 nc->va = page ? page_address(page) : NULL;
3951 void *__alloc_page_frag(struct page_frag_cache *nc,
3952 unsigned int fragsz, gfp_t gfp_mask)
3954 unsigned int size = PAGE_SIZE;
3958 if (unlikely(!nc->va)) {
3960 page = __page_frag_refill(nc, gfp_mask);
3964 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3965 /* if size can vary use size else just use PAGE_SIZE */
3968 /* Even if we own the page, we do not use atomic_set().
3969 * This would break get_page_unless_zero() users.
3971 page_ref_add(page, size - 1);
3973 /* reset page count bias and offset to start of new frag */
3974 nc->pfmemalloc = page_is_pfmemalloc(page);
3975 nc->pagecnt_bias = size;
3979 offset = nc->offset - fragsz;
3980 if (unlikely(offset < 0)) {
3981 page = virt_to_page(nc->va);
3983 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3986 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3987 /* if size can vary use size else just use PAGE_SIZE */
3990 /* OK, page count is 0, we can safely set it */
3991 set_page_count(page, size);
3993 /* reset page count bias and offset to start of new frag */
3994 nc->pagecnt_bias = size;
3995 offset = size - fragsz;
3999 nc->offset = offset;
4001 return nc->va + offset;
4003 EXPORT_SYMBOL(__alloc_page_frag);
4006 * Frees a page fragment allocated out of either a compound or order 0 page.
4008 void __free_page_frag(void *addr)
4010 struct page *page = virt_to_head_page(addr);
4012 if (unlikely(put_page_testzero(page)))
4013 __free_pages_ok(page, compound_order(page));
4015 EXPORT_SYMBOL(__free_page_frag);
4018 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
4019 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
4020 * equivalent to alloc_pages.
4022 * It should be used when the caller would like to use kmalloc, but since the
4023 * allocation is large, it has to fall back to the page allocator.
4025 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
4029 page = alloc_pages(gfp_mask, order);
4030 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4031 __free_pages(page, order);
4037 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
4041 page = alloc_pages_node(nid, gfp_mask, order);
4042 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4043 __free_pages(page, order);
4050 * __free_kmem_pages and free_kmem_pages will free pages allocated with
4053 void __free_kmem_pages(struct page *page, unsigned int order)
4055 memcg_kmem_uncharge(page, order);
4056 __free_pages(page, order);
4059 void free_kmem_pages(unsigned long addr, unsigned int order)
4062 VM_BUG_ON(!virt_addr_valid((void *)addr));
4063 __free_kmem_pages(virt_to_page((void *)addr), order);
4067 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4071 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4072 unsigned long used = addr + PAGE_ALIGN(size);
4074 split_page(virt_to_page((void *)addr), order);
4075 while (used < alloc_end) {
4080 return (void *)addr;
4084 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4085 * @size: the number of bytes to allocate
4086 * @gfp_mask: GFP flags for the allocation
4088 * This function is similar to alloc_pages(), except that it allocates the
4089 * minimum number of pages to satisfy the request. alloc_pages() can only
4090 * allocate memory in power-of-two pages.
4092 * This function is also limited by MAX_ORDER.
4094 * Memory allocated by this function must be released by free_pages_exact().
4096 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4098 unsigned int order = get_order(size);
4101 addr = __get_free_pages(gfp_mask, order);
4102 return make_alloc_exact(addr, order, size);
4104 EXPORT_SYMBOL(alloc_pages_exact);
4107 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4109 * @nid: the preferred node ID where memory should be allocated
4110 * @size: the number of bytes to allocate
4111 * @gfp_mask: GFP flags for the allocation
4113 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4116 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4118 unsigned int order = get_order(size);
4119 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4122 return make_alloc_exact((unsigned long)page_address(p), order, size);
4126 * free_pages_exact - release memory allocated via alloc_pages_exact()
4127 * @virt: the value returned by alloc_pages_exact.
4128 * @size: size of allocation, same value as passed to alloc_pages_exact().
4130 * Release the memory allocated by a previous call to alloc_pages_exact.
4132 void free_pages_exact(void *virt, size_t size)
4134 unsigned long addr = (unsigned long)virt;
4135 unsigned long end = addr + PAGE_ALIGN(size);
4137 while (addr < end) {
4142 EXPORT_SYMBOL(free_pages_exact);
4145 * nr_free_zone_pages - count number of pages beyond high watermark
4146 * @offset: The zone index of the highest zone
4148 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4149 * high watermark within all zones at or below a given zone index. For each
4150 * zone, the number of pages is calculated as:
4151 * managed_pages - high_pages
4153 static unsigned long nr_free_zone_pages(int offset)
4158 /* Just pick one node, since fallback list is circular */
4159 unsigned long sum = 0;
4161 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4163 for_each_zone_zonelist(zone, z, zonelist, offset) {
4164 unsigned long size = zone->managed_pages;
4165 unsigned long high = high_wmark_pages(zone);
4174 * nr_free_buffer_pages - count number of pages beyond high watermark
4176 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4177 * watermark within ZONE_DMA and ZONE_NORMAL.
4179 unsigned long nr_free_buffer_pages(void)
4181 return nr_free_zone_pages(gfp_zone(GFP_USER));
4183 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4186 * nr_free_pagecache_pages - count number of pages beyond high watermark
4188 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4189 * high watermark within all zones.
4191 unsigned long nr_free_pagecache_pages(void)
4193 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4196 static inline void show_node(struct zone *zone)
4198 if (IS_ENABLED(CONFIG_NUMA))
4199 printk("Node %d ", zone_to_nid(zone));
4202 long si_mem_available(void)
4205 unsigned long pagecache;
4206 unsigned long wmark_low = 0;
4207 unsigned long pages[NR_LRU_LISTS];
4211 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4212 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4215 wmark_low += zone->watermark[WMARK_LOW];
4218 * Estimate the amount of memory available for userspace allocations,
4219 * without causing swapping.
4221 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4224 * Not all the page cache can be freed, otherwise the system will
4225 * start swapping. Assume at least half of the page cache, or the
4226 * low watermark worth of cache, needs to stay.
4228 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4229 pagecache -= min(pagecache / 2, wmark_low);
4230 available += pagecache;
4233 * Part of the reclaimable slab consists of items that are in use,
4234 * and cannot be freed. Cap this estimate at the low watermark.
4236 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4237 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4243 EXPORT_SYMBOL_GPL(si_mem_available);
4245 void si_meminfo(struct sysinfo *val)
4247 val->totalram = totalram_pages;
4248 val->sharedram = global_page_state(NR_SHMEM);
4249 val->freeram = global_page_state(NR_FREE_PAGES);
4250 val->bufferram = nr_blockdev_pages();
4251 val->totalhigh = totalhigh_pages;
4252 val->freehigh = nr_free_highpages();
4253 val->mem_unit = PAGE_SIZE;
4256 EXPORT_SYMBOL(si_meminfo);
4259 void si_meminfo_node(struct sysinfo *val, int nid)
4261 int zone_type; /* needs to be signed */
4262 unsigned long managed_pages = 0;
4263 unsigned long managed_highpages = 0;
4264 unsigned long free_highpages = 0;
4265 pg_data_t *pgdat = NODE_DATA(nid);
4267 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4268 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4269 val->totalram = managed_pages;
4270 val->sharedram = node_page_state(nid, NR_SHMEM);
4271 val->freeram = node_page_state(nid, NR_FREE_PAGES);
4272 #ifdef CONFIG_HIGHMEM
4273 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4274 struct zone *zone = &pgdat->node_zones[zone_type];
4276 if (is_highmem(zone)) {
4277 managed_highpages += zone->managed_pages;
4278 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4281 val->totalhigh = managed_highpages;
4282 val->freehigh = free_highpages;
4284 val->totalhigh = managed_highpages;
4285 val->freehigh = free_highpages;
4287 val->mem_unit = PAGE_SIZE;
4292 * Determine whether the node should be displayed or not, depending on whether
4293 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4295 bool skip_free_areas_node(unsigned int flags, int nid)
4298 unsigned int cpuset_mems_cookie;
4300 if (!(flags & SHOW_MEM_FILTER_NODES))
4304 cpuset_mems_cookie = read_mems_allowed_begin();
4305 ret = !node_isset(nid, cpuset_current_mems_allowed);
4306 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4311 #define K(x) ((x) << (PAGE_SHIFT-10))
4313 static void show_migration_types(unsigned char type)
4315 static const char types[MIGRATE_TYPES] = {
4316 [MIGRATE_UNMOVABLE] = 'U',
4317 [MIGRATE_MOVABLE] = 'M',
4318 [MIGRATE_RECLAIMABLE] = 'E',
4319 [MIGRATE_HIGHATOMIC] = 'H',
4321 [MIGRATE_CMA] = 'C',
4323 #ifdef CONFIG_MEMORY_ISOLATION
4324 [MIGRATE_ISOLATE] = 'I',
4327 char tmp[MIGRATE_TYPES + 1];
4331 for (i = 0; i < MIGRATE_TYPES; i++) {
4332 if (type & (1 << i))
4337 printk("(%s) ", tmp);
4341 * Show free area list (used inside shift_scroll-lock stuff)
4342 * We also calculate the percentage fragmentation. We do this by counting the
4343 * memory on each free list with the exception of the first item on the list.
4346 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4349 void show_free_areas(unsigned int filter)
4351 unsigned long free_pcp = 0;
4355 for_each_populated_zone(zone) {
4356 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4359 for_each_online_cpu(cpu)
4360 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4363 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4364 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4365 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4366 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4367 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4368 " free:%lu free_pcp:%lu free_cma:%lu\n",
4369 global_page_state(NR_ACTIVE_ANON),
4370 global_page_state(NR_INACTIVE_ANON),
4371 global_page_state(NR_ISOLATED_ANON),
4372 global_page_state(NR_ACTIVE_FILE),
4373 global_page_state(NR_INACTIVE_FILE),
4374 global_page_state(NR_ISOLATED_FILE),
4375 global_page_state(NR_UNEVICTABLE),
4376 global_page_state(NR_FILE_DIRTY),
4377 global_page_state(NR_WRITEBACK),
4378 global_page_state(NR_UNSTABLE_NFS),
4379 global_page_state(NR_SLAB_RECLAIMABLE),
4380 global_page_state(NR_SLAB_UNRECLAIMABLE),
4381 global_page_state(NR_FILE_MAPPED),
4382 global_page_state(NR_SHMEM),
4383 global_page_state(NR_PAGETABLE),
4384 global_page_state(NR_BOUNCE),
4385 global_page_state(NR_FREE_PAGES),
4387 global_page_state(NR_FREE_CMA_PAGES));
4389 for_each_populated_zone(zone) {
4392 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4396 for_each_online_cpu(cpu)
4397 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4405 " active_anon:%lukB"
4406 " inactive_anon:%lukB"
4407 " active_file:%lukB"
4408 " inactive_file:%lukB"
4409 " unevictable:%lukB"
4410 " isolated(anon):%lukB"
4411 " isolated(file):%lukB"
4419 " slab_reclaimable:%lukB"
4420 " slab_unreclaimable:%lukB"
4421 " kernel_stack:%lukB"
4428 " writeback_tmp:%lukB"
4429 " pages_scanned:%lu"
4430 " all_unreclaimable? %s"
4433 K(zone_page_state(zone, NR_FREE_PAGES)),
4434 K(min_wmark_pages(zone)),
4435 K(low_wmark_pages(zone)),
4436 K(high_wmark_pages(zone)),
4437 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4438 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4439 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4440 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4441 K(zone_page_state(zone, NR_UNEVICTABLE)),
4442 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4443 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4444 K(zone->present_pages),
4445 K(zone->managed_pages),
4446 K(zone_page_state(zone, NR_MLOCK)),
4447 K(zone_page_state(zone, NR_FILE_DIRTY)),
4448 K(zone_page_state(zone, NR_WRITEBACK)),
4449 K(zone_page_state(zone, NR_FILE_MAPPED)),
4450 K(zone_page_state(zone, NR_SHMEM)),
4451 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4452 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4453 zone_page_state(zone, NR_KERNEL_STACK) *
4455 K(zone_page_state(zone, NR_PAGETABLE)),
4456 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4457 K(zone_page_state(zone, NR_BOUNCE)),
4459 K(this_cpu_read(zone->pageset->pcp.count)),
4460 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4461 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4462 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4463 (!zone_reclaimable(zone) ? "yes" : "no")
4465 printk("lowmem_reserve[]:");
4466 for (i = 0; i < MAX_NR_ZONES; i++)
4467 printk(" %ld", zone->lowmem_reserve[i]);
4471 for_each_populated_zone(zone) {
4473 unsigned long nr[MAX_ORDER], flags, total = 0;
4474 unsigned char types[MAX_ORDER];
4476 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4479 printk("%s: ", zone->name);
4481 spin_lock_irqsave(&zone->lock, flags);
4482 for (order = 0; order < MAX_ORDER; order++) {
4483 struct free_area *area = &zone->free_area[order];
4486 nr[order] = area->nr_free;
4487 total += nr[order] << order;
4490 for (type = 0; type < MIGRATE_TYPES; type++) {
4491 if (!list_empty(&area->free_list[type]))
4492 types[order] |= 1 << type;
4495 spin_unlock_irqrestore(&zone->lock, flags);
4496 for (order = 0; order < MAX_ORDER; order++) {
4497 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4499 show_migration_types(types[order]);
4501 printk("= %lukB\n", K(total));
4504 hugetlb_show_meminfo();
4506 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4508 show_swap_cache_info();
4511 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4513 zoneref->zone = zone;
4514 zoneref->zone_idx = zone_idx(zone);
4518 * Builds allocation fallback zone lists.
4520 * Add all populated zones of a node to the zonelist.
4522 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4526 enum zone_type zone_type = MAX_NR_ZONES;
4530 zone = pgdat->node_zones + zone_type;
4531 if (populated_zone(zone)) {
4532 zoneref_set_zone(zone,
4533 &zonelist->_zonerefs[nr_zones++]);
4534 check_highest_zone(zone_type);
4536 } while (zone_type);
4544 * 0 = automatic detection of better ordering.
4545 * 1 = order by ([node] distance, -zonetype)
4546 * 2 = order by (-zonetype, [node] distance)
4548 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4549 * the same zonelist. So only NUMA can configure this param.
4551 #define ZONELIST_ORDER_DEFAULT 0
4552 #define ZONELIST_ORDER_NODE 1
4553 #define ZONELIST_ORDER_ZONE 2
4555 /* zonelist order in the kernel.
4556 * set_zonelist_order() will set this to NODE or ZONE.
4558 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4559 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4563 /* The value user specified ....changed by config */
4564 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4565 /* string for sysctl */
4566 #define NUMA_ZONELIST_ORDER_LEN 16
4567 char numa_zonelist_order[16] = "default";
4570 * interface for configure zonelist ordering.
4571 * command line option "numa_zonelist_order"
4572 * = "[dD]efault - default, automatic configuration.
4573 * = "[nN]ode - order by node locality, then by zone within node
4574 * = "[zZ]one - order by zone, then by locality within zone
4577 static int __parse_numa_zonelist_order(char *s)
4579 if (*s == 'd' || *s == 'D') {
4580 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4581 } else if (*s == 'n' || *s == 'N') {
4582 user_zonelist_order = ZONELIST_ORDER_NODE;
4583 } else if (*s == 'z' || *s == 'Z') {
4584 user_zonelist_order = ZONELIST_ORDER_ZONE;
4586 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4592 static __init int setup_numa_zonelist_order(char *s)
4599 ret = __parse_numa_zonelist_order(s);
4601 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4605 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4608 * sysctl handler for numa_zonelist_order
4610 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4611 void __user *buffer, size_t *length,
4614 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4616 static DEFINE_MUTEX(zl_order_mutex);
4618 mutex_lock(&zl_order_mutex);
4620 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4624 strcpy(saved_string, (char *)table->data);
4626 ret = proc_dostring(table, write, buffer, length, ppos);
4630 int oldval = user_zonelist_order;
4632 ret = __parse_numa_zonelist_order((char *)table->data);
4635 * bogus value. restore saved string
4637 strncpy((char *)table->data, saved_string,
4638 NUMA_ZONELIST_ORDER_LEN);
4639 user_zonelist_order = oldval;
4640 } else if (oldval != user_zonelist_order) {
4641 mutex_lock(&zonelists_mutex);
4642 build_all_zonelists(NULL, NULL);
4643 mutex_unlock(&zonelists_mutex);
4647 mutex_unlock(&zl_order_mutex);
4652 #define MAX_NODE_LOAD (nr_online_nodes)
4653 static int node_load[MAX_NUMNODES];
4656 * find_next_best_node - find the next node that should appear in a given node's fallback list
4657 * @node: node whose fallback list we're appending
4658 * @used_node_mask: nodemask_t of already used nodes
4660 * We use a number of factors to determine which is the next node that should
4661 * appear on a given node's fallback list. The node should not have appeared
4662 * already in @node's fallback list, and it should be the next closest node
4663 * according to the distance array (which contains arbitrary distance values
4664 * from each node to each node in the system), and should also prefer nodes
4665 * with no CPUs, since presumably they'll have very little allocation pressure
4666 * on them otherwise.
4667 * It returns -1 if no node is found.
4669 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4672 int min_val = INT_MAX;
4673 int best_node = NUMA_NO_NODE;
4674 const struct cpumask *tmp = cpumask_of_node(0);
4676 /* Use the local node if we haven't already */
4677 if (!node_isset(node, *used_node_mask)) {
4678 node_set(node, *used_node_mask);
4682 for_each_node_state(n, N_MEMORY) {
4684 /* Don't want a node to appear more than once */
4685 if (node_isset(n, *used_node_mask))
4688 /* Use the distance array to find the distance */
4689 val = node_distance(node, n);
4691 /* Penalize nodes under us ("prefer the next node") */
4694 /* Give preference to headless and unused nodes */
4695 tmp = cpumask_of_node(n);
4696 if (!cpumask_empty(tmp))
4697 val += PENALTY_FOR_NODE_WITH_CPUS;
4699 /* Slight preference for less loaded node */
4700 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4701 val += node_load[n];
4703 if (val < min_val) {
4710 node_set(best_node, *used_node_mask);
4717 * Build zonelists ordered by node and zones within node.
4718 * This results in maximum locality--normal zone overflows into local
4719 * DMA zone, if any--but risks exhausting DMA zone.
4721 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4724 struct zonelist *zonelist;
4726 zonelist = &pgdat->node_zonelists[0];
4727 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4729 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4730 zonelist->_zonerefs[j].zone = NULL;
4731 zonelist->_zonerefs[j].zone_idx = 0;
4735 * Build gfp_thisnode zonelists
4737 static void build_thisnode_zonelists(pg_data_t *pgdat)
4740 struct zonelist *zonelist;
4742 zonelist = &pgdat->node_zonelists[1];
4743 j = build_zonelists_node(pgdat, zonelist, 0);
4744 zonelist->_zonerefs[j].zone = NULL;
4745 zonelist->_zonerefs[j].zone_idx = 0;
4749 * Build zonelists ordered by zone and nodes within zones.
4750 * This results in conserving DMA zone[s] until all Normal memory is
4751 * exhausted, but results in overflowing to remote node while memory
4752 * may still exist in local DMA zone.
4754 static int node_order[MAX_NUMNODES];
4756 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4759 int zone_type; /* needs to be signed */
4761 struct zonelist *zonelist;
4763 zonelist = &pgdat->node_zonelists[0];
4765 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4766 for (j = 0; j < nr_nodes; j++) {
4767 node = node_order[j];
4768 z = &NODE_DATA(node)->node_zones[zone_type];
4769 if (populated_zone(z)) {
4771 &zonelist->_zonerefs[pos++]);
4772 check_highest_zone(zone_type);
4776 zonelist->_zonerefs[pos].zone = NULL;
4777 zonelist->_zonerefs[pos].zone_idx = 0;
4780 #if defined(CONFIG_64BIT)
4782 * Devices that require DMA32/DMA are relatively rare and do not justify a
4783 * penalty to every machine in case the specialised case applies. Default
4784 * to Node-ordering on 64-bit NUMA machines
4786 static int default_zonelist_order(void)
4788 return ZONELIST_ORDER_NODE;
4792 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4793 * by the kernel. If processes running on node 0 deplete the low memory zone
4794 * then reclaim will occur more frequency increasing stalls and potentially
4795 * be easier to OOM if a large percentage of the zone is under writeback or
4796 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4797 * Hence, default to zone ordering on 32-bit.
4799 static int default_zonelist_order(void)
4801 return ZONELIST_ORDER_ZONE;
4803 #endif /* CONFIG_64BIT */
4805 static void set_zonelist_order(void)
4807 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4808 current_zonelist_order = default_zonelist_order();
4810 current_zonelist_order = user_zonelist_order;
4813 static void build_zonelists(pg_data_t *pgdat)
4816 nodemask_t used_mask;
4817 int local_node, prev_node;
4818 struct zonelist *zonelist;
4819 unsigned int order = current_zonelist_order;
4821 /* initialize zonelists */
4822 for (i = 0; i < MAX_ZONELISTS; i++) {
4823 zonelist = pgdat->node_zonelists + i;
4824 zonelist->_zonerefs[0].zone = NULL;
4825 zonelist->_zonerefs[0].zone_idx = 0;
4828 /* NUMA-aware ordering of nodes */
4829 local_node = pgdat->node_id;
4830 load = nr_online_nodes;
4831 prev_node = local_node;
4832 nodes_clear(used_mask);
4834 memset(node_order, 0, sizeof(node_order));
4837 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4839 * We don't want to pressure a particular node.
4840 * So adding penalty to the first node in same
4841 * distance group to make it round-robin.
4843 if (node_distance(local_node, node) !=
4844 node_distance(local_node, prev_node))
4845 node_load[node] = load;
4849 if (order == ZONELIST_ORDER_NODE)
4850 build_zonelists_in_node_order(pgdat, node);
4852 node_order[i++] = node; /* remember order */
4855 if (order == ZONELIST_ORDER_ZONE) {
4856 /* calculate node order -- i.e., DMA last! */
4857 build_zonelists_in_zone_order(pgdat, i);
4860 build_thisnode_zonelists(pgdat);
4863 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4865 * Return node id of node used for "local" allocations.
4866 * I.e., first node id of first zone in arg node's generic zonelist.
4867 * Used for initializing percpu 'numa_mem', which is used primarily
4868 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4870 int local_memory_node(int node)
4874 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4875 gfp_zone(GFP_KERNEL),
4877 return z->zone->node;
4881 #else /* CONFIG_NUMA */
4883 static void set_zonelist_order(void)
4885 current_zonelist_order = ZONELIST_ORDER_ZONE;
4888 static void build_zonelists(pg_data_t *pgdat)
4890 int node, local_node;
4892 struct zonelist *zonelist;
4894 local_node = pgdat->node_id;
4896 zonelist = &pgdat->node_zonelists[0];
4897 j = build_zonelists_node(pgdat, zonelist, 0);
4900 * Now we build the zonelist so that it contains the zones
4901 * of all the other nodes.
4902 * We don't want to pressure a particular node, so when
4903 * building the zones for node N, we make sure that the
4904 * zones coming right after the local ones are those from
4905 * node N+1 (modulo N)
4907 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4908 if (!node_online(node))
4910 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4912 for (node = 0; node < local_node; node++) {
4913 if (!node_online(node))
4915 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4918 zonelist->_zonerefs[j].zone = NULL;
4919 zonelist->_zonerefs[j].zone_idx = 0;
4922 #endif /* CONFIG_NUMA */
4925 * Boot pageset table. One per cpu which is going to be used for all
4926 * zones and all nodes. The parameters will be set in such a way
4927 * that an item put on a list will immediately be handed over to
4928 * the buddy list. This is safe since pageset manipulation is done
4929 * with interrupts disabled.
4931 * The boot_pagesets must be kept even after bootup is complete for
4932 * unused processors and/or zones. They do play a role for bootstrapping
4933 * hotplugged processors.
4935 * zoneinfo_show() and maybe other functions do
4936 * not check if the processor is online before following the pageset pointer.
4937 * Other parts of the kernel may not check if the zone is available.
4939 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4940 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4941 static void setup_zone_pageset(struct zone *zone);
4944 * Global mutex to protect against size modification of zonelists
4945 * as well as to serialize pageset setup for the new populated zone.
4947 DEFINE_MUTEX(zonelists_mutex);
4949 /* return values int ....just for stop_machine() */
4950 static int __build_all_zonelists(void *data)
4954 pg_data_t *self = data;
4957 memset(node_load, 0, sizeof(node_load));
4960 if (self && !node_online(self->node_id)) {
4961 build_zonelists(self);
4964 for_each_online_node(nid) {
4965 pg_data_t *pgdat = NODE_DATA(nid);
4967 build_zonelists(pgdat);
4971 * Initialize the boot_pagesets that are going to be used
4972 * for bootstrapping processors. The real pagesets for
4973 * each zone will be allocated later when the per cpu
4974 * allocator is available.
4976 * boot_pagesets are used also for bootstrapping offline
4977 * cpus if the system is already booted because the pagesets
4978 * are needed to initialize allocators on a specific cpu too.
4979 * F.e. the percpu allocator needs the page allocator which
4980 * needs the percpu allocator in order to allocate its pagesets
4981 * (a chicken-egg dilemma).
4983 for_each_possible_cpu(cpu) {
4984 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4986 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4988 * We now know the "local memory node" for each node--
4989 * i.e., the node of the first zone in the generic zonelist.
4990 * Set up numa_mem percpu variable for on-line cpus. During
4991 * boot, only the boot cpu should be on-line; we'll init the
4992 * secondary cpus' numa_mem as they come on-line. During
4993 * node/memory hotplug, we'll fixup all on-line cpus.
4995 if (cpu_online(cpu))
4996 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5003 static noinline void __init
5004 build_all_zonelists_init(void)
5006 __build_all_zonelists(NULL);
5007 mminit_verify_zonelist();
5008 cpuset_init_current_mems_allowed();
5012 * Called with zonelists_mutex held always
5013 * unless system_state == SYSTEM_BOOTING.
5015 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5016 * [we're only called with non-NULL zone through __meminit paths] and
5017 * (2) call of __init annotated helper build_all_zonelists_init
5018 * [protected by SYSTEM_BOOTING].
5020 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5022 set_zonelist_order();
5024 if (system_state == SYSTEM_BOOTING) {
5025 build_all_zonelists_init();
5027 #ifdef CONFIG_MEMORY_HOTPLUG
5029 setup_zone_pageset(zone);
5031 /* we have to stop all cpus to guarantee there is no user
5033 stop_machine(__build_all_zonelists, pgdat, NULL);
5034 /* cpuset refresh routine should be here */
5036 vm_total_pages = nr_free_pagecache_pages();
5038 * Disable grouping by mobility if the number of pages in the
5039 * system is too low to allow the mechanism to work. It would be
5040 * more accurate, but expensive to check per-zone. This check is
5041 * made on memory-hotadd so a system can start with mobility
5042 * disabled and enable it later
5044 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5045 page_group_by_mobility_disabled = 1;
5047 page_group_by_mobility_disabled = 0;
5049 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5051 zonelist_order_name[current_zonelist_order],
5052 page_group_by_mobility_disabled ? "off" : "on",
5055 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5060 * Helper functions to size the waitqueue hash table.
5061 * Essentially these want to choose hash table sizes sufficiently
5062 * large so that collisions trying to wait on pages are rare.
5063 * But in fact, the number of active page waitqueues on typical
5064 * systems is ridiculously low, less than 200. So this is even
5065 * conservative, even though it seems large.
5067 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5068 * waitqueues, i.e. the size of the waitq table given the number of pages.
5070 #define PAGES_PER_WAITQUEUE 256
5072 #ifndef CONFIG_MEMORY_HOTPLUG
5073 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5075 unsigned long size = 1;
5077 pages /= PAGES_PER_WAITQUEUE;
5079 while (size < pages)
5083 * Once we have dozens or even hundreds of threads sleeping
5084 * on IO we've got bigger problems than wait queue collision.
5085 * Limit the size of the wait table to a reasonable size.
5087 size = min(size, 4096UL);
5089 return max(size, 4UL);
5093 * A zone's size might be changed by hot-add, so it is not possible to determine
5094 * a suitable size for its wait_table. So we use the maximum size now.
5096 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5098 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5099 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5100 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5102 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5103 * or more by the traditional way. (See above). It equals:
5105 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5106 * ia64(16K page size) : = ( 8G + 4M)byte.
5107 * powerpc (64K page size) : = (32G +16M)byte.
5109 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5116 * This is an integer logarithm so that shifts can be used later
5117 * to extract the more random high bits from the multiplicative
5118 * hash function before the remainder is taken.
5120 static inline unsigned long wait_table_bits(unsigned long size)
5126 * Initially all pages are reserved - free ones are freed
5127 * up by free_all_bootmem() once the early boot process is
5128 * done. Non-atomic initialization, single-pass.
5130 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5131 unsigned long start_pfn, enum memmap_context context)
5133 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5134 unsigned long end_pfn = start_pfn + size;
5135 pg_data_t *pgdat = NODE_DATA(nid);
5137 unsigned long nr_initialised = 0;
5138 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5139 struct memblock_region *r = NULL, *tmp;
5142 if (highest_memmap_pfn < end_pfn - 1)
5143 highest_memmap_pfn = end_pfn - 1;
5146 * Honor reservation requested by the driver for this ZONE_DEVICE
5149 if (altmap && start_pfn == altmap->base_pfn)
5150 start_pfn += altmap->reserve;
5152 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5154 * There can be holes in boot-time mem_map[]s handed to this
5155 * function. They do not exist on hotplugged memory.
5157 if (context != MEMMAP_EARLY)
5160 if (!early_pfn_valid(pfn))
5162 if (!early_pfn_in_nid(pfn, nid))
5164 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5167 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5169 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5170 * from zone_movable_pfn[nid] to end of each node should be
5171 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5173 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5174 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5178 * Check given memblock attribute by firmware which can affect
5179 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5180 * mirrored, it's an overlapped memmap init. skip it.
5182 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5183 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5184 for_each_memblock(memory, tmp)
5185 if (pfn < memblock_region_memory_end_pfn(tmp))
5189 if (pfn >= memblock_region_memory_base_pfn(r) &&
5190 memblock_is_mirror(r)) {
5191 /* already initialized as NORMAL */
5192 pfn = memblock_region_memory_end_pfn(r);
5200 * Mark the block movable so that blocks are reserved for
5201 * movable at startup. This will force kernel allocations
5202 * to reserve their blocks rather than leaking throughout
5203 * the address space during boot when many long-lived
5204 * kernel allocations are made.
5206 * bitmap is created for zone's valid pfn range. but memmap
5207 * can be created for invalid pages (for alignment)
5208 * check here not to call set_pageblock_migratetype() against
5211 if (!(pfn & (pageblock_nr_pages - 1))) {
5212 struct page *page = pfn_to_page(pfn);
5214 __init_single_page(page, pfn, zone, nid);
5215 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5217 __init_single_pfn(pfn, zone, nid);
5222 static void __meminit zone_init_free_lists(struct zone *zone)
5224 unsigned int order, t;
5225 for_each_migratetype_order(order, t) {
5226 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5227 zone->free_area[order].nr_free = 0;
5231 #ifndef __HAVE_ARCH_MEMMAP_INIT
5232 #define memmap_init(size, nid, zone, start_pfn) \
5233 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5236 static int zone_batchsize(struct zone *zone)
5242 * The per-cpu-pages pools are set to around 1000th of the
5243 * size of the zone. But no more than 1/2 of a meg.
5245 * OK, so we don't know how big the cache is. So guess.
5247 batch = zone->managed_pages / 1024;
5248 if (batch * PAGE_SIZE > 512 * 1024)
5249 batch = (512 * 1024) / PAGE_SIZE;
5250 batch /= 4; /* We effectively *= 4 below */
5255 * Clamp the batch to a 2^n - 1 value. Having a power
5256 * of 2 value was found to be more likely to have
5257 * suboptimal cache aliasing properties in some cases.
5259 * For example if 2 tasks are alternately allocating
5260 * batches of pages, one task can end up with a lot
5261 * of pages of one half of the possible page colors
5262 * and the other with pages of the other colors.
5264 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5269 /* The deferral and batching of frees should be suppressed under NOMMU
5272 * The problem is that NOMMU needs to be able to allocate large chunks
5273 * of contiguous memory as there's no hardware page translation to
5274 * assemble apparent contiguous memory from discontiguous pages.
5276 * Queueing large contiguous runs of pages for batching, however,
5277 * causes the pages to actually be freed in smaller chunks. As there
5278 * can be a significant delay between the individual batches being
5279 * recycled, this leads to the once large chunks of space being
5280 * fragmented and becoming unavailable for high-order allocations.
5287 * pcp->high and pcp->batch values are related and dependent on one another:
5288 * ->batch must never be higher then ->high.
5289 * The following function updates them in a safe manner without read side
5292 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5293 * those fields changing asynchronously (acording the the above rule).
5295 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5296 * outside of boot time (or some other assurance that no concurrent updaters
5299 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5300 unsigned long batch)
5302 /* start with a fail safe value for batch */
5306 /* Update high, then batch, in order */
5313 /* a companion to pageset_set_high() */
5314 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5316 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5319 static void pageset_init(struct per_cpu_pageset *p)
5321 struct per_cpu_pages *pcp;
5324 memset(p, 0, sizeof(*p));
5328 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5329 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5332 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5335 pageset_set_batch(p, batch);
5339 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5340 * to the value high for the pageset p.
5342 static void pageset_set_high(struct per_cpu_pageset *p,
5345 unsigned long batch = max(1UL, high / 4);
5346 if ((high / 4) > (PAGE_SHIFT * 8))
5347 batch = PAGE_SHIFT * 8;
5349 pageset_update(&p->pcp, high, batch);
5352 static void pageset_set_high_and_batch(struct zone *zone,
5353 struct per_cpu_pageset *pcp)
5355 if (percpu_pagelist_fraction)
5356 pageset_set_high(pcp,
5357 (zone->managed_pages /
5358 percpu_pagelist_fraction));
5360 pageset_set_batch(pcp, zone_batchsize(zone));
5363 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5365 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5368 pageset_set_high_and_batch(zone, pcp);
5371 static void __meminit setup_zone_pageset(struct zone *zone)
5374 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5375 for_each_possible_cpu(cpu)
5376 zone_pageset_init(zone, cpu);
5380 * Allocate per cpu pagesets and initialize them.
5381 * Before this call only boot pagesets were available.
5383 void __init setup_per_cpu_pageset(void)
5387 for_each_populated_zone(zone)
5388 setup_zone_pageset(zone);
5391 static noinline __init_refok
5392 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5398 * The per-page waitqueue mechanism uses hashed waitqueues
5401 zone->wait_table_hash_nr_entries =
5402 wait_table_hash_nr_entries(zone_size_pages);
5403 zone->wait_table_bits =
5404 wait_table_bits(zone->wait_table_hash_nr_entries);
5405 alloc_size = zone->wait_table_hash_nr_entries
5406 * sizeof(wait_queue_head_t);
5408 if (!slab_is_available()) {
5409 zone->wait_table = (wait_queue_head_t *)
5410 memblock_virt_alloc_node_nopanic(
5411 alloc_size, zone->zone_pgdat->node_id);
5414 * This case means that a zone whose size was 0 gets new memory
5415 * via memory hot-add.
5416 * But it may be the case that a new node was hot-added. In
5417 * this case vmalloc() will not be able to use this new node's
5418 * memory - this wait_table must be initialized to use this new
5419 * node itself as well.
5420 * To use this new node's memory, further consideration will be
5423 zone->wait_table = vmalloc(alloc_size);
5425 if (!zone->wait_table)
5428 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5429 init_waitqueue_head(zone->wait_table + i);
5434 static __meminit void zone_pcp_init(struct zone *zone)
5437 * per cpu subsystem is not up at this point. The following code
5438 * relies on the ability of the linker to provide the
5439 * offset of a (static) per cpu variable into the per cpu area.
5441 zone->pageset = &boot_pageset;
5443 if (populated_zone(zone))
5444 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5445 zone->name, zone->present_pages,
5446 zone_batchsize(zone));
5449 int __meminit init_currently_empty_zone(struct zone *zone,
5450 unsigned long zone_start_pfn,
5453 struct pglist_data *pgdat = zone->zone_pgdat;
5455 ret = zone_wait_table_init(zone, size);
5458 pgdat->nr_zones = zone_idx(zone) + 1;
5460 zone->zone_start_pfn = zone_start_pfn;
5462 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5463 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5465 (unsigned long)zone_idx(zone),
5466 zone_start_pfn, (zone_start_pfn + size));
5468 zone_init_free_lists(zone);
5473 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5474 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5477 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5479 int __meminit __early_pfn_to_nid(unsigned long pfn,
5480 struct mminit_pfnnid_cache *state)
5482 unsigned long start_pfn, end_pfn;
5485 if (state->last_start <= pfn && pfn < state->last_end)
5486 return state->last_nid;
5488 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5490 state->last_start = start_pfn;
5491 state->last_end = end_pfn;
5492 state->last_nid = nid;
5497 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5500 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5501 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5502 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5504 * If an architecture guarantees that all ranges registered contain no holes
5505 * and may be freed, this this function may be used instead of calling
5506 * memblock_free_early_nid() manually.
5508 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5510 unsigned long start_pfn, end_pfn;
5513 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5514 start_pfn = min(start_pfn, max_low_pfn);
5515 end_pfn = min(end_pfn, max_low_pfn);
5517 if (start_pfn < end_pfn)
5518 memblock_free_early_nid(PFN_PHYS(start_pfn),
5519 (end_pfn - start_pfn) << PAGE_SHIFT,
5525 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5526 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5528 * If an architecture guarantees that all ranges registered contain no holes and may
5529 * be freed, this function may be used instead of calling memory_present() manually.
5531 void __init sparse_memory_present_with_active_regions(int nid)
5533 unsigned long start_pfn, end_pfn;
5536 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5537 memory_present(this_nid, start_pfn, end_pfn);
5541 * get_pfn_range_for_nid - Return the start and end page frames for a node
5542 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5543 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5544 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5546 * It returns the start and end page frame of a node based on information
5547 * provided by memblock_set_node(). If called for a node
5548 * with no available memory, a warning is printed and the start and end
5551 void __meminit get_pfn_range_for_nid(unsigned int nid,
5552 unsigned long *start_pfn, unsigned long *end_pfn)
5554 unsigned long this_start_pfn, this_end_pfn;
5560 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5561 *start_pfn = min(*start_pfn, this_start_pfn);
5562 *end_pfn = max(*end_pfn, this_end_pfn);
5565 if (*start_pfn == -1UL)
5570 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5571 * assumption is made that zones within a node are ordered in monotonic
5572 * increasing memory addresses so that the "highest" populated zone is used
5574 static void __init find_usable_zone_for_movable(void)
5577 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5578 if (zone_index == ZONE_MOVABLE)
5581 if (arch_zone_highest_possible_pfn[zone_index] >
5582 arch_zone_lowest_possible_pfn[zone_index])
5586 VM_BUG_ON(zone_index == -1);
5587 movable_zone = zone_index;
5591 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5592 * because it is sized independent of architecture. Unlike the other zones,
5593 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5594 * in each node depending on the size of each node and how evenly kernelcore
5595 * is distributed. This helper function adjusts the zone ranges
5596 * provided by the architecture for a given node by using the end of the
5597 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5598 * zones within a node are in order of monotonic increases memory addresses
5600 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5601 unsigned long zone_type,
5602 unsigned long node_start_pfn,
5603 unsigned long node_end_pfn,
5604 unsigned long *zone_start_pfn,
5605 unsigned long *zone_end_pfn)
5607 /* Only adjust if ZONE_MOVABLE is on this node */
5608 if (zone_movable_pfn[nid]) {
5609 /* Size ZONE_MOVABLE */
5610 if (zone_type == ZONE_MOVABLE) {
5611 *zone_start_pfn = zone_movable_pfn[nid];
5612 *zone_end_pfn = min(node_end_pfn,
5613 arch_zone_highest_possible_pfn[movable_zone]);
5615 /* Check if this whole range is within ZONE_MOVABLE */
5616 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5617 *zone_start_pfn = *zone_end_pfn;
5622 * Return the number of pages a zone spans in a node, including holes
5623 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5625 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5626 unsigned long zone_type,
5627 unsigned long node_start_pfn,
5628 unsigned long node_end_pfn,
5629 unsigned long *zone_start_pfn,
5630 unsigned long *zone_end_pfn,
5631 unsigned long *ignored)
5633 /* When hotadd a new node from cpu_up(), the node should be empty */
5634 if (!node_start_pfn && !node_end_pfn)
5637 /* Get the start and end of the zone */
5638 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5639 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5640 adjust_zone_range_for_zone_movable(nid, zone_type,
5641 node_start_pfn, node_end_pfn,
5642 zone_start_pfn, zone_end_pfn);
5644 /* Check that this node has pages within the zone's required range */
5645 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5648 /* Move the zone boundaries inside the node if necessary */
5649 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5650 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5652 /* Return the spanned pages */
5653 return *zone_end_pfn - *zone_start_pfn;
5657 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5658 * then all holes in the requested range will be accounted for.
5660 unsigned long __meminit __absent_pages_in_range(int nid,
5661 unsigned long range_start_pfn,
5662 unsigned long range_end_pfn)
5664 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5665 unsigned long start_pfn, end_pfn;
5668 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5669 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5670 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5671 nr_absent -= end_pfn - start_pfn;
5677 * absent_pages_in_range - Return number of page frames in holes within a range
5678 * @start_pfn: The start PFN to start searching for holes
5679 * @end_pfn: The end PFN to stop searching for holes
5681 * It returns the number of pages frames in memory holes within a range.
5683 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5684 unsigned long end_pfn)
5686 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5689 /* Return the number of page frames in holes in a zone on a node */
5690 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5691 unsigned long zone_type,
5692 unsigned long node_start_pfn,
5693 unsigned long node_end_pfn,
5694 unsigned long *ignored)
5696 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5697 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5698 unsigned long zone_start_pfn, zone_end_pfn;
5699 unsigned long nr_absent;
5701 /* When hotadd a new node from cpu_up(), the node should be empty */
5702 if (!node_start_pfn && !node_end_pfn)
5705 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5706 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5708 adjust_zone_range_for_zone_movable(nid, zone_type,
5709 node_start_pfn, node_end_pfn,
5710 &zone_start_pfn, &zone_end_pfn);
5711 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5714 * ZONE_MOVABLE handling.
5715 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5718 if (zone_movable_pfn[nid]) {
5719 if (mirrored_kernelcore) {
5720 unsigned long start_pfn, end_pfn;
5721 struct memblock_region *r;
5723 for_each_memblock(memory, r) {
5724 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5725 zone_start_pfn, zone_end_pfn);
5726 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5727 zone_start_pfn, zone_end_pfn);
5729 if (zone_type == ZONE_MOVABLE &&
5730 memblock_is_mirror(r))
5731 nr_absent += end_pfn - start_pfn;
5733 if (zone_type == ZONE_NORMAL &&
5734 !memblock_is_mirror(r))
5735 nr_absent += end_pfn - start_pfn;
5738 if (zone_type == ZONE_NORMAL)
5739 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5746 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5747 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5748 unsigned long zone_type,
5749 unsigned long node_start_pfn,
5750 unsigned long node_end_pfn,
5751 unsigned long *zone_start_pfn,
5752 unsigned long *zone_end_pfn,
5753 unsigned long *zones_size)
5757 *zone_start_pfn = node_start_pfn;
5758 for (zone = 0; zone < zone_type; zone++)
5759 *zone_start_pfn += zones_size[zone];
5761 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5763 return zones_size[zone_type];
5766 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5767 unsigned long zone_type,
5768 unsigned long node_start_pfn,
5769 unsigned long node_end_pfn,
5770 unsigned long *zholes_size)
5775 return zholes_size[zone_type];
5778 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5780 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5781 unsigned long node_start_pfn,
5782 unsigned long node_end_pfn,
5783 unsigned long *zones_size,
5784 unsigned long *zholes_size)
5786 unsigned long realtotalpages = 0, totalpages = 0;
5789 for (i = 0; i < MAX_NR_ZONES; i++) {
5790 struct zone *zone = pgdat->node_zones + i;
5791 unsigned long zone_start_pfn, zone_end_pfn;
5792 unsigned long size, real_size;
5794 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5800 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5801 node_start_pfn, node_end_pfn,
5804 zone->zone_start_pfn = zone_start_pfn;
5806 zone->zone_start_pfn = 0;
5807 zone->spanned_pages = size;
5808 zone->present_pages = real_size;
5811 realtotalpages += real_size;
5814 pgdat->node_spanned_pages = totalpages;
5815 pgdat->node_present_pages = realtotalpages;
5816 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5820 #ifndef CONFIG_SPARSEMEM
5822 * Calculate the size of the zone->blockflags rounded to an unsigned long
5823 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5824 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5825 * round what is now in bits to nearest long in bits, then return it in
5828 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5830 unsigned long usemapsize;
5832 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5833 usemapsize = roundup(zonesize, pageblock_nr_pages);
5834 usemapsize = usemapsize >> pageblock_order;
5835 usemapsize *= NR_PAGEBLOCK_BITS;
5836 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5838 return usemapsize / 8;
5841 static void __init setup_usemap(struct pglist_data *pgdat,
5843 unsigned long zone_start_pfn,
5844 unsigned long zonesize)
5846 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5847 zone->pageblock_flags = NULL;
5849 zone->pageblock_flags =
5850 memblock_virt_alloc_node_nopanic(usemapsize,
5854 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5855 unsigned long zone_start_pfn, unsigned long zonesize) {}
5856 #endif /* CONFIG_SPARSEMEM */
5858 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5860 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5861 void __paginginit set_pageblock_order(void)
5865 /* Check that pageblock_nr_pages has not already been setup */
5866 if (pageblock_order)
5869 if (HPAGE_SHIFT > PAGE_SHIFT)
5870 order = HUGETLB_PAGE_ORDER;
5872 order = MAX_ORDER - 1;
5875 * Assume the largest contiguous order of interest is a huge page.
5876 * This value may be variable depending on boot parameters on IA64 and
5879 pageblock_order = order;
5881 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5884 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5885 * is unused as pageblock_order is set at compile-time. See
5886 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5889 void __paginginit set_pageblock_order(void)
5893 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5895 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5896 unsigned long present_pages)
5898 unsigned long pages = spanned_pages;
5901 * Provide a more accurate estimation if there are holes within
5902 * the zone and SPARSEMEM is in use. If there are holes within the
5903 * zone, each populated memory region may cost us one or two extra
5904 * memmap pages due to alignment because memmap pages for each
5905 * populated regions may not naturally algined on page boundary.
5906 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5908 if (spanned_pages > present_pages + (present_pages >> 4) &&
5909 IS_ENABLED(CONFIG_SPARSEMEM))
5910 pages = present_pages;
5912 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5916 * Set up the zone data structures:
5917 * - mark all pages reserved
5918 * - mark all memory queues empty
5919 * - clear the memory bitmaps
5921 * NOTE: pgdat should get zeroed by caller.
5923 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5926 int nid = pgdat->node_id;
5929 pgdat_resize_init(pgdat);
5930 #ifdef CONFIG_NUMA_BALANCING
5931 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5932 pgdat->numabalancing_migrate_nr_pages = 0;
5933 pgdat->numabalancing_migrate_next_window = jiffies;
5935 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5936 spin_lock_init(&pgdat->split_queue_lock);
5937 INIT_LIST_HEAD(&pgdat->split_queue);
5938 pgdat->split_queue_len = 0;
5940 init_waitqueue_head(&pgdat->kswapd_wait);
5941 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5942 #ifdef CONFIG_COMPACTION
5943 init_waitqueue_head(&pgdat->kcompactd_wait);
5945 pgdat_page_ext_init(pgdat);
5947 for (j = 0; j < MAX_NR_ZONES; j++) {
5948 struct zone *zone = pgdat->node_zones + j;
5949 unsigned long size, realsize, freesize, memmap_pages;
5950 unsigned long zone_start_pfn = zone->zone_start_pfn;
5952 size = zone->spanned_pages;
5953 realsize = freesize = zone->present_pages;
5956 * Adjust freesize so that it accounts for how much memory
5957 * is used by this zone for memmap. This affects the watermark
5958 * and per-cpu initialisations
5960 memmap_pages = calc_memmap_size(size, realsize);
5961 if (!is_highmem_idx(j)) {
5962 if (freesize >= memmap_pages) {
5963 freesize -= memmap_pages;
5966 " %s zone: %lu pages used for memmap\n",
5967 zone_names[j], memmap_pages);
5969 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5970 zone_names[j], memmap_pages, freesize);
5973 /* Account for reserved pages */
5974 if (j == 0 && freesize > dma_reserve) {
5975 freesize -= dma_reserve;
5976 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5977 zone_names[0], dma_reserve);
5980 if (!is_highmem_idx(j))
5981 nr_kernel_pages += freesize;
5982 /* Charge for highmem memmap if there are enough kernel pages */
5983 else if (nr_kernel_pages > memmap_pages * 2)
5984 nr_kernel_pages -= memmap_pages;
5985 nr_all_pages += freesize;
5988 * Set an approximate value for lowmem here, it will be adjusted
5989 * when the bootmem allocator frees pages into the buddy system.
5990 * And all highmem pages will be managed by the buddy system.
5992 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5995 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5997 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5999 zone->name = zone_names[j];
6000 spin_lock_init(&zone->lock);
6001 spin_lock_init(&zone->lru_lock);
6002 zone_seqlock_init(zone);
6003 zone->zone_pgdat = pgdat;
6004 zone_pcp_init(zone);
6006 /* For bootup, initialized properly in watermark setup */
6007 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
6009 lruvec_init(&zone->lruvec);
6013 set_pageblock_order();
6014 setup_usemap(pgdat, zone, zone_start_pfn, size);
6015 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6017 memmap_init(size, nid, j, zone_start_pfn);
6021 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
6023 unsigned long __maybe_unused start = 0;
6024 unsigned long __maybe_unused offset = 0;
6026 /* Skip empty nodes */
6027 if (!pgdat->node_spanned_pages)
6030 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6031 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6032 offset = pgdat->node_start_pfn - start;
6033 /* ia64 gets its own node_mem_map, before this, without bootmem */
6034 if (!pgdat->node_mem_map) {
6035 unsigned long size, end;
6039 * The zone's endpoints aren't required to be MAX_ORDER
6040 * aligned but the node_mem_map endpoints must be in order
6041 * for the buddy allocator to function correctly.
6043 end = pgdat_end_pfn(pgdat);
6044 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6045 size = (end - start) * sizeof(struct page);
6046 map = alloc_remap(pgdat->node_id, size);
6048 map = memblock_virt_alloc_node_nopanic(size,
6050 pgdat->node_mem_map = map + offset;
6052 #ifndef CONFIG_NEED_MULTIPLE_NODES
6054 * With no DISCONTIG, the global mem_map is just set as node 0's
6056 if (pgdat == NODE_DATA(0)) {
6057 mem_map = NODE_DATA(0)->node_mem_map;
6058 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6059 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6061 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6064 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6067 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6068 unsigned long node_start_pfn, unsigned long *zholes_size)
6070 pg_data_t *pgdat = NODE_DATA(nid);
6071 unsigned long start_pfn = 0;
6072 unsigned long end_pfn = 0;
6074 /* pg_data_t should be reset to zero when it's allocated */
6075 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6077 reset_deferred_meminit(pgdat);
6078 pgdat->node_id = nid;
6079 pgdat->node_start_pfn = node_start_pfn;
6080 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6081 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6082 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6083 (u64)start_pfn << PAGE_SHIFT,
6084 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6086 start_pfn = node_start_pfn;
6088 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6089 zones_size, zholes_size);
6091 alloc_node_mem_map(pgdat);
6092 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6093 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6094 nid, (unsigned long)pgdat,
6095 (unsigned long)pgdat->node_mem_map);
6098 free_area_init_core(pgdat);
6101 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6103 #if MAX_NUMNODES > 1
6105 * Figure out the number of possible node ids.
6107 void __init setup_nr_node_ids(void)
6109 unsigned int highest;
6111 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6112 nr_node_ids = highest + 1;
6117 * node_map_pfn_alignment - determine the maximum internode alignment
6119 * This function should be called after node map is populated and sorted.
6120 * It calculates the maximum power of two alignment which can distinguish
6123 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6124 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6125 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6126 * shifted, 1GiB is enough and this function will indicate so.
6128 * This is used to test whether pfn -> nid mapping of the chosen memory
6129 * model has fine enough granularity to avoid incorrect mapping for the
6130 * populated node map.
6132 * Returns the determined alignment in pfn's. 0 if there is no alignment
6133 * requirement (single node).
6135 unsigned long __init node_map_pfn_alignment(void)
6137 unsigned long accl_mask = 0, last_end = 0;
6138 unsigned long start, end, mask;
6142 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6143 if (!start || last_nid < 0 || last_nid == nid) {
6150 * Start with a mask granular enough to pin-point to the
6151 * start pfn and tick off bits one-by-one until it becomes
6152 * too coarse to separate the current node from the last.
6154 mask = ~((1 << __ffs(start)) - 1);
6155 while (mask && last_end <= (start & (mask << 1)))
6158 /* accumulate all internode masks */
6162 /* convert mask to number of pages */
6163 return ~accl_mask + 1;
6166 /* Find the lowest pfn for a node */
6167 static unsigned long __init find_min_pfn_for_node(int nid)
6169 unsigned long min_pfn = ULONG_MAX;
6170 unsigned long start_pfn;
6173 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6174 min_pfn = min(min_pfn, start_pfn);
6176 if (min_pfn == ULONG_MAX) {
6177 pr_warn("Could not find start_pfn for node %d\n", nid);
6185 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6187 * It returns the minimum PFN based on information provided via
6188 * memblock_set_node().
6190 unsigned long __init find_min_pfn_with_active_regions(void)
6192 return find_min_pfn_for_node(MAX_NUMNODES);
6196 * early_calculate_totalpages()
6197 * Sum pages in active regions for movable zone.
6198 * Populate N_MEMORY for calculating usable_nodes.
6200 static unsigned long __init early_calculate_totalpages(void)
6202 unsigned long totalpages = 0;
6203 unsigned long start_pfn, end_pfn;
6206 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6207 unsigned long pages = end_pfn - start_pfn;
6209 totalpages += pages;
6211 node_set_state(nid, N_MEMORY);
6217 * Find the PFN the Movable zone begins in each node. Kernel memory
6218 * is spread evenly between nodes as long as the nodes have enough
6219 * memory. When they don't, some nodes will have more kernelcore than
6222 static void __init find_zone_movable_pfns_for_nodes(void)
6225 unsigned long usable_startpfn;
6226 unsigned long kernelcore_node, kernelcore_remaining;
6227 /* save the state before borrow the nodemask */
6228 nodemask_t saved_node_state = node_states[N_MEMORY];
6229 unsigned long totalpages = early_calculate_totalpages();
6230 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6231 struct memblock_region *r;
6233 /* Need to find movable_zone earlier when movable_node is specified. */
6234 find_usable_zone_for_movable();
6237 * If movable_node is specified, ignore kernelcore and movablecore
6240 if (movable_node_is_enabled()) {
6241 for_each_memblock(memory, r) {
6242 if (!memblock_is_hotpluggable(r))
6247 usable_startpfn = PFN_DOWN(r->base);
6248 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6249 min(usable_startpfn, zone_movable_pfn[nid]) :
6257 * If kernelcore=mirror is specified, ignore movablecore option
6259 if (mirrored_kernelcore) {
6260 bool mem_below_4gb_not_mirrored = false;
6262 for_each_memblock(memory, r) {
6263 if (memblock_is_mirror(r))
6268 usable_startpfn = memblock_region_memory_base_pfn(r);
6270 if (usable_startpfn < 0x100000) {
6271 mem_below_4gb_not_mirrored = true;
6275 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6276 min(usable_startpfn, zone_movable_pfn[nid]) :
6280 if (mem_below_4gb_not_mirrored)
6281 pr_warn("This configuration results in unmirrored kernel memory.");
6287 * If movablecore=nn[KMG] was specified, calculate what size of
6288 * kernelcore that corresponds so that memory usable for
6289 * any allocation type is evenly spread. If both kernelcore
6290 * and movablecore are specified, then the value of kernelcore
6291 * will be used for required_kernelcore if it's greater than
6292 * what movablecore would have allowed.
6294 if (required_movablecore) {
6295 unsigned long corepages;
6298 * Round-up so that ZONE_MOVABLE is at least as large as what
6299 * was requested by the user
6301 required_movablecore =
6302 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6303 required_movablecore = min(totalpages, required_movablecore);
6304 corepages = totalpages - required_movablecore;
6306 required_kernelcore = max(required_kernelcore, corepages);
6310 * If kernelcore was not specified or kernelcore size is larger
6311 * than totalpages, there is no ZONE_MOVABLE.
6313 if (!required_kernelcore || required_kernelcore >= totalpages)
6316 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6317 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6320 /* Spread kernelcore memory as evenly as possible throughout nodes */
6321 kernelcore_node = required_kernelcore / usable_nodes;
6322 for_each_node_state(nid, N_MEMORY) {
6323 unsigned long start_pfn, end_pfn;
6326 * Recalculate kernelcore_node if the division per node
6327 * now exceeds what is necessary to satisfy the requested
6328 * amount of memory for the kernel
6330 if (required_kernelcore < kernelcore_node)
6331 kernelcore_node = required_kernelcore / usable_nodes;
6334 * As the map is walked, we track how much memory is usable
6335 * by the kernel using kernelcore_remaining. When it is
6336 * 0, the rest of the node is usable by ZONE_MOVABLE
6338 kernelcore_remaining = kernelcore_node;
6340 /* Go through each range of PFNs within this node */
6341 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6342 unsigned long size_pages;
6344 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6345 if (start_pfn >= end_pfn)
6348 /* Account for what is only usable for kernelcore */
6349 if (start_pfn < usable_startpfn) {
6350 unsigned long kernel_pages;
6351 kernel_pages = min(end_pfn, usable_startpfn)
6354 kernelcore_remaining -= min(kernel_pages,
6355 kernelcore_remaining);
6356 required_kernelcore -= min(kernel_pages,
6357 required_kernelcore);
6359 /* Continue if range is now fully accounted */
6360 if (end_pfn <= usable_startpfn) {
6363 * Push zone_movable_pfn to the end so
6364 * that if we have to rebalance
6365 * kernelcore across nodes, we will
6366 * not double account here
6368 zone_movable_pfn[nid] = end_pfn;
6371 start_pfn = usable_startpfn;
6375 * The usable PFN range for ZONE_MOVABLE is from
6376 * start_pfn->end_pfn. Calculate size_pages as the
6377 * number of pages used as kernelcore
6379 size_pages = end_pfn - start_pfn;
6380 if (size_pages > kernelcore_remaining)
6381 size_pages = kernelcore_remaining;
6382 zone_movable_pfn[nid] = start_pfn + size_pages;
6385 * Some kernelcore has been met, update counts and
6386 * break if the kernelcore for this node has been
6389 required_kernelcore -= min(required_kernelcore,
6391 kernelcore_remaining -= size_pages;
6392 if (!kernelcore_remaining)
6398 * If there is still required_kernelcore, we do another pass with one
6399 * less node in the count. This will push zone_movable_pfn[nid] further
6400 * along on the nodes that still have memory until kernelcore is
6404 if (usable_nodes && required_kernelcore > usable_nodes)
6408 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6409 for (nid = 0; nid < MAX_NUMNODES; nid++)
6410 zone_movable_pfn[nid] =
6411 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6414 /* restore the node_state */
6415 node_states[N_MEMORY] = saved_node_state;
6418 /* Any regular or high memory on that node ? */
6419 static void check_for_memory(pg_data_t *pgdat, int nid)
6421 enum zone_type zone_type;
6423 if (N_MEMORY == N_NORMAL_MEMORY)
6426 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6427 struct zone *zone = &pgdat->node_zones[zone_type];
6428 if (populated_zone(zone)) {
6429 node_set_state(nid, N_HIGH_MEMORY);
6430 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6431 zone_type <= ZONE_NORMAL)
6432 node_set_state(nid, N_NORMAL_MEMORY);
6439 * free_area_init_nodes - Initialise all pg_data_t and zone data
6440 * @max_zone_pfn: an array of max PFNs for each zone
6442 * This will call free_area_init_node() for each active node in the system.
6443 * Using the page ranges provided by memblock_set_node(), the size of each
6444 * zone in each node and their holes is calculated. If the maximum PFN
6445 * between two adjacent zones match, it is assumed that the zone is empty.
6446 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6447 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6448 * starts where the previous one ended. For example, ZONE_DMA32 starts
6449 * at arch_max_dma_pfn.
6451 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6453 unsigned long start_pfn, end_pfn;
6456 /* Record where the zone boundaries are */
6457 memset(arch_zone_lowest_possible_pfn, 0,
6458 sizeof(arch_zone_lowest_possible_pfn));
6459 memset(arch_zone_highest_possible_pfn, 0,
6460 sizeof(arch_zone_highest_possible_pfn));
6461 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6462 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6463 for (i = 1; i < MAX_NR_ZONES; i++) {
6464 if (i == ZONE_MOVABLE)
6466 arch_zone_lowest_possible_pfn[i] =
6467 arch_zone_highest_possible_pfn[i-1];
6468 arch_zone_highest_possible_pfn[i] =
6469 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6471 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6472 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6474 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6475 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6476 find_zone_movable_pfns_for_nodes();
6478 /* Print out the zone ranges */
6479 pr_info("Zone ranges:\n");
6480 for (i = 0; i < MAX_NR_ZONES; i++) {
6481 if (i == ZONE_MOVABLE)
6483 pr_info(" %-8s ", zone_names[i]);
6484 if (arch_zone_lowest_possible_pfn[i] ==
6485 arch_zone_highest_possible_pfn[i])
6488 pr_cont("[mem %#018Lx-%#018Lx]\n",
6489 (u64)arch_zone_lowest_possible_pfn[i]
6491 ((u64)arch_zone_highest_possible_pfn[i]
6492 << PAGE_SHIFT) - 1);
6495 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6496 pr_info("Movable zone start for each node\n");
6497 for (i = 0; i < MAX_NUMNODES; i++) {
6498 if (zone_movable_pfn[i])
6499 pr_info(" Node %d: %#018Lx\n", i,
6500 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6503 /* Print out the early node map */
6504 pr_info("Early memory node ranges\n");
6505 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6506 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6507 (u64)start_pfn << PAGE_SHIFT,
6508 ((u64)end_pfn << PAGE_SHIFT) - 1);
6510 /* Initialise every node */
6511 mminit_verify_pageflags_layout();
6512 setup_nr_node_ids();
6513 for_each_online_node(nid) {
6514 pg_data_t *pgdat = NODE_DATA(nid);
6515 free_area_init_node(nid, NULL,
6516 find_min_pfn_for_node(nid), NULL);
6518 /* Any memory on that node */
6519 if (pgdat->node_present_pages)
6520 node_set_state(nid, N_MEMORY);
6521 check_for_memory(pgdat, nid);
6525 static int __init cmdline_parse_core(char *p, unsigned long *core)
6527 unsigned long long coremem;
6531 coremem = memparse(p, &p);
6532 *core = coremem >> PAGE_SHIFT;
6534 /* Paranoid check that UL is enough for the coremem value */
6535 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6541 * kernelcore=size sets the amount of memory for use for allocations that
6542 * cannot be reclaimed or migrated.
6544 static int __init cmdline_parse_kernelcore(char *p)
6546 /* parse kernelcore=mirror */
6547 if (parse_option_str(p, "mirror")) {
6548 mirrored_kernelcore = true;
6552 return cmdline_parse_core(p, &required_kernelcore);
6556 * movablecore=size sets the amount of memory for use for allocations that
6557 * can be reclaimed or migrated.
6559 static int __init cmdline_parse_movablecore(char *p)
6561 return cmdline_parse_core(p, &required_movablecore);
6564 early_param("kernelcore", cmdline_parse_kernelcore);
6565 early_param("movablecore", cmdline_parse_movablecore);
6567 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6569 void adjust_managed_page_count(struct page *page, long count)
6571 spin_lock(&managed_page_count_lock);
6572 page_zone(page)->managed_pages += count;
6573 totalram_pages += count;
6574 #ifdef CONFIG_HIGHMEM
6575 if (PageHighMem(page))
6576 totalhigh_pages += count;
6578 spin_unlock(&managed_page_count_lock);
6580 EXPORT_SYMBOL(adjust_managed_page_count);
6582 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6585 unsigned long pages = 0;
6587 start = (void *)PAGE_ALIGN((unsigned long)start);
6588 end = (void *)((unsigned long)end & PAGE_MASK);
6589 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6590 if ((unsigned int)poison <= 0xFF)
6591 memset(pos, poison, PAGE_SIZE);
6592 free_reserved_page(virt_to_page(pos));
6596 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6597 s, pages << (PAGE_SHIFT - 10), start, end);
6601 EXPORT_SYMBOL(free_reserved_area);
6603 #ifdef CONFIG_HIGHMEM
6604 void free_highmem_page(struct page *page)
6606 __free_reserved_page(page);
6608 page_zone(page)->managed_pages++;
6614 void __init mem_init_print_info(const char *str)
6616 unsigned long physpages, codesize, datasize, rosize, bss_size;
6617 unsigned long init_code_size, init_data_size;
6619 physpages = get_num_physpages();
6620 codesize = _etext - _stext;
6621 datasize = _edata - _sdata;
6622 rosize = __end_rodata - __start_rodata;
6623 bss_size = __bss_stop - __bss_start;
6624 init_data_size = __init_end - __init_begin;
6625 init_code_size = _einittext - _sinittext;
6628 * Detect special cases and adjust section sizes accordingly:
6629 * 1) .init.* may be embedded into .data sections
6630 * 2) .init.text.* may be out of [__init_begin, __init_end],
6631 * please refer to arch/tile/kernel/vmlinux.lds.S.
6632 * 3) .rodata.* may be embedded into .text or .data sections.
6634 #define adj_init_size(start, end, size, pos, adj) \
6636 if (start <= pos && pos < end && size > adj) \
6640 adj_init_size(__init_begin, __init_end, init_data_size,
6641 _sinittext, init_code_size);
6642 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6643 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6644 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6645 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6647 #undef adj_init_size
6649 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6650 #ifdef CONFIG_HIGHMEM
6654 nr_free_pages() << (PAGE_SHIFT - 10),
6655 physpages << (PAGE_SHIFT - 10),
6656 codesize >> 10, datasize >> 10, rosize >> 10,
6657 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6658 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6659 totalcma_pages << (PAGE_SHIFT - 10),
6660 #ifdef CONFIG_HIGHMEM
6661 totalhigh_pages << (PAGE_SHIFT - 10),
6663 str ? ", " : "", str ? str : "");
6667 * set_dma_reserve - set the specified number of pages reserved in the first zone
6668 * @new_dma_reserve: The number of pages to mark reserved
6670 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6671 * In the DMA zone, a significant percentage may be consumed by kernel image
6672 * and other unfreeable allocations which can skew the watermarks badly. This
6673 * function may optionally be used to account for unfreeable pages in the
6674 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6675 * smaller per-cpu batchsize.
6677 void __init set_dma_reserve(unsigned long new_dma_reserve)
6679 dma_reserve = new_dma_reserve;
6682 void __init free_area_init(unsigned long *zones_size)
6684 free_area_init_node(0, zones_size,
6685 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6688 static int page_alloc_cpu_notify(struct notifier_block *self,
6689 unsigned long action, void *hcpu)
6691 int cpu = (unsigned long)hcpu;
6693 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6694 lru_add_drain_cpu(cpu);
6698 * Spill the event counters of the dead processor
6699 * into the current processors event counters.
6700 * This artificially elevates the count of the current
6703 vm_events_fold_cpu(cpu);
6706 * Zero the differential counters of the dead processor
6707 * so that the vm statistics are consistent.
6709 * This is only okay since the processor is dead and cannot
6710 * race with what we are doing.
6712 cpu_vm_stats_fold(cpu);
6717 void __init page_alloc_init(void)
6719 hotcpu_notifier(page_alloc_cpu_notify, 0);
6723 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6724 * or min_free_kbytes changes.
6726 static void calculate_totalreserve_pages(void)
6728 struct pglist_data *pgdat;
6729 unsigned long reserve_pages = 0;
6730 enum zone_type i, j;
6732 for_each_online_pgdat(pgdat) {
6733 for (i = 0; i < MAX_NR_ZONES; i++) {
6734 struct zone *zone = pgdat->node_zones + i;
6737 /* Find valid and maximum lowmem_reserve in the zone */
6738 for (j = i; j < MAX_NR_ZONES; j++) {
6739 if (zone->lowmem_reserve[j] > max)
6740 max = zone->lowmem_reserve[j];
6743 /* we treat the high watermark as reserved pages. */
6744 max += high_wmark_pages(zone);
6746 if (max > zone->managed_pages)
6747 max = zone->managed_pages;
6749 zone->totalreserve_pages = max;
6751 reserve_pages += max;
6754 totalreserve_pages = reserve_pages;
6758 * setup_per_zone_lowmem_reserve - called whenever
6759 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6760 * has a correct pages reserved value, so an adequate number of
6761 * pages are left in the zone after a successful __alloc_pages().
6763 static void setup_per_zone_lowmem_reserve(void)
6765 struct pglist_data *pgdat;
6766 enum zone_type j, idx;
6768 for_each_online_pgdat(pgdat) {
6769 for (j = 0; j < MAX_NR_ZONES; j++) {
6770 struct zone *zone = pgdat->node_zones + j;
6771 unsigned long managed_pages = zone->managed_pages;
6773 zone->lowmem_reserve[j] = 0;
6777 struct zone *lower_zone;
6781 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6782 sysctl_lowmem_reserve_ratio[idx] = 1;
6784 lower_zone = pgdat->node_zones + idx;
6785 lower_zone->lowmem_reserve[j] = managed_pages /
6786 sysctl_lowmem_reserve_ratio[idx];
6787 managed_pages += lower_zone->managed_pages;
6792 /* update totalreserve_pages */
6793 calculate_totalreserve_pages();
6796 static void __setup_per_zone_wmarks(void)
6798 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6799 unsigned long lowmem_pages = 0;
6801 unsigned long flags;
6803 /* Calculate total number of !ZONE_HIGHMEM pages */
6804 for_each_zone(zone) {
6805 if (!is_highmem(zone))
6806 lowmem_pages += zone->managed_pages;
6809 for_each_zone(zone) {
6812 spin_lock_irqsave(&zone->lock, flags);
6813 tmp = (u64)pages_min * zone->managed_pages;
6814 do_div(tmp, lowmem_pages);
6815 if (is_highmem(zone)) {
6817 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6818 * need highmem pages, so cap pages_min to a small
6821 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6822 * deltas control asynch page reclaim, and so should
6823 * not be capped for highmem.
6825 unsigned long min_pages;
6827 min_pages = zone->managed_pages / 1024;
6828 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6829 zone->watermark[WMARK_MIN] = min_pages;
6832 * If it's a lowmem zone, reserve a number of pages
6833 * proportionate to the zone's size.
6835 zone->watermark[WMARK_MIN] = tmp;
6839 * Set the kswapd watermarks distance according to the
6840 * scale factor in proportion to available memory, but
6841 * ensure a minimum size on small systems.
6843 tmp = max_t(u64, tmp >> 2,
6844 mult_frac(zone->managed_pages,
6845 watermark_scale_factor, 10000));
6847 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6848 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6850 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6851 high_wmark_pages(zone) - low_wmark_pages(zone) -
6852 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6854 spin_unlock_irqrestore(&zone->lock, flags);
6857 /* update totalreserve_pages */
6858 calculate_totalreserve_pages();
6862 * setup_per_zone_wmarks - called when min_free_kbytes changes
6863 * or when memory is hot-{added|removed}
6865 * Ensures that the watermark[min,low,high] values for each zone are set
6866 * correctly with respect to min_free_kbytes.
6868 void setup_per_zone_wmarks(void)
6870 mutex_lock(&zonelists_mutex);
6871 __setup_per_zone_wmarks();
6872 mutex_unlock(&zonelists_mutex);
6876 * Initialise min_free_kbytes.
6878 * For small machines we want it small (128k min). For large machines
6879 * we want it large (64MB max). But it is not linear, because network
6880 * bandwidth does not increase linearly with machine size. We use
6882 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6883 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6899 int __meminit init_per_zone_wmark_min(void)
6901 unsigned long lowmem_kbytes;
6902 int new_min_free_kbytes;
6904 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6905 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6907 if (new_min_free_kbytes > user_min_free_kbytes) {
6908 min_free_kbytes = new_min_free_kbytes;
6909 if (min_free_kbytes < 128)
6910 min_free_kbytes = 128;
6911 if (min_free_kbytes > 65536)
6912 min_free_kbytes = 65536;
6914 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6915 new_min_free_kbytes, user_min_free_kbytes);
6917 setup_per_zone_wmarks();
6918 refresh_zone_stat_thresholds();
6919 setup_per_zone_lowmem_reserve();
6922 core_initcall(init_per_zone_wmark_min)
6925 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6926 * that we can call two helper functions whenever min_free_kbytes
6929 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6930 void __user *buffer, size_t *length, loff_t *ppos)
6934 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6939 user_min_free_kbytes = min_free_kbytes;
6940 setup_per_zone_wmarks();
6945 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6946 void __user *buffer, size_t *length, loff_t *ppos)
6950 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6955 setup_per_zone_wmarks();
6961 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6962 void __user *buffer, size_t *length, loff_t *ppos)
6967 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6972 zone->min_unmapped_pages = (zone->managed_pages *
6973 sysctl_min_unmapped_ratio) / 100;
6977 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6978 void __user *buffer, size_t *length, loff_t *ppos)
6983 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6988 zone->min_slab_pages = (zone->managed_pages *
6989 sysctl_min_slab_ratio) / 100;
6995 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6996 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6997 * whenever sysctl_lowmem_reserve_ratio changes.
6999 * The reserve ratio obviously has absolutely no relation with the
7000 * minimum watermarks. The lowmem reserve ratio can only make sense
7001 * if in function of the boot time zone sizes.
7003 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7004 void __user *buffer, size_t *length, loff_t *ppos)
7006 proc_dointvec_minmax(table, write, buffer, length, ppos);
7007 setup_per_zone_lowmem_reserve();
7012 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7013 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7014 * pagelist can have before it gets flushed back to buddy allocator.
7016 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7017 void __user *buffer, size_t *length, loff_t *ppos)
7020 int old_percpu_pagelist_fraction;
7023 mutex_lock(&pcp_batch_high_lock);
7024 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7026 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7027 if (!write || ret < 0)
7030 /* Sanity checking to avoid pcp imbalance */
7031 if (percpu_pagelist_fraction &&
7032 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7033 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7039 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7042 for_each_populated_zone(zone) {
7045 for_each_possible_cpu(cpu)
7046 pageset_set_high_and_batch(zone,
7047 per_cpu_ptr(zone->pageset, cpu));
7050 mutex_unlock(&pcp_batch_high_lock);
7055 int hashdist = HASHDIST_DEFAULT;
7057 static int __init set_hashdist(char *str)
7061 hashdist = simple_strtoul(str, &str, 0);
7064 __setup("hashdist=", set_hashdist);
7068 * allocate a large system hash table from bootmem
7069 * - it is assumed that the hash table must contain an exact power-of-2
7070 * quantity of entries
7071 * - limit is the number of hash buckets, not the total allocation size
7073 void *__init alloc_large_system_hash(const char *tablename,
7074 unsigned long bucketsize,
7075 unsigned long numentries,
7078 unsigned int *_hash_shift,
7079 unsigned int *_hash_mask,
7080 unsigned long low_limit,
7081 unsigned long high_limit)
7083 unsigned long long max = high_limit;
7084 unsigned long log2qty, size;
7087 /* allow the kernel cmdline to have a say */
7089 /* round applicable memory size up to nearest megabyte */
7090 numentries = nr_kernel_pages;
7092 /* It isn't necessary when PAGE_SIZE >= 1MB */
7093 if (PAGE_SHIFT < 20)
7094 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7096 /* limit to 1 bucket per 2^scale bytes of low memory */
7097 if (scale > PAGE_SHIFT)
7098 numentries >>= (scale - PAGE_SHIFT);
7100 numentries <<= (PAGE_SHIFT - scale);
7102 /* Make sure we've got at least a 0-order allocation.. */
7103 if (unlikely(flags & HASH_SMALL)) {
7104 /* Makes no sense without HASH_EARLY */
7105 WARN_ON(!(flags & HASH_EARLY));
7106 if (!(numentries >> *_hash_shift)) {
7107 numentries = 1UL << *_hash_shift;
7108 BUG_ON(!numentries);
7110 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7111 numentries = PAGE_SIZE / bucketsize;
7113 numentries = roundup_pow_of_two(numentries);
7115 /* limit allocation size to 1/16 total memory by default */
7117 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7118 do_div(max, bucketsize);
7120 max = min(max, 0x80000000ULL);
7122 if (numentries < low_limit)
7123 numentries = low_limit;
7124 if (numentries > max)
7127 log2qty = ilog2(numentries);
7130 size = bucketsize << log2qty;
7131 if (flags & HASH_EARLY)
7132 table = memblock_virt_alloc_nopanic(size, 0);
7134 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7137 * If bucketsize is not a power-of-two, we may free
7138 * some pages at the end of hash table which
7139 * alloc_pages_exact() automatically does
7141 if (get_order(size) < MAX_ORDER) {
7142 table = alloc_pages_exact(size, GFP_ATOMIC);
7143 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7146 } while (!table && size > PAGE_SIZE && --log2qty);
7149 panic("Failed to allocate %s hash table\n", tablename);
7151 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7152 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7155 *_hash_shift = log2qty;
7157 *_hash_mask = (1 << log2qty) - 1;
7163 * This function checks whether pageblock includes unmovable pages or not.
7164 * If @count is not zero, it is okay to include less @count unmovable pages
7166 * PageLRU check without isolation or lru_lock could race so that
7167 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7168 * expect this function should be exact.
7170 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7171 bool skip_hwpoisoned_pages)
7173 unsigned long pfn, iter, found;
7177 * For avoiding noise data, lru_add_drain_all() should be called
7178 * If ZONE_MOVABLE, the zone never contains unmovable pages
7180 if (zone_idx(zone) == ZONE_MOVABLE)
7182 mt = get_pageblock_migratetype(page);
7183 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7186 pfn = page_to_pfn(page);
7187 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7188 unsigned long check = pfn + iter;
7190 if (!pfn_valid_within(check))
7193 page = pfn_to_page(check);
7196 * Hugepages are not in LRU lists, but they're movable.
7197 * We need not scan over tail pages bacause we don't
7198 * handle each tail page individually in migration.
7200 if (PageHuge(page)) {
7201 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7206 * We can't use page_count without pin a page
7207 * because another CPU can free compound page.
7208 * This check already skips compound tails of THP
7209 * because their page->_refcount is zero at all time.
7211 if (!page_ref_count(page)) {
7212 if (PageBuddy(page))
7213 iter += (1 << page_order(page)) - 1;
7218 * The HWPoisoned page may be not in buddy system, and
7219 * page_count() is not 0.
7221 if (skip_hwpoisoned_pages && PageHWPoison(page))
7227 * If there are RECLAIMABLE pages, we need to check
7228 * it. But now, memory offline itself doesn't call
7229 * shrink_node_slabs() and it still to be fixed.
7232 * If the page is not RAM, page_count()should be 0.
7233 * we don't need more check. This is an _used_ not-movable page.
7235 * The problematic thing here is PG_reserved pages. PG_reserved
7236 * is set to both of a memory hole page and a _used_ kernel
7245 bool is_pageblock_removable_nolock(struct page *page)
7251 * We have to be careful here because we are iterating over memory
7252 * sections which are not zone aware so we might end up outside of
7253 * the zone but still within the section.
7254 * We have to take care about the node as well. If the node is offline
7255 * its NODE_DATA will be NULL - see page_zone.
7257 if (!node_online(page_to_nid(page)))
7260 zone = page_zone(page);
7261 pfn = page_to_pfn(page);
7262 if (!zone_spans_pfn(zone, pfn))
7265 return !has_unmovable_pages(zone, page, 0, true);
7268 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7270 static unsigned long pfn_max_align_down(unsigned long pfn)
7272 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7273 pageblock_nr_pages) - 1);
7276 static unsigned long pfn_max_align_up(unsigned long pfn)
7278 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7279 pageblock_nr_pages));
7282 /* [start, end) must belong to a single zone. */
7283 static int __alloc_contig_migrate_range(struct compact_control *cc,
7284 unsigned long start, unsigned long end)
7286 /* This function is based on compact_zone() from compaction.c. */
7287 unsigned long nr_reclaimed;
7288 unsigned long pfn = start;
7289 unsigned int tries = 0;
7294 while (pfn < end || !list_empty(&cc->migratepages)) {
7295 if (fatal_signal_pending(current)) {
7300 if (list_empty(&cc->migratepages)) {
7301 cc->nr_migratepages = 0;
7302 pfn = isolate_migratepages_range(cc, pfn, end);
7308 } else if (++tries == 5) {
7309 ret = ret < 0 ? ret : -EBUSY;
7313 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7315 cc->nr_migratepages -= nr_reclaimed;
7317 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7318 NULL, 0, cc->mode, MR_CMA);
7321 putback_movable_pages(&cc->migratepages);
7328 * alloc_contig_range() -- tries to allocate given range of pages
7329 * @start: start PFN to allocate
7330 * @end: one-past-the-last PFN to allocate
7331 * @migratetype: migratetype of the underlaying pageblocks (either
7332 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7333 * in range must have the same migratetype and it must
7334 * be either of the two.
7336 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7337 * aligned, however it's the caller's responsibility to guarantee that
7338 * we are the only thread that changes migrate type of pageblocks the
7341 * The PFN range must belong to a single zone.
7343 * Returns zero on success or negative error code. On success all
7344 * pages which PFN is in [start, end) are allocated for the caller and
7345 * need to be freed with free_contig_range().
7347 int alloc_contig_range(unsigned long start, unsigned long end,
7348 unsigned migratetype)
7350 unsigned long outer_start, outer_end;
7354 struct compact_control cc = {
7355 .nr_migratepages = 0,
7357 .zone = page_zone(pfn_to_page(start)),
7358 .mode = MIGRATE_SYNC,
7359 .ignore_skip_hint = true,
7361 INIT_LIST_HEAD(&cc.migratepages);
7364 * What we do here is we mark all pageblocks in range as
7365 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7366 * have different sizes, and due to the way page allocator
7367 * work, we align the range to biggest of the two pages so
7368 * that page allocator won't try to merge buddies from
7369 * different pageblocks and change MIGRATE_ISOLATE to some
7370 * other migration type.
7372 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7373 * migrate the pages from an unaligned range (ie. pages that
7374 * we are interested in). This will put all the pages in
7375 * range back to page allocator as MIGRATE_ISOLATE.
7377 * When this is done, we take the pages in range from page
7378 * allocator removing them from the buddy system. This way
7379 * page allocator will never consider using them.
7381 * This lets us mark the pageblocks back as
7382 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7383 * aligned range but not in the unaligned, original range are
7384 * put back to page allocator so that buddy can use them.
7387 ret = start_isolate_page_range(pfn_max_align_down(start),
7388 pfn_max_align_up(end), migratetype,
7394 * In case of -EBUSY, we'd like to know which page causes problem.
7395 * So, just fall through. We will check it in test_pages_isolated().
7397 ret = __alloc_contig_migrate_range(&cc, start, end);
7398 if (ret && ret != -EBUSY)
7402 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7403 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7404 * more, all pages in [start, end) are free in page allocator.
7405 * What we are going to do is to allocate all pages from
7406 * [start, end) (that is remove them from page allocator).
7408 * The only problem is that pages at the beginning and at the
7409 * end of interesting range may be not aligned with pages that
7410 * page allocator holds, ie. they can be part of higher order
7411 * pages. Because of this, we reserve the bigger range and
7412 * once this is done free the pages we are not interested in.
7414 * We don't have to hold zone->lock here because the pages are
7415 * isolated thus they won't get removed from buddy.
7418 lru_add_drain_all();
7419 drain_all_pages(cc.zone);
7422 outer_start = start;
7423 while (!PageBuddy(pfn_to_page(outer_start))) {
7424 if (++order >= MAX_ORDER) {
7425 outer_start = start;
7428 outer_start &= ~0UL << order;
7431 if (outer_start != start) {
7432 order = page_order(pfn_to_page(outer_start));
7435 * outer_start page could be small order buddy page and
7436 * it doesn't include start page. Adjust outer_start
7437 * in this case to report failed page properly
7438 * on tracepoint in test_pages_isolated()
7440 if (outer_start + (1UL << order) <= start)
7441 outer_start = start;
7444 /* Make sure the range is really isolated. */
7445 if (test_pages_isolated(outer_start, end, false)) {
7446 pr_info("%s: [%lx, %lx) PFNs busy\n",
7447 __func__, outer_start, end);
7452 /* Grab isolated pages from freelists. */
7453 outer_end = isolate_freepages_range(&cc, outer_start, end);
7459 /* Free head and tail (if any) */
7460 if (start != outer_start)
7461 free_contig_range(outer_start, start - outer_start);
7462 if (end != outer_end)
7463 free_contig_range(end, outer_end - end);
7466 undo_isolate_page_range(pfn_max_align_down(start),
7467 pfn_max_align_up(end), migratetype);
7471 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7473 unsigned int count = 0;
7475 for (; nr_pages--; pfn++) {
7476 struct page *page = pfn_to_page(pfn);
7478 count += page_count(page) != 1;
7481 WARN(count != 0, "%d pages are still in use!\n", count);
7485 #ifdef CONFIG_MEMORY_HOTPLUG
7487 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7488 * page high values need to be recalulated.
7490 void __meminit zone_pcp_update(struct zone *zone)
7493 mutex_lock(&pcp_batch_high_lock);
7494 for_each_possible_cpu(cpu)
7495 pageset_set_high_and_batch(zone,
7496 per_cpu_ptr(zone->pageset, cpu));
7497 mutex_unlock(&pcp_batch_high_lock);
7501 void zone_pcp_reset(struct zone *zone)
7503 unsigned long flags;
7505 struct per_cpu_pageset *pset;
7507 /* avoid races with drain_pages() */
7508 local_irq_save(flags);
7509 if (zone->pageset != &boot_pageset) {
7510 for_each_online_cpu(cpu) {
7511 pset = per_cpu_ptr(zone->pageset, cpu);
7512 drain_zonestat(zone, pset);
7514 free_percpu(zone->pageset);
7515 zone->pageset = &boot_pageset;
7517 local_irq_restore(flags);
7520 #ifdef CONFIG_MEMORY_HOTREMOVE
7522 * All pages in the range must be in a single zone and isolated
7523 * before calling this.
7526 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7530 unsigned int order, i;
7532 unsigned long flags;
7533 /* find the first valid pfn */
7534 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7539 zone = page_zone(pfn_to_page(pfn));
7540 spin_lock_irqsave(&zone->lock, flags);
7542 while (pfn < end_pfn) {
7543 if (!pfn_valid(pfn)) {
7547 page = pfn_to_page(pfn);
7549 * The HWPoisoned page may be not in buddy system, and
7550 * page_count() is not 0.
7552 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7554 SetPageReserved(page);
7558 BUG_ON(page_count(page));
7559 BUG_ON(!PageBuddy(page));
7560 order = page_order(page);
7561 #ifdef CONFIG_DEBUG_VM
7562 pr_info("remove from free list %lx %d %lx\n",
7563 pfn, 1 << order, end_pfn);
7565 list_del(&page->lru);
7566 rmv_page_order(page);
7567 zone->free_area[order].nr_free--;
7568 for (i = 0; i < (1 << order); i++)
7569 SetPageReserved((page+i));
7570 pfn += (1 << order);
7572 spin_unlock_irqrestore(&zone->lock, flags);
7576 bool is_free_buddy_page(struct page *page)
7578 struct zone *zone = page_zone(page);
7579 unsigned long pfn = page_to_pfn(page);
7580 unsigned long flags;
7583 spin_lock_irqsave(&zone->lock, flags);
7584 for (order = 0; order < MAX_ORDER; order++) {
7585 struct page *page_head = page - (pfn & ((1 << order) - 1));
7587 if (PageBuddy(page_head) && page_order(page_head) >= order)
7590 spin_unlock_irqrestore(&zone->lock, flags);
7592 return order < MAX_ORDER;