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;
526 * Allow a burst of 60 reports, then keep quiet for that minute;
527 * or allow a steady drip of one report per second.
529 if (nr_shown == 60) {
530 if (time_before(jiffies, resume)) {
536 "BUG: Bad page state: %lu messages suppressed\n",
543 resume = jiffies + 60 * HZ;
545 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
546 current->comm, page_to_pfn(page));
547 __dump_page(page, reason);
548 bad_flags &= page->flags;
550 pr_alert("bad because of flags: %#lx(%pGp)\n",
551 bad_flags, &bad_flags);
552 dump_page_owner(page);
557 /* Leave bad fields for debug, except PageBuddy could make trouble */
558 page_mapcount_reset(page); /* remove PageBuddy */
559 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
563 * Higher-order pages are called "compound pages". They are structured thusly:
565 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
567 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
568 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
570 * The first tail page's ->compound_dtor holds the offset in array of compound
571 * page destructors. See compound_page_dtors.
573 * The first tail page's ->compound_order holds the order of allocation.
574 * This usage means that zero-order pages may not be compound.
577 void free_compound_page(struct page *page)
579 __free_pages_ok(page, compound_order(page));
582 void prep_compound_page(struct page *page, unsigned int order)
585 int nr_pages = 1 << order;
587 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
588 set_compound_order(page, order);
590 for (i = 1; i < nr_pages; i++) {
591 struct page *p = page + i;
592 set_page_count(p, 0);
593 p->mapping = TAIL_MAPPING;
594 set_compound_head(p, page);
596 atomic_set(compound_mapcount_ptr(page), -1);
599 #ifdef CONFIG_DEBUG_PAGEALLOC
600 unsigned int _debug_guardpage_minorder;
601 bool _debug_pagealloc_enabled __read_mostly
602 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
603 EXPORT_SYMBOL(_debug_pagealloc_enabled);
604 bool _debug_guardpage_enabled __read_mostly;
606 static int __init early_debug_pagealloc(char *buf)
610 return kstrtobool(buf, &_debug_pagealloc_enabled);
612 early_param("debug_pagealloc", early_debug_pagealloc);
614 static bool need_debug_guardpage(void)
616 /* If we don't use debug_pagealloc, we don't need guard page */
617 if (!debug_pagealloc_enabled())
623 static void init_debug_guardpage(void)
625 if (!debug_pagealloc_enabled())
628 _debug_guardpage_enabled = true;
631 struct page_ext_operations debug_guardpage_ops = {
632 .need = need_debug_guardpage,
633 .init = init_debug_guardpage,
636 static int __init debug_guardpage_minorder_setup(char *buf)
640 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
641 pr_err("Bad debug_guardpage_minorder value\n");
644 _debug_guardpage_minorder = res;
645 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
648 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
650 static inline void set_page_guard(struct zone *zone, struct page *page,
651 unsigned int order, int migratetype)
653 struct page_ext *page_ext;
655 if (!debug_guardpage_enabled())
658 page_ext = lookup_page_ext(page);
659 if (unlikely(!page_ext))
662 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
664 INIT_LIST_HEAD(&page->lru);
665 set_page_private(page, order);
666 /* Guard pages are not available for any usage */
667 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
670 static inline void clear_page_guard(struct zone *zone, struct page *page,
671 unsigned int order, int migratetype)
673 struct page_ext *page_ext;
675 if (!debug_guardpage_enabled())
678 page_ext = lookup_page_ext(page);
679 if (unlikely(!page_ext))
682 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
684 set_page_private(page, 0);
685 if (!is_migrate_isolate(migratetype))
686 __mod_zone_freepage_state(zone, (1 << order), migratetype);
689 struct page_ext_operations debug_guardpage_ops = { NULL, };
690 static inline void set_page_guard(struct zone *zone, struct page *page,
691 unsigned int order, int migratetype) {}
692 static inline void clear_page_guard(struct zone *zone, struct page *page,
693 unsigned int order, int migratetype) {}
696 static inline void set_page_order(struct page *page, unsigned int order)
698 set_page_private(page, order);
699 __SetPageBuddy(page);
702 static inline void rmv_page_order(struct page *page)
704 __ClearPageBuddy(page);
705 set_page_private(page, 0);
709 * This function checks whether a page is free && is the buddy
710 * we can do coalesce a page and its buddy if
711 * (a) the buddy is not in a hole &&
712 * (b) the buddy is in the buddy system &&
713 * (c) a page and its buddy have the same order &&
714 * (d) a page and its buddy are in the same zone.
716 * For recording whether a page is in the buddy system, we set ->_mapcount
717 * PAGE_BUDDY_MAPCOUNT_VALUE.
718 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
719 * serialized by zone->lock.
721 * For recording page's order, we use page_private(page).
723 static inline int page_is_buddy(struct page *page, struct page *buddy,
726 if (!pfn_valid_within(page_to_pfn(buddy)))
729 if (page_is_guard(buddy) && page_order(buddy) == order) {
730 if (page_zone_id(page) != page_zone_id(buddy))
733 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
738 if (PageBuddy(buddy) && page_order(buddy) == order) {
740 * zone check is done late to avoid uselessly
741 * calculating zone/node ids for pages that could
744 if (page_zone_id(page) != page_zone_id(buddy))
747 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
755 * Freeing function for a buddy system allocator.
757 * The concept of a buddy system is to maintain direct-mapped table
758 * (containing bit values) for memory blocks of various "orders".
759 * The bottom level table contains the map for the smallest allocatable
760 * units of memory (here, pages), and each level above it describes
761 * pairs of units from the levels below, hence, "buddies".
762 * At a high level, all that happens here is marking the table entry
763 * at the bottom level available, and propagating the changes upward
764 * as necessary, plus some accounting needed to play nicely with other
765 * parts of the VM system.
766 * At each level, we keep a list of pages, which are heads of continuous
767 * free pages of length of (1 << order) and marked with _mapcount
768 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
770 * So when we are allocating or freeing one, we can derive the state of the
771 * other. That is, if we allocate a small block, and both were
772 * free, the remainder of the region must be split into blocks.
773 * If a block is freed, and its buddy is also free, then this
774 * triggers coalescing into a block of larger size.
779 static inline void __free_one_page(struct page *page,
781 struct zone *zone, unsigned int order,
784 unsigned long page_idx;
785 unsigned long combined_idx;
786 unsigned long uninitialized_var(buddy_idx);
788 unsigned int max_order;
790 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
792 VM_BUG_ON(!zone_is_initialized(zone));
793 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
795 VM_BUG_ON(migratetype == -1);
796 if (likely(!is_migrate_isolate(migratetype)))
797 __mod_zone_freepage_state(zone, 1 << order, migratetype);
799 page_idx = pfn & ((1 << MAX_ORDER) - 1);
801 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
802 VM_BUG_ON_PAGE(bad_range(zone, page), page);
805 while (order < max_order - 1) {
806 buddy_idx = __find_buddy_index(page_idx, order);
807 buddy = page + (buddy_idx - page_idx);
808 if (!page_is_buddy(page, buddy, order))
811 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
812 * merge with it and move up one order.
814 if (page_is_guard(buddy)) {
815 clear_page_guard(zone, buddy, order, migratetype);
817 list_del(&buddy->lru);
818 zone->free_area[order].nr_free--;
819 rmv_page_order(buddy);
821 combined_idx = buddy_idx & page_idx;
822 page = page + (combined_idx - page_idx);
823 page_idx = combined_idx;
826 if (max_order < MAX_ORDER) {
827 /* If we are here, it means order is >= pageblock_order.
828 * We want to prevent merge between freepages on isolate
829 * pageblock and normal pageblock. Without this, pageblock
830 * isolation could cause incorrect freepage or CMA accounting.
832 * We don't want to hit this code for the more frequent
835 if (unlikely(has_isolate_pageblock(zone))) {
838 buddy_idx = __find_buddy_index(page_idx, order);
839 buddy = page + (buddy_idx - page_idx);
840 buddy_mt = get_pageblock_migratetype(buddy);
842 if (migratetype != buddy_mt
843 && (is_migrate_isolate(migratetype) ||
844 is_migrate_isolate(buddy_mt)))
848 goto continue_merging;
852 set_page_order(page, order);
855 * If this is not the largest possible page, check if the buddy
856 * of the next-highest order is free. If it is, it's possible
857 * that pages are being freed that will coalesce soon. In case,
858 * that is happening, add the free page to the tail of the list
859 * so it's less likely to be used soon and more likely to be merged
860 * as a higher order page
862 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
863 struct page *higher_page, *higher_buddy;
864 combined_idx = buddy_idx & page_idx;
865 higher_page = page + (combined_idx - page_idx);
866 buddy_idx = __find_buddy_index(combined_idx, order + 1);
867 higher_buddy = higher_page + (buddy_idx - combined_idx);
868 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
869 list_add_tail(&page->lru,
870 &zone->free_area[order].free_list[migratetype]);
875 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
877 zone->free_area[order].nr_free++;
881 * A bad page could be due to a number of fields. Instead of multiple branches,
882 * try and check multiple fields with one check. The caller must do a detailed
883 * check if necessary.
885 static inline bool page_expected_state(struct page *page,
886 unsigned long check_flags)
888 if (unlikely(atomic_read(&page->_mapcount) != -1))
891 if (unlikely((unsigned long)page->mapping |
892 page_ref_count(page) |
894 (unsigned long)page->mem_cgroup |
896 (page->flags & check_flags)))
902 static void free_pages_check_bad(struct page *page)
904 const char *bad_reason;
905 unsigned long bad_flags;
910 if (unlikely(atomic_read(&page->_mapcount) != -1))
911 bad_reason = "nonzero mapcount";
912 if (unlikely(page->mapping != NULL))
913 bad_reason = "non-NULL mapping";
914 if (unlikely(page_ref_count(page) != 0))
915 bad_reason = "nonzero _refcount";
916 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
917 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
918 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
921 if (unlikely(page->mem_cgroup))
922 bad_reason = "page still charged to cgroup";
924 bad_page(page, bad_reason, bad_flags);
927 static inline int free_pages_check(struct page *page)
929 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
932 /* Something has gone sideways, find it */
933 free_pages_check_bad(page);
937 static int free_tail_pages_check(struct page *head_page, struct page *page)
942 * We rely page->lru.next never has bit 0 set, unless the page
943 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
945 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
947 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
951 switch (page - head_page) {
953 /* the first tail page: ->mapping is compound_mapcount() */
954 if (unlikely(compound_mapcount(page))) {
955 bad_page(page, "nonzero compound_mapcount", 0);
961 * the second tail page: ->mapping is
962 * page_deferred_list().next -- ignore value.
966 if (page->mapping != TAIL_MAPPING) {
967 bad_page(page, "corrupted mapping in tail page", 0);
972 if (unlikely(!PageTail(page))) {
973 bad_page(page, "PageTail not set", 0);
976 if (unlikely(compound_head(page) != head_page)) {
977 bad_page(page, "compound_head not consistent", 0);
982 page->mapping = NULL;
983 clear_compound_head(page);
987 static __always_inline bool free_pages_prepare(struct page *page,
988 unsigned int order, bool check_free)
992 VM_BUG_ON_PAGE(PageTail(page), page);
994 trace_mm_page_free(page, order);
995 kmemcheck_free_shadow(page, order);
998 * Check tail pages before head page information is cleared to
999 * avoid checking PageCompound for order-0 pages.
1001 if (unlikely(order)) {
1002 bool compound = PageCompound(page);
1005 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1007 for (i = 1; i < (1 << order); i++) {
1009 bad += free_tail_pages_check(page, page + i);
1010 if (unlikely(free_pages_check(page + i))) {
1014 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1017 if (PageAnonHead(page))
1018 page->mapping = NULL;
1020 bad += free_pages_check(page);
1024 page_cpupid_reset_last(page);
1025 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1026 reset_page_owner(page, order);
1028 if (!PageHighMem(page)) {
1029 debug_check_no_locks_freed(page_address(page),
1030 PAGE_SIZE << order);
1031 debug_check_no_obj_freed(page_address(page),
1032 PAGE_SIZE << order);
1034 arch_free_page(page, order);
1035 kernel_poison_pages(page, 1 << order, 0);
1036 kernel_map_pages(page, 1 << order, 0);
1037 kasan_free_pages(page, order);
1042 #ifdef CONFIG_DEBUG_VM
1043 static inline bool free_pcp_prepare(struct page *page)
1045 return free_pages_prepare(page, 0, true);
1048 static inline bool bulkfree_pcp_prepare(struct page *page)
1053 static bool free_pcp_prepare(struct page *page)
1055 return free_pages_prepare(page, 0, false);
1058 static bool bulkfree_pcp_prepare(struct page *page)
1060 return free_pages_check(page);
1062 #endif /* CONFIG_DEBUG_VM */
1065 * Frees a number of pages from the PCP lists
1066 * Assumes all pages on list are in same zone, and of same order.
1067 * count is the number of pages to free.
1069 * If the zone was previously in an "all pages pinned" state then look to
1070 * see if this freeing clears that state.
1072 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1073 * pinned" detection logic.
1075 static void free_pcppages_bulk(struct zone *zone, int count,
1076 struct per_cpu_pages *pcp)
1078 int migratetype = 0;
1080 unsigned long nr_scanned;
1081 bool isolated_pageblocks;
1083 spin_lock(&zone->lock);
1084 isolated_pageblocks = has_isolate_pageblock(zone);
1085 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1087 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1091 struct list_head *list;
1094 * Remove pages from lists in a round-robin fashion. A
1095 * batch_free count is maintained that is incremented when an
1096 * empty list is encountered. This is so more pages are freed
1097 * off fuller lists instead of spinning excessively around empty
1102 if (++migratetype == MIGRATE_PCPTYPES)
1104 list = &pcp->lists[migratetype];
1105 } while (list_empty(list));
1107 /* This is the only non-empty list. Free them all. */
1108 if (batch_free == MIGRATE_PCPTYPES)
1112 int mt; /* migratetype of the to-be-freed page */
1114 page = list_last_entry(list, struct page, lru);
1115 /* must delete as __free_one_page list manipulates */
1116 list_del(&page->lru);
1118 mt = get_pcppage_migratetype(page);
1119 /* MIGRATE_ISOLATE page should not go to pcplists */
1120 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1121 /* Pageblock could have been isolated meanwhile */
1122 if (unlikely(isolated_pageblocks))
1123 mt = get_pageblock_migratetype(page);
1125 if (bulkfree_pcp_prepare(page))
1128 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1129 trace_mm_page_pcpu_drain(page, 0, mt);
1130 } while (--count && --batch_free && !list_empty(list));
1132 spin_unlock(&zone->lock);
1135 static void free_one_page(struct zone *zone,
1136 struct page *page, unsigned long pfn,
1140 unsigned long nr_scanned;
1141 spin_lock(&zone->lock);
1142 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1144 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1146 if (unlikely(has_isolate_pageblock(zone) ||
1147 is_migrate_isolate(migratetype))) {
1148 migratetype = get_pfnblock_migratetype(page, pfn);
1150 __free_one_page(page, pfn, zone, order, migratetype);
1151 spin_unlock(&zone->lock);
1154 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1155 unsigned long zone, int nid)
1157 set_page_links(page, zone, nid, pfn);
1158 init_page_count(page);
1159 page_mapcount_reset(page);
1160 page_cpupid_reset_last(page);
1162 INIT_LIST_HEAD(&page->lru);
1163 #ifdef WANT_PAGE_VIRTUAL
1164 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1165 if (!is_highmem_idx(zone))
1166 set_page_address(page, __va(pfn << PAGE_SHIFT));
1170 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1173 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1176 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1177 static void init_reserved_page(unsigned long pfn)
1182 if (!early_page_uninitialised(pfn))
1185 nid = early_pfn_to_nid(pfn);
1186 pgdat = NODE_DATA(nid);
1188 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1189 struct zone *zone = &pgdat->node_zones[zid];
1191 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1194 __init_single_pfn(pfn, zid, nid);
1197 static inline void init_reserved_page(unsigned long pfn)
1200 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1203 * Initialised pages do not have PageReserved set. This function is
1204 * called for each range allocated by the bootmem allocator and
1205 * marks the pages PageReserved. The remaining valid pages are later
1206 * sent to the buddy page allocator.
1208 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1210 unsigned long start_pfn = PFN_DOWN(start);
1211 unsigned long end_pfn = PFN_UP(end);
1213 for (; start_pfn < end_pfn; start_pfn++) {
1214 if (pfn_valid(start_pfn)) {
1215 struct page *page = pfn_to_page(start_pfn);
1217 init_reserved_page(start_pfn);
1219 /* Avoid false-positive PageTail() */
1220 INIT_LIST_HEAD(&page->lru);
1222 SetPageReserved(page);
1227 static void __free_pages_ok(struct page *page, unsigned int order)
1229 unsigned long flags;
1231 unsigned long pfn = page_to_pfn(page);
1233 if (!free_pages_prepare(page, order, true))
1236 migratetype = get_pfnblock_migratetype(page, pfn);
1237 local_irq_save(flags);
1238 __count_vm_events(PGFREE, 1 << order);
1239 free_one_page(page_zone(page), page, pfn, order, migratetype);
1240 local_irq_restore(flags);
1243 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1245 unsigned int nr_pages = 1 << order;
1246 struct page *p = page;
1250 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1252 __ClearPageReserved(p);
1253 set_page_count(p, 0);
1255 __ClearPageReserved(p);
1256 set_page_count(p, 0);
1258 page_zone(page)->managed_pages += nr_pages;
1259 set_page_refcounted(page);
1260 __free_pages(page, order);
1263 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1264 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1266 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1268 int __meminit early_pfn_to_nid(unsigned long pfn)
1270 static DEFINE_SPINLOCK(early_pfn_lock);
1273 spin_lock(&early_pfn_lock);
1274 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1277 spin_unlock(&early_pfn_lock);
1283 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1284 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1285 struct mminit_pfnnid_cache *state)
1289 nid = __early_pfn_to_nid(pfn, state);
1290 if (nid >= 0 && nid != node)
1295 /* Only safe to use early in boot when initialisation is single-threaded */
1296 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1298 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1303 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1307 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1308 struct mminit_pfnnid_cache *state)
1315 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1318 if (early_page_uninitialised(pfn))
1320 return __free_pages_boot_core(page, order);
1324 * Check that the whole (or subset of) a pageblock given by the interval of
1325 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1326 * with the migration of free compaction scanner. The scanners then need to
1327 * use only pfn_valid_within() check for arches that allow holes within
1330 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1332 * It's possible on some configurations to have a setup like node0 node1 node0
1333 * i.e. it's possible that all pages within a zones range of pages do not
1334 * belong to a single zone. We assume that a border between node0 and node1
1335 * can occur within a single pageblock, but not a node0 node1 node0
1336 * interleaving within a single pageblock. It is therefore sufficient to check
1337 * the first and last page of a pageblock and avoid checking each individual
1338 * page in a pageblock.
1340 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1341 unsigned long end_pfn, struct zone *zone)
1343 struct page *start_page;
1344 struct page *end_page;
1346 /* end_pfn is one past the range we are checking */
1349 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1352 start_page = pfn_to_page(start_pfn);
1354 if (page_zone(start_page) != zone)
1357 end_page = pfn_to_page(end_pfn);
1359 /* This gives a shorter code than deriving page_zone(end_page) */
1360 if (page_zone_id(start_page) != page_zone_id(end_page))
1366 void set_zone_contiguous(struct zone *zone)
1368 unsigned long block_start_pfn = zone->zone_start_pfn;
1369 unsigned long block_end_pfn;
1371 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1372 for (; block_start_pfn < zone_end_pfn(zone);
1373 block_start_pfn = block_end_pfn,
1374 block_end_pfn += pageblock_nr_pages) {
1376 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1378 if (!__pageblock_pfn_to_page(block_start_pfn,
1379 block_end_pfn, zone))
1383 /* We confirm that there is no hole */
1384 zone->contiguous = true;
1387 void clear_zone_contiguous(struct zone *zone)
1389 zone->contiguous = false;
1392 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1393 static void __init deferred_free_range(struct page *page,
1394 unsigned long pfn, int nr_pages)
1401 /* Free a large naturally-aligned chunk if possible */
1402 if (nr_pages == MAX_ORDER_NR_PAGES &&
1403 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1404 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1405 __free_pages_boot_core(page, MAX_ORDER-1);
1409 for (i = 0; i < nr_pages; i++, page++)
1410 __free_pages_boot_core(page, 0);
1413 /* Completion tracking for deferred_init_memmap() threads */
1414 static atomic_t pgdat_init_n_undone __initdata;
1415 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1417 static inline void __init pgdat_init_report_one_done(void)
1419 if (atomic_dec_and_test(&pgdat_init_n_undone))
1420 complete(&pgdat_init_all_done_comp);
1423 /* Initialise remaining memory on a node */
1424 static int __init deferred_init_memmap(void *data)
1426 pg_data_t *pgdat = data;
1427 int nid = pgdat->node_id;
1428 struct mminit_pfnnid_cache nid_init_state = { };
1429 unsigned long start = jiffies;
1430 unsigned long nr_pages = 0;
1431 unsigned long walk_start, walk_end;
1434 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1435 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1437 if (first_init_pfn == ULONG_MAX) {
1438 pgdat_init_report_one_done();
1442 /* Bind memory initialisation thread to a local node if possible */
1443 if (!cpumask_empty(cpumask))
1444 set_cpus_allowed_ptr(current, cpumask);
1446 /* Sanity check boundaries */
1447 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1448 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1449 pgdat->first_deferred_pfn = ULONG_MAX;
1451 /* Only the highest zone is deferred so find it */
1452 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1453 zone = pgdat->node_zones + zid;
1454 if (first_init_pfn < zone_end_pfn(zone))
1458 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1459 unsigned long pfn, end_pfn;
1460 struct page *page = NULL;
1461 struct page *free_base_page = NULL;
1462 unsigned long free_base_pfn = 0;
1465 end_pfn = min(walk_end, zone_end_pfn(zone));
1466 pfn = first_init_pfn;
1467 if (pfn < walk_start)
1469 if (pfn < zone->zone_start_pfn)
1470 pfn = zone->zone_start_pfn;
1472 for (; pfn < end_pfn; pfn++) {
1473 if (!pfn_valid_within(pfn))
1477 * Ensure pfn_valid is checked every
1478 * MAX_ORDER_NR_PAGES for memory holes
1480 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1481 if (!pfn_valid(pfn)) {
1487 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1492 /* Minimise pfn page lookups and scheduler checks */
1493 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1496 nr_pages += nr_to_free;
1497 deferred_free_range(free_base_page,
1498 free_base_pfn, nr_to_free);
1499 free_base_page = NULL;
1500 free_base_pfn = nr_to_free = 0;
1502 page = pfn_to_page(pfn);
1507 VM_BUG_ON(page_zone(page) != zone);
1511 __init_single_page(page, pfn, zid, nid);
1512 if (!free_base_page) {
1513 free_base_page = page;
1514 free_base_pfn = pfn;
1519 /* Where possible, batch up pages for a single free */
1522 /* Free the current block of pages to allocator */
1523 nr_pages += nr_to_free;
1524 deferred_free_range(free_base_page, free_base_pfn,
1526 free_base_page = NULL;
1527 free_base_pfn = nr_to_free = 0;
1530 first_init_pfn = max(end_pfn, first_init_pfn);
1533 /* Sanity check that the next zone really is unpopulated */
1534 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1536 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1537 jiffies_to_msecs(jiffies - start));
1539 pgdat_init_report_one_done();
1542 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1544 void __init page_alloc_init_late(void)
1548 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1551 /* There will be num_node_state(N_MEMORY) threads */
1552 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1553 for_each_node_state(nid, N_MEMORY) {
1554 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1557 /* Block until all are initialised */
1558 wait_for_completion(&pgdat_init_all_done_comp);
1560 /* Reinit limits that are based on free pages after the kernel is up */
1561 files_maxfiles_init();
1564 for_each_populated_zone(zone)
1565 set_zone_contiguous(zone);
1569 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1570 void __init init_cma_reserved_pageblock(struct page *page)
1572 unsigned i = pageblock_nr_pages;
1573 struct page *p = page;
1576 __ClearPageReserved(p);
1577 set_page_count(p, 0);
1580 set_pageblock_migratetype(page, MIGRATE_CMA);
1582 if (pageblock_order >= MAX_ORDER) {
1583 i = pageblock_nr_pages;
1586 set_page_refcounted(p);
1587 __free_pages(p, MAX_ORDER - 1);
1588 p += MAX_ORDER_NR_PAGES;
1589 } while (i -= MAX_ORDER_NR_PAGES);
1591 set_page_refcounted(page);
1592 __free_pages(page, pageblock_order);
1595 adjust_managed_page_count(page, pageblock_nr_pages);
1600 * The order of subdivision here is critical for the IO subsystem.
1601 * Please do not alter this order without good reasons and regression
1602 * testing. Specifically, as large blocks of memory are subdivided,
1603 * the order in which smaller blocks are delivered depends on the order
1604 * they're subdivided in this function. This is the primary factor
1605 * influencing the order in which pages are delivered to the IO
1606 * subsystem according to empirical testing, and this is also justified
1607 * by considering the behavior of a buddy system containing a single
1608 * large block of memory acted on by a series of small allocations.
1609 * This behavior is a critical factor in sglist merging's success.
1613 static inline void expand(struct zone *zone, struct page *page,
1614 int low, int high, struct free_area *area,
1617 unsigned long size = 1 << high;
1619 while (high > low) {
1623 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1625 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1626 debug_guardpage_enabled() &&
1627 high < debug_guardpage_minorder()) {
1629 * Mark as guard pages (or page), that will allow to
1630 * merge back to allocator when buddy will be freed.
1631 * Corresponding page table entries will not be touched,
1632 * pages will stay not present in virtual address space
1634 set_page_guard(zone, &page[size], high, migratetype);
1637 list_add(&page[size].lru, &area->free_list[migratetype]);
1639 set_page_order(&page[size], high);
1643 static void check_new_page_bad(struct page *page)
1645 const char *bad_reason = NULL;
1646 unsigned long bad_flags = 0;
1648 if (unlikely(atomic_read(&page->_mapcount) != -1))
1649 bad_reason = "nonzero mapcount";
1650 if (unlikely(page->mapping != NULL))
1651 bad_reason = "non-NULL mapping";
1652 if (unlikely(page_ref_count(page) != 0))
1653 bad_reason = "nonzero _count";
1654 if (unlikely(page->flags & __PG_HWPOISON)) {
1655 bad_reason = "HWPoisoned (hardware-corrupted)";
1656 bad_flags = __PG_HWPOISON;
1657 /* Don't complain about hwpoisoned pages */
1658 page_mapcount_reset(page); /* remove PageBuddy */
1661 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1662 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1663 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1666 if (unlikely(page->mem_cgroup))
1667 bad_reason = "page still charged to cgroup";
1669 bad_page(page, bad_reason, bad_flags);
1673 * This page is about to be returned from the page allocator
1675 static inline int check_new_page(struct page *page)
1677 if (likely(page_expected_state(page,
1678 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1681 check_new_page_bad(page);
1685 static inline bool free_pages_prezeroed(bool poisoned)
1687 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1688 page_poisoning_enabled() && poisoned;
1691 #ifdef CONFIG_DEBUG_VM
1692 static bool check_pcp_refill(struct page *page)
1697 static bool check_new_pcp(struct page *page)
1699 return check_new_page(page);
1702 static bool check_pcp_refill(struct page *page)
1704 return check_new_page(page);
1706 static bool check_new_pcp(struct page *page)
1710 #endif /* CONFIG_DEBUG_VM */
1712 static bool check_new_pages(struct page *page, unsigned int order)
1715 for (i = 0; i < (1 << order); i++) {
1716 struct page *p = page + i;
1718 if (unlikely(check_new_page(p)))
1725 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1726 unsigned int alloc_flags)
1729 bool poisoned = true;
1731 for (i = 0; i < (1 << order); i++) {
1732 struct page *p = page + i;
1734 poisoned &= page_is_poisoned(p);
1737 set_page_private(page, 0);
1738 set_page_refcounted(page);
1740 arch_alloc_page(page, order);
1741 kernel_map_pages(page, 1 << order, 1);
1742 kernel_poison_pages(page, 1 << order, 1);
1743 kasan_alloc_pages(page, order);
1745 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1746 for (i = 0; i < (1 << order); i++)
1747 clear_highpage(page + i);
1749 if (order && (gfp_flags & __GFP_COMP))
1750 prep_compound_page(page, order);
1752 set_page_owner(page, order, gfp_flags);
1755 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1756 * allocate the page. The expectation is that the caller is taking
1757 * steps that will free more memory. The caller should avoid the page
1758 * being used for !PFMEMALLOC purposes.
1760 if (alloc_flags & ALLOC_NO_WATERMARKS)
1761 set_page_pfmemalloc(page);
1763 clear_page_pfmemalloc(page);
1767 * Go through the free lists for the given migratetype and remove
1768 * the smallest available page from the freelists
1771 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1774 unsigned int current_order;
1775 struct free_area *area;
1778 /* Find a page of the appropriate size in the preferred list */
1779 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1780 area = &(zone->free_area[current_order]);
1781 page = list_first_entry_or_null(&area->free_list[migratetype],
1785 list_del(&page->lru);
1786 rmv_page_order(page);
1788 expand(zone, page, order, current_order, area, migratetype);
1789 set_pcppage_migratetype(page, migratetype);
1798 * This array describes the order lists are fallen back to when
1799 * the free lists for the desirable migrate type are depleted
1801 static int fallbacks[MIGRATE_TYPES][4] = {
1802 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1803 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1804 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1806 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1808 #ifdef CONFIG_MEMORY_ISOLATION
1809 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1814 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1817 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1820 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1821 unsigned int order) { return NULL; }
1825 * Move the free pages in a range to the free lists of the requested type.
1826 * Note that start_page and end_pages are not aligned on a pageblock
1827 * boundary. If alignment is required, use move_freepages_block()
1829 int move_freepages(struct zone *zone,
1830 struct page *start_page, struct page *end_page,
1835 int pages_moved = 0;
1837 #ifndef CONFIG_HOLES_IN_ZONE
1839 * page_zone is not safe to call in this context when
1840 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1841 * anyway as we check zone boundaries in move_freepages_block().
1842 * Remove at a later date when no bug reports exist related to
1843 * grouping pages by mobility
1845 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1848 for (page = start_page; page <= end_page;) {
1849 /* Make sure we are not inadvertently changing nodes */
1850 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1852 if (!pfn_valid_within(page_to_pfn(page))) {
1857 if (!PageBuddy(page)) {
1862 order = page_order(page);
1863 list_move(&page->lru,
1864 &zone->free_area[order].free_list[migratetype]);
1866 pages_moved += 1 << order;
1872 int move_freepages_block(struct zone *zone, struct page *page,
1875 unsigned long start_pfn, end_pfn;
1876 struct page *start_page, *end_page;
1878 start_pfn = page_to_pfn(page);
1879 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1880 start_page = pfn_to_page(start_pfn);
1881 end_page = start_page + pageblock_nr_pages - 1;
1882 end_pfn = start_pfn + pageblock_nr_pages - 1;
1884 /* Do not cross zone boundaries */
1885 if (!zone_spans_pfn(zone, start_pfn))
1887 if (!zone_spans_pfn(zone, end_pfn))
1890 return move_freepages(zone, start_page, end_page, migratetype);
1893 static void change_pageblock_range(struct page *pageblock_page,
1894 int start_order, int migratetype)
1896 int nr_pageblocks = 1 << (start_order - pageblock_order);
1898 while (nr_pageblocks--) {
1899 set_pageblock_migratetype(pageblock_page, migratetype);
1900 pageblock_page += pageblock_nr_pages;
1905 * When we are falling back to another migratetype during allocation, try to
1906 * steal extra free pages from the same pageblocks to satisfy further
1907 * allocations, instead of polluting multiple pageblocks.
1909 * If we are stealing a relatively large buddy page, it is likely there will
1910 * be more free pages in the pageblock, so try to steal them all. For
1911 * reclaimable and unmovable allocations, we steal regardless of page size,
1912 * as fragmentation caused by those allocations polluting movable pageblocks
1913 * is worse than movable allocations stealing from unmovable and reclaimable
1916 static bool can_steal_fallback(unsigned int order, int start_mt)
1919 * Leaving this order check is intended, although there is
1920 * relaxed order check in next check. The reason is that
1921 * we can actually steal whole pageblock if this condition met,
1922 * but, below check doesn't guarantee it and that is just heuristic
1923 * so could be changed anytime.
1925 if (order >= pageblock_order)
1928 if (order >= pageblock_order / 2 ||
1929 start_mt == MIGRATE_RECLAIMABLE ||
1930 start_mt == MIGRATE_UNMOVABLE ||
1931 page_group_by_mobility_disabled)
1938 * This function implements actual steal behaviour. If order is large enough,
1939 * we can steal whole pageblock. If not, we first move freepages in this
1940 * pageblock and check whether half of pages are moved or not. If half of
1941 * pages are moved, we can change migratetype of pageblock and permanently
1942 * use it's pages as requested migratetype in the future.
1944 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1947 unsigned int current_order = page_order(page);
1950 /* Take ownership for orders >= pageblock_order */
1951 if (current_order >= pageblock_order) {
1952 change_pageblock_range(page, current_order, start_type);
1956 pages = move_freepages_block(zone, page, start_type);
1958 /* Claim the whole block if over half of it is free */
1959 if (pages >= (1 << (pageblock_order-1)) ||
1960 page_group_by_mobility_disabled)
1961 set_pageblock_migratetype(page, start_type);
1965 * Check whether there is a suitable fallback freepage with requested order.
1966 * If only_stealable is true, this function returns fallback_mt only if
1967 * we can steal other freepages all together. This would help to reduce
1968 * fragmentation due to mixed migratetype pages in one pageblock.
1970 int find_suitable_fallback(struct free_area *area, unsigned int order,
1971 int migratetype, bool only_stealable, bool *can_steal)
1976 if (area->nr_free == 0)
1981 fallback_mt = fallbacks[migratetype][i];
1982 if (fallback_mt == MIGRATE_TYPES)
1985 if (list_empty(&area->free_list[fallback_mt]))
1988 if (can_steal_fallback(order, migratetype))
1991 if (!only_stealable)
2002 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2003 * there are no empty page blocks that contain a page with a suitable order
2005 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2006 unsigned int alloc_order)
2009 unsigned long max_managed, flags;
2012 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2013 * Check is race-prone but harmless.
2015 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2016 if (zone->nr_reserved_highatomic >= max_managed)
2019 spin_lock_irqsave(&zone->lock, flags);
2021 /* Recheck the nr_reserved_highatomic limit under the lock */
2022 if (zone->nr_reserved_highatomic >= max_managed)
2026 mt = get_pageblock_migratetype(page);
2027 if (mt != MIGRATE_HIGHATOMIC &&
2028 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2029 zone->nr_reserved_highatomic += pageblock_nr_pages;
2030 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2031 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2035 spin_unlock_irqrestore(&zone->lock, flags);
2039 * Used when an allocation is about to fail under memory pressure. This
2040 * potentially hurts the reliability of high-order allocations when under
2041 * intense memory pressure but failed atomic allocations should be easier
2042 * to recover from than an OOM.
2044 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2046 struct zonelist *zonelist = ac->zonelist;
2047 unsigned long flags;
2053 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2055 /* Preserve at least one pageblock */
2056 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2059 spin_lock_irqsave(&zone->lock, flags);
2060 for (order = 0; order < MAX_ORDER; order++) {
2061 struct free_area *area = &(zone->free_area[order]);
2063 page = list_first_entry_or_null(
2064 &area->free_list[MIGRATE_HIGHATOMIC],
2070 * It should never happen but changes to locking could
2071 * inadvertently allow a per-cpu drain to add pages
2072 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2073 * and watch for underflows.
2075 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2076 zone->nr_reserved_highatomic);
2079 * Convert to ac->migratetype and avoid the normal
2080 * pageblock stealing heuristics. Minimally, the caller
2081 * is doing the work and needs the pages. More
2082 * importantly, if the block was always converted to
2083 * MIGRATE_UNMOVABLE or another type then the number
2084 * of pageblocks that cannot be completely freed
2087 set_pageblock_migratetype(page, ac->migratetype);
2088 move_freepages_block(zone, page, ac->migratetype);
2089 spin_unlock_irqrestore(&zone->lock, flags);
2092 spin_unlock_irqrestore(&zone->lock, flags);
2096 /* Remove an element from the buddy allocator from the fallback list */
2097 static inline struct page *
2098 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2100 struct free_area *area;
2101 unsigned int current_order;
2106 /* Find the largest possible block of pages in the other list */
2107 for (current_order = MAX_ORDER-1;
2108 current_order >= order && current_order <= MAX_ORDER-1;
2110 area = &(zone->free_area[current_order]);
2111 fallback_mt = find_suitable_fallback(area, current_order,
2112 start_migratetype, false, &can_steal);
2113 if (fallback_mt == -1)
2116 page = list_first_entry(&area->free_list[fallback_mt],
2119 steal_suitable_fallback(zone, page, start_migratetype);
2121 /* Remove the page from the freelists */
2123 list_del(&page->lru);
2124 rmv_page_order(page);
2126 expand(zone, page, order, current_order, area,
2129 * The pcppage_migratetype may differ from pageblock's
2130 * migratetype depending on the decisions in
2131 * find_suitable_fallback(). This is OK as long as it does not
2132 * differ for MIGRATE_CMA pageblocks. Those can be used as
2133 * fallback only via special __rmqueue_cma_fallback() function
2135 set_pcppage_migratetype(page, start_migratetype);
2137 trace_mm_page_alloc_extfrag(page, order, current_order,
2138 start_migratetype, fallback_mt);
2147 * Do the hard work of removing an element from the buddy allocator.
2148 * Call me with the zone->lock already held.
2150 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2155 page = __rmqueue_smallest(zone, order, migratetype);
2156 if (unlikely(!page)) {
2157 if (migratetype == MIGRATE_MOVABLE)
2158 page = __rmqueue_cma_fallback(zone, order);
2161 page = __rmqueue_fallback(zone, order, migratetype);
2164 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2169 * Obtain a specified number of elements from the buddy allocator, all under
2170 * a single hold of the lock, for efficiency. Add them to the supplied list.
2171 * Returns the number of new pages which were placed at *list.
2173 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2174 unsigned long count, struct list_head *list,
2175 int migratetype, bool cold)
2179 spin_lock(&zone->lock);
2180 for (i = 0; i < count; ++i) {
2181 struct page *page = __rmqueue(zone, order, migratetype);
2182 if (unlikely(page == NULL))
2185 if (unlikely(check_pcp_refill(page)))
2189 * Split buddy pages returned by expand() are received here
2190 * in physical page order. The page is added to the callers and
2191 * list and the list head then moves forward. From the callers
2192 * perspective, the linked list is ordered by page number in
2193 * some conditions. This is useful for IO devices that can
2194 * merge IO requests if the physical pages are ordered
2198 list_add(&page->lru, list);
2200 list_add_tail(&page->lru, list);
2202 if (is_migrate_cma(get_pcppage_migratetype(page)))
2203 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2206 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2207 spin_unlock(&zone->lock);
2213 * Called from the vmstat counter updater to drain pagesets of this
2214 * currently executing processor on remote nodes after they have
2217 * Note that this function must be called with the thread pinned to
2218 * a single processor.
2220 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2222 unsigned long flags;
2223 int to_drain, batch;
2225 local_irq_save(flags);
2226 batch = READ_ONCE(pcp->batch);
2227 to_drain = min(pcp->count, batch);
2229 free_pcppages_bulk(zone, to_drain, pcp);
2230 pcp->count -= to_drain;
2232 local_irq_restore(flags);
2237 * Drain pcplists of the indicated processor and zone.
2239 * The processor must either be the current processor and the
2240 * thread pinned to the current processor or a processor that
2243 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2245 unsigned long flags;
2246 struct per_cpu_pageset *pset;
2247 struct per_cpu_pages *pcp;
2249 local_irq_save(flags);
2250 pset = per_cpu_ptr(zone->pageset, cpu);
2254 free_pcppages_bulk(zone, pcp->count, pcp);
2257 local_irq_restore(flags);
2261 * Drain pcplists of all zones on the indicated processor.
2263 * The processor must either be the current processor and the
2264 * thread pinned to the current processor or a processor that
2267 static void drain_pages(unsigned int cpu)
2271 for_each_populated_zone(zone) {
2272 drain_pages_zone(cpu, zone);
2277 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2279 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2280 * the single zone's pages.
2282 void drain_local_pages(struct zone *zone)
2284 int cpu = smp_processor_id();
2287 drain_pages_zone(cpu, zone);
2293 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2295 * When zone parameter is non-NULL, spill just the single zone's pages.
2297 * Note that this code is protected against sending an IPI to an offline
2298 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2299 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2300 * nothing keeps CPUs from showing up after we populated the cpumask and
2301 * before the call to on_each_cpu_mask().
2303 void drain_all_pages(struct zone *zone)
2308 * Allocate in the BSS so we wont require allocation in
2309 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2311 static cpumask_t cpus_with_pcps;
2314 * We don't care about racing with CPU hotplug event
2315 * as offline notification will cause the notified
2316 * cpu to drain that CPU pcps and on_each_cpu_mask
2317 * disables preemption as part of its processing
2319 for_each_online_cpu(cpu) {
2320 struct per_cpu_pageset *pcp;
2322 bool has_pcps = false;
2325 pcp = per_cpu_ptr(zone->pageset, cpu);
2329 for_each_populated_zone(z) {
2330 pcp = per_cpu_ptr(z->pageset, cpu);
2331 if (pcp->pcp.count) {
2339 cpumask_set_cpu(cpu, &cpus_with_pcps);
2341 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2343 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2347 #ifdef CONFIG_HIBERNATION
2349 void mark_free_pages(struct zone *zone)
2351 unsigned long pfn, max_zone_pfn;
2352 unsigned long flags;
2353 unsigned int order, t;
2356 if (zone_is_empty(zone))
2359 spin_lock_irqsave(&zone->lock, flags);
2361 max_zone_pfn = zone_end_pfn(zone);
2362 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2363 if (pfn_valid(pfn)) {
2364 page = pfn_to_page(pfn);
2366 if (page_zone(page) != zone)
2369 if (!swsusp_page_is_forbidden(page))
2370 swsusp_unset_page_free(page);
2373 for_each_migratetype_order(order, t) {
2374 list_for_each_entry(page,
2375 &zone->free_area[order].free_list[t], lru) {
2378 pfn = page_to_pfn(page);
2379 for (i = 0; i < (1UL << order); i++)
2380 swsusp_set_page_free(pfn_to_page(pfn + i));
2383 spin_unlock_irqrestore(&zone->lock, flags);
2385 #endif /* CONFIG_PM */
2388 * Free a 0-order page
2389 * cold == true ? free a cold page : free a hot page
2391 void free_hot_cold_page(struct page *page, bool cold)
2393 struct zone *zone = page_zone(page);
2394 struct per_cpu_pages *pcp;
2395 unsigned long flags;
2396 unsigned long pfn = page_to_pfn(page);
2399 if (!free_pcp_prepare(page))
2402 migratetype = get_pfnblock_migratetype(page, pfn);
2403 set_pcppage_migratetype(page, migratetype);
2404 local_irq_save(flags);
2405 __count_vm_event(PGFREE);
2408 * We only track unmovable, reclaimable and movable on pcp lists.
2409 * Free ISOLATE pages back to the allocator because they are being
2410 * offlined but treat RESERVE as movable pages so we can get those
2411 * areas back if necessary. Otherwise, we may have to free
2412 * excessively into the page allocator
2414 if (migratetype >= MIGRATE_PCPTYPES) {
2415 if (unlikely(is_migrate_isolate(migratetype))) {
2416 free_one_page(zone, page, pfn, 0, migratetype);
2419 migratetype = MIGRATE_MOVABLE;
2422 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2424 list_add(&page->lru, &pcp->lists[migratetype]);
2426 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2428 if (pcp->count >= pcp->high) {
2429 unsigned long batch = READ_ONCE(pcp->batch);
2430 free_pcppages_bulk(zone, batch, pcp);
2431 pcp->count -= batch;
2435 local_irq_restore(flags);
2439 * Free a list of 0-order pages
2441 void free_hot_cold_page_list(struct list_head *list, bool cold)
2443 struct page *page, *next;
2445 list_for_each_entry_safe(page, next, list, lru) {
2446 trace_mm_page_free_batched(page, cold);
2447 free_hot_cold_page(page, cold);
2452 * split_page takes a non-compound higher-order page, and splits it into
2453 * n (1<<order) sub-pages: page[0..n]
2454 * Each sub-page must be freed individually.
2456 * Note: this is probably too low level an operation for use in drivers.
2457 * Please consult with lkml before using this in your driver.
2459 void split_page(struct page *page, unsigned int order)
2464 VM_BUG_ON_PAGE(PageCompound(page), page);
2465 VM_BUG_ON_PAGE(!page_count(page), page);
2467 #ifdef CONFIG_KMEMCHECK
2469 * Split shadow pages too, because free(page[0]) would
2470 * otherwise free the whole shadow.
2472 if (kmemcheck_page_is_tracked(page))
2473 split_page(virt_to_page(page[0].shadow), order);
2476 gfp_mask = get_page_owner_gfp(page);
2477 set_page_owner(page, 0, gfp_mask);
2478 for (i = 1; i < (1 << order); i++) {
2479 set_page_refcounted(page + i);
2480 set_page_owner(page + i, 0, gfp_mask);
2483 EXPORT_SYMBOL_GPL(split_page);
2485 int __isolate_free_page(struct page *page, unsigned int order)
2487 unsigned long watermark;
2491 BUG_ON(!PageBuddy(page));
2493 zone = page_zone(page);
2494 mt = get_pageblock_migratetype(page);
2496 if (!is_migrate_isolate(mt)) {
2497 /* Obey watermarks as if the page was being allocated */
2498 watermark = low_wmark_pages(zone) + (1 << order);
2499 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2502 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2505 /* Remove page from free list */
2506 list_del(&page->lru);
2507 zone->free_area[order].nr_free--;
2508 rmv_page_order(page);
2510 set_page_owner(page, order, __GFP_MOVABLE);
2512 /* Set the pageblock if the isolated page is at least a pageblock */
2513 if (order >= pageblock_order - 1) {
2514 struct page *endpage = page + (1 << order) - 1;
2515 for (; page < endpage; page += pageblock_nr_pages) {
2516 int mt = get_pageblock_migratetype(page);
2517 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2518 set_pageblock_migratetype(page,
2524 return 1UL << order;
2528 * Similar to split_page except the page is already free. As this is only
2529 * being used for migration, the migratetype of the block also changes.
2530 * As this is called with interrupts disabled, the caller is responsible
2531 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2534 * Note: this is probably too low level an operation for use in drivers.
2535 * Please consult with lkml before using this in your driver.
2537 int split_free_page(struct page *page)
2542 order = page_order(page);
2544 nr_pages = __isolate_free_page(page, order);
2548 /* Split into individual pages */
2549 set_page_refcounted(page);
2550 split_page(page, order);
2555 * Update NUMA hit/miss statistics
2557 * Must be called with interrupts disabled.
2559 * When __GFP_OTHER_NODE is set assume the node of the preferred
2560 * zone is the local node. This is useful for daemons who allocate
2561 * memory on behalf of other processes.
2563 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2567 int local_nid = numa_node_id();
2568 enum zone_stat_item local_stat = NUMA_LOCAL;
2570 if (unlikely(flags & __GFP_OTHER_NODE)) {
2571 local_stat = NUMA_OTHER;
2572 local_nid = preferred_zone->node;
2575 if (z->node == local_nid) {
2576 __inc_zone_state(z, NUMA_HIT);
2577 __inc_zone_state(z, local_stat);
2579 __inc_zone_state(z, NUMA_MISS);
2580 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2586 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2589 struct page *buffered_rmqueue(struct zone *preferred_zone,
2590 struct zone *zone, unsigned int order,
2591 gfp_t gfp_flags, unsigned int alloc_flags,
2594 unsigned long flags;
2596 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2598 if (likely(order == 0)) {
2599 struct per_cpu_pages *pcp;
2600 struct list_head *list;
2602 local_irq_save(flags);
2604 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2605 list = &pcp->lists[migratetype];
2606 if (list_empty(list)) {
2607 pcp->count += rmqueue_bulk(zone, 0,
2610 if (unlikely(list_empty(list)))
2615 page = list_last_entry(list, struct page, lru);
2617 page = list_first_entry(list, struct page, lru);
2619 __dec_zone_state(zone, NR_ALLOC_BATCH);
2620 list_del(&page->lru);
2623 } while (check_new_pcp(page));
2626 * We most definitely don't want callers attempting to
2627 * allocate greater than order-1 page units with __GFP_NOFAIL.
2629 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2630 spin_lock_irqsave(&zone->lock, flags);
2634 if (alloc_flags & ALLOC_HARDER) {
2635 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2637 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2640 page = __rmqueue(zone, order, migratetype);
2641 } while (page && check_new_pages(page, order));
2642 spin_unlock(&zone->lock);
2645 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2646 __mod_zone_freepage_state(zone, -(1 << order),
2647 get_pcppage_migratetype(page));
2650 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2651 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2652 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2654 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2655 zone_statistics(preferred_zone, zone, gfp_flags);
2656 local_irq_restore(flags);
2658 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2662 local_irq_restore(flags);
2666 #ifdef CONFIG_FAIL_PAGE_ALLOC
2669 struct fault_attr attr;
2671 bool ignore_gfp_highmem;
2672 bool ignore_gfp_reclaim;
2674 } fail_page_alloc = {
2675 .attr = FAULT_ATTR_INITIALIZER,
2676 .ignore_gfp_reclaim = true,
2677 .ignore_gfp_highmem = true,
2681 static int __init setup_fail_page_alloc(char *str)
2683 return setup_fault_attr(&fail_page_alloc.attr, str);
2685 __setup("fail_page_alloc=", setup_fail_page_alloc);
2687 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2689 if (order < fail_page_alloc.min_order)
2691 if (gfp_mask & __GFP_NOFAIL)
2693 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2695 if (fail_page_alloc.ignore_gfp_reclaim &&
2696 (gfp_mask & __GFP_DIRECT_RECLAIM))
2699 return should_fail(&fail_page_alloc.attr, 1 << order);
2702 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2704 static int __init fail_page_alloc_debugfs(void)
2706 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2709 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2710 &fail_page_alloc.attr);
2712 return PTR_ERR(dir);
2714 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2715 &fail_page_alloc.ignore_gfp_reclaim))
2717 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2718 &fail_page_alloc.ignore_gfp_highmem))
2720 if (!debugfs_create_u32("min-order", mode, dir,
2721 &fail_page_alloc.min_order))
2726 debugfs_remove_recursive(dir);
2731 late_initcall(fail_page_alloc_debugfs);
2733 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2735 #else /* CONFIG_FAIL_PAGE_ALLOC */
2737 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2742 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2745 * Return true if free base pages are above 'mark'. For high-order checks it
2746 * will return true of the order-0 watermark is reached and there is at least
2747 * one free page of a suitable size. Checking now avoids taking the zone lock
2748 * to check in the allocation paths if no pages are free.
2750 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2751 int classzone_idx, unsigned int alloc_flags,
2756 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2758 /* free_pages may go negative - that's OK */
2759 free_pages -= (1 << order) - 1;
2761 if (alloc_flags & ALLOC_HIGH)
2765 * If the caller does not have rights to ALLOC_HARDER then subtract
2766 * the high-atomic reserves. This will over-estimate the size of the
2767 * atomic reserve but it avoids a search.
2769 if (likely(!alloc_harder))
2770 free_pages -= z->nr_reserved_highatomic;
2775 /* If allocation can't use CMA areas don't use free CMA pages */
2776 if (!(alloc_flags & ALLOC_CMA))
2777 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2781 * Check watermarks for an order-0 allocation request. If these
2782 * are not met, then a high-order request also cannot go ahead
2783 * even if a suitable page happened to be free.
2785 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2788 /* If this is an order-0 request then the watermark is fine */
2792 /* For a high-order request, check at least one suitable page is free */
2793 for (o = order; o < MAX_ORDER; o++) {
2794 struct free_area *area = &z->free_area[o];
2803 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2804 if (!list_empty(&area->free_list[mt]))
2809 if ((alloc_flags & ALLOC_CMA) &&
2810 !list_empty(&area->free_list[MIGRATE_CMA])) {
2818 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2819 int classzone_idx, unsigned int alloc_flags)
2821 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2822 zone_page_state(z, NR_FREE_PAGES));
2825 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2826 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2828 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2832 /* If allocation can't use CMA areas don't use free CMA pages */
2833 if (!(alloc_flags & ALLOC_CMA))
2834 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2838 * Fast check for order-0 only. If this fails then the reserves
2839 * need to be calculated. There is a corner case where the check
2840 * passes but only the high-order atomic reserve are free. If
2841 * the caller is !atomic then it'll uselessly search the free
2842 * list. That corner case is then slower but it is harmless.
2844 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2847 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2851 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2852 unsigned long mark, int classzone_idx)
2854 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2856 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2857 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2859 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2864 static bool zone_local(struct zone *local_zone, struct zone *zone)
2866 return local_zone->node == zone->node;
2869 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2871 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2874 #else /* CONFIG_NUMA */
2875 static bool zone_local(struct zone *local_zone, struct zone *zone)
2880 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2884 #endif /* CONFIG_NUMA */
2886 static void reset_alloc_batches(struct zone *preferred_zone)
2888 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2891 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2892 high_wmark_pages(zone) - low_wmark_pages(zone) -
2893 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2894 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2895 } while (zone++ != preferred_zone);
2899 * get_page_from_freelist goes through the zonelist trying to allocate
2902 static struct page *
2903 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2904 const struct alloc_context *ac)
2906 struct zoneref *z = ac->preferred_zoneref;
2908 bool fair_skipped = false;
2909 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2913 * Scan zonelist, looking for a zone with enough free.
2914 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2916 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2921 if (cpusets_enabled() &&
2922 (alloc_flags & ALLOC_CPUSET) &&
2923 !__cpuset_zone_allowed(zone, gfp_mask))
2926 * Distribute pages in proportion to the individual
2927 * zone size to ensure fair page aging. The zone a
2928 * page was allocated in should have no effect on the
2929 * time the page has in memory before being reclaimed.
2932 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2933 fair_skipped = true;
2936 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2943 * When allocating a page cache page for writing, we
2944 * want to get it from a zone that is within its dirty
2945 * limit, such that no single zone holds more than its
2946 * proportional share of globally allowed dirty pages.
2947 * The dirty limits take into account the zone's
2948 * lowmem reserves and high watermark so that kswapd
2949 * should be able to balance it without having to
2950 * write pages from its LRU list.
2952 * This may look like it could increase pressure on
2953 * lower zones by failing allocations in higher zones
2954 * before they are full. But the pages that do spill
2955 * over are limited as the lower zones are protected
2956 * by this very same mechanism. It should not become
2957 * a practical burden to them.
2959 * XXX: For now, allow allocations to potentially
2960 * exceed the per-zone dirty limit in the slowpath
2961 * (spread_dirty_pages unset) before going into reclaim,
2962 * which is important when on a NUMA setup the allowed
2963 * zones are together not big enough to reach the
2964 * global limit. The proper fix for these situations
2965 * will require awareness of zones in the
2966 * dirty-throttling and the flusher threads.
2968 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2971 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2972 if (!zone_watermark_fast(zone, order, mark,
2973 ac_classzone_idx(ac), alloc_flags)) {
2976 /* Checked here to keep the fast path fast */
2977 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2978 if (alloc_flags & ALLOC_NO_WATERMARKS)
2981 if (zone_reclaim_mode == 0 ||
2982 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2985 ret = zone_reclaim(zone, gfp_mask, order);
2987 case ZONE_RECLAIM_NOSCAN:
2990 case ZONE_RECLAIM_FULL:
2991 /* scanned but unreclaimable */
2994 /* did we reclaim enough */
2995 if (zone_watermark_ok(zone, order, mark,
2996 ac_classzone_idx(ac), alloc_flags))
3004 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
3005 gfp_mask, alloc_flags, ac->migratetype);
3007 prep_new_page(page, order, gfp_mask, alloc_flags);
3010 * If this is a high-order atomic allocation then check
3011 * if the pageblock should be reserved for the future
3013 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3014 reserve_highatomic_pageblock(page, zone, order);
3021 * The first pass makes sure allocations are spread fairly within the
3022 * local node. However, the local node might have free pages left
3023 * after the fairness batches are exhausted, and remote zones haven't
3024 * even been considered yet. Try once more without fairness, and
3025 * include remote zones now, before entering the slowpath and waking
3026 * kswapd: prefer spilling to a remote zone over swapping locally.
3031 fair_skipped = false;
3032 reset_alloc_batches(ac->preferred_zoneref->zone);
3040 * Large machines with many possible nodes should not always dump per-node
3041 * meminfo in irq context.
3043 static inline bool should_suppress_show_mem(void)
3048 ret = in_interrupt();
3053 static DEFINE_RATELIMIT_STATE(nopage_rs,
3054 DEFAULT_RATELIMIT_INTERVAL,
3055 DEFAULT_RATELIMIT_BURST);
3057 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3059 unsigned int filter = SHOW_MEM_FILTER_NODES;
3061 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3062 debug_guardpage_minorder() > 0)
3066 * This documents exceptions given to allocations in certain
3067 * contexts that are allowed to allocate outside current's set
3070 if (!(gfp_mask & __GFP_NOMEMALLOC))
3071 if (test_thread_flag(TIF_MEMDIE) ||
3072 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3073 filter &= ~SHOW_MEM_FILTER_NODES;
3074 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3075 filter &= ~SHOW_MEM_FILTER_NODES;
3078 struct va_format vaf;
3081 va_start(args, fmt);
3086 pr_warn("%pV", &vaf);
3091 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3092 current->comm, order, gfp_mask, &gfp_mask);
3094 if (!should_suppress_show_mem())
3098 static inline struct page *
3099 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3100 const struct alloc_context *ac, unsigned long *did_some_progress)
3102 struct oom_control oc = {
3103 .zonelist = ac->zonelist,
3104 .nodemask = ac->nodemask,
3105 .gfp_mask = gfp_mask,
3110 *did_some_progress = 0;
3113 * Acquire the oom lock. If that fails, somebody else is
3114 * making progress for us.
3116 if (!mutex_trylock(&oom_lock)) {
3117 *did_some_progress = 1;
3118 schedule_timeout_uninterruptible(1);
3123 * Go through the zonelist yet one more time, keep very high watermark
3124 * here, this is only to catch a parallel oom killing, we must fail if
3125 * we're still under heavy pressure.
3127 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3128 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3132 if (!(gfp_mask & __GFP_NOFAIL)) {
3133 /* Coredumps can quickly deplete all memory reserves */
3134 if (current->flags & PF_DUMPCORE)
3136 /* The OOM killer will not help higher order allocs */
3137 if (order > PAGE_ALLOC_COSTLY_ORDER)
3139 /* The OOM killer does not needlessly kill tasks for lowmem */
3140 if (ac->high_zoneidx < ZONE_NORMAL)
3142 if (pm_suspended_storage())
3145 * XXX: GFP_NOFS allocations should rather fail than rely on
3146 * other request to make a forward progress.
3147 * We are in an unfortunate situation where out_of_memory cannot
3148 * do much for this context but let's try it to at least get
3149 * access to memory reserved if the current task is killed (see
3150 * out_of_memory). Once filesystems are ready to handle allocation
3151 * failures more gracefully we should just bail out here.
3154 /* The OOM killer may not free memory on a specific node */
3155 if (gfp_mask & __GFP_THISNODE)
3158 /* Exhausted what can be done so it's blamo time */
3159 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3160 *did_some_progress = 1;
3162 if (gfp_mask & __GFP_NOFAIL) {
3163 page = get_page_from_freelist(gfp_mask, order,
3164 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3166 * fallback to ignore cpuset restriction if our nodes
3170 page = get_page_from_freelist(gfp_mask, order,
3171 ALLOC_NO_WATERMARKS, ac);
3175 mutex_unlock(&oom_lock);
3181 * Maximum number of compaction retries wit a progress before OOM
3182 * killer is consider as the only way to move forward.
3184 #define MAX_COMPACT_RETRIES 16
3186 #ifdef CONFIG_COMPACTION
3187 /* Try memory compaction for high-order allocations before reclaim */
3188 static struct page *
3189 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3190 unsigned int alloc_flags, const struct alloc_context *ac,
3191 enum migrate_mode mode, enum compact_result *compact_result)
3194 int contended_compaction;
3199 current->flags |= PF_MEMALLOC;
3200 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3201 mode, &contended_compaction);
3202 current->flags &= ~PF_MEMALLOC;
3204 if (*compact_result <= COMPACT_INACTIVE)
3208 * At least in one zone compaction wasn't deferred or skipped, so let's
3209 * count a compaction stall
3211 count_vm_event(COMPACTSTALL);
3213 page = get_page_from_freelist(gfp_mask, order,
3214 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3217 struct zone *zone = page_zone(page);
3219 zone->compact_blockskip_flush = false;
3220 compaction_defer_reset(zone, order, true);
3221 count_vm_event(COMPACTSUCCESS);
3226 * It's bad if compaction run occurs and fails. The most likely reason
3227 * is that pages exist, but not enough to satisfy watermarks.
3229 count_vm_event(COMPACTFAIL);
3232 * In all zones where compaction was attempted (and not
3233 * deferred or skipped), lock contention has been detected.
3234 * For THP allocation we do not want to disrupt the others
3235 * so we fallback to base pages instead.
3237 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3238 *compact_result = COMPACT_CONTENDED;
3241 * If compaction was aborted due to need_resched(), we do not
3242 * want to further increase allocation latency, unless it is
3243 * khugepaged trying to collapse.
3245 if (contended_compaction == COMPACT_CONTENDED_SCHED
3246 && !(current->flags & PF_KTHREAD))
3247 *compact_result = COMPACT_CONTENDED;
3255 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3256 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3257 int compaction_retries)
3259 int max_retries = MAX_COMPACT_RETRIES;
3265 * compaction considers all the zone as desperately out of memory
3266 * so it doesn't really make much sense to retry except when the
3267 * failure could be caused by weak migration mode.
3269 if (compaction_failed(compact_result)) {
3270 if (*migrate_mode == MIGRATE_ASYNC) {
3271 *migrate_mode = MIGRATE_SYNC_LIGHT;
3278 * make sure the compaction wasn't deferred or didn't bail out early
3279 * due to locks contention before we declare that we should give up.
3280 * But do not retry if the given zonelist is not suitable for
3283 if (compaction_withdrawn(compact_result))
3284 return compaction_zonelist_suitable(ac, order, alloc_flags);
3287 * !costly requests are much more important than __GFP_REPEAT
3288 * costly ones because they are de facto nofail and invoke OOM
3289 * killer to move on while costly can fail and users are ready
3290 * to cope with that. 1/4 retries is rather arbitrary but we
3291 * would need much more detailed feedback from compaction to
3292 * make a better decision.
3294 if (order > PAGE_ALLOC_COSTLY_ORDER)
3296 if (compaction_retries <= max_retries)
3302 static inline struct page *
3303 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3304 unsigned int alloc_flags, const struct alloc_context *ac,
3305 enum migrate_mode mode, enum compact_result *compact_result)
3307 *compact_result = COMPACT_SKIPPED;
3312 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3313 enum compact_result compact_result,
3314 enum migrate_mode *migrate_mode,
3315 int compaction_retries)
3320 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3324 * There are setups with compaction disabled which would prefer to loop
3325 * inside the allocator rather than hit the oom killer prematurely.
3326 * Let's give them a good hope and keep retrying while the order-0
3327 * watermarks are OK.
3329 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3331 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3332 ac_classzone_idx(ac), alloc_flags))
3337 #endif /* CONFIG_COMPACTION */
3339 /* Perform direct synchronous page reclaim */
3341 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3342 const struct alloc_context *ac)
3344 struct reclaim_state reclaim_state;
3349 /* We now go into synchronous reclaim */
3350 cpuset_memory_pressure_bump();
3351 current->flags |= PF_MEMALLOC;
3352 lockdep_set_current_reclaim_state(gfp_mask);
3353 reclaim_state.reclaimed_slab = 0;
3354 current->reclaim_state = &reclaim_state;
3356 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3359 current->reclaim_state = NULL;
3360 lockdep_clear_current_reclaim_state();
3361 current->flags &= ~PF_MEMALLOC;
3368 /* The really slow allocator path where we enter direct reclaim */
3369 static inline struct page *
3370 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3371 unsigned int alloc_flags, const struct alloc_context *ac,
3372 unsigned long *did_some_progress)
3374 struct page *page = NULL;
3375 bool drained = false;
3377 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3378 if (unlikely(!(*did_some_progress)))
3382 page = get_page_from_freelist(gfp_mask, order,
3383 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3386 * If an allocation failed after direct reclaim, it could be because
3387 * pages are pinned on the per-cpu lists or in high alloc reserves.
3388 * Shrink them them and try again
3390 if (!page && !drained) {
3391 unreserve_highatomic_pageblock(ac);
3392 drain_all_pages(NULL);
3400 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3405 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3406 ac->high_zoneidx, ac->nodemask)
3407 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3410 static inline unsigned int
3411 gfp_to_alloc_flags(gfp_t gfp_mask)
3413 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3415 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3416 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3419 * The caller may dip into page reserves a bit more if the caller
3420 * cannot run direct reclaim, or if the caller has realtime scheduling
3421 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3422 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3424 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3426 if (gfp_mask & __GFP_ATOMIC) {
3428 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3429 * if it can't schedule.
3431 if (!(gfp_mask & __GFP_NOMEMALLOC))
3432 alloc_flags |= ALLOC_HARDER;
3434 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3435 * comment for __cpuset_node_allowed().
3437 alloc_flags &= ~ALLOC_CPUSET;
3438 } else if (unlikely(rt_task(current)) && !in_interrupt())
3439 alloc_flags |= ALLOC_HARDER;
3441 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3442 if (gfp_mask & __GFP_MEMALLOC)
3443 alloc_flags |= ALLOC_NO_WATERMARKS;
3444 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3445 alloc_flags |= ALLOC_NO_WATERMARKS;
3446 else if (!in_interrupt() &&
3447 ((current->flags & PF_MEMALLOC) ||
3448 unlikely(test_thread_flag(TIF_MEMDIE))))
3449 alloc_flags |= ALLOC_NO_WATERMARKS;
3452 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3453 alloc_flags |= ALLOC_CMA;
3458 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3460 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3463 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3465 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3469 * Maximum number of reclaim retries without any progress before OOM killer
3470 * is consider as the only way to move forward.
3472 #define MAX_RECLAIM_RETRIES 16
3475 * Checks whether it makes sense to retry the reclaim to make a forward progress
3476 * for the given allocation request.
3477 * The reclaim feedback represented by did_some_progress (any progress during
3478 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3479 * any progress in a row) is considered as well as the reclaimable pages on the
3480 * applicable zone list (with a backoff mechanism which is a function of
3481 * no_progress_loops).
3483 * Returns true if a retry is viable or false to enter the oom path.
3486 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3487 struct alloc_context *ac, int alloc_flags,
3488 bool did_some_progress, int no_progress_loops)
3494 * Make sure we converge to OOM if we cannot make any progress
3495 * several times in the row.
3497 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3501 * Keep reclaiming pages while there is a chance this will lead somewhere.
3502 * If none of the target zones can satisfy our allocation request even
3503 * if all reclaimable pages are considered then we are screwed and have
3506 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3508 unsigned long available;
3509 unsigned long reclaimable;
3511 available = reclaimable = zone_reclaimable_pages(zone);
3512 available -= DIV_ROUND_UP(no_progress_loops * available,
3513 MAX_RECLAIM_RETRIES);
3514 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3517 * Would the allocation succeed if we reclaimed the whole
3520 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3521 ac_classzone_idx(ac), alloc_flags, available)) {
3523 * If we didn't make any progress and have a lot of
3524 * dirty + writeback pages then we should wait for
3525 * an IO to complete to slow down the reclaim and
3526 * prevent from pre mature OOM
3528 if (!did_some_progress) {
3529 unsigned long writeback;
3530 unsigned long dirty;
3532 writeback = zone_page_state_snapshot(zone,
3534 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3536 if (2*(writeback + dirty) > reclaimable) {
3537 congestion_wait(BLK_RW_ASYNC, HZ/10);
3543 * Memory allocation/reclaim might be called from a WQ
3544 * context and the current implementation of the WQ
3545 * concurrency control doesn't recognize that
3546 * a particular WQ is congested if the worker thread is
3547 * looping without ever sleeping. Therefore we have to
3548 * do a short sleep here rather than calling
3551 if (current->flags & PF_WQ_WORKER)
3552 schedule_timeout_uninterruptible(1);
3563 static inline struct page *
3564 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3565 struct alloc_context *ac)
3567 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3568 struct page *page = NULL;
3569 unsigned int alloc_flags;
3570 unsigned long did_some_progress;
3571 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3572 enum compact_result compact_result;
3573 int compaction_retries = 0;
3574 int no_progress_loops = 0;
3577 * In the slowpath, we sanity check order to avoid ever trying to
3578 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3579 * be using allocators in order of preference for an area that is
3582 if (order >= MAX_ORDER) {
3583 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3588 * We also sanity check to catch abuse of atomic reserves being used by
3589 * callers that are not in atomic context.
3591 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3592 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3593 gfp_mask &= ~__GFP_ATOMIC;
3596 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3597 wake_all_kswapds(order, ac);
3600 * OK, we're below the kswapd watermark and have kicked background
3601 * reclaim. Now things get more complex, so set up alloc_flags according
3602 * to how we want to proceed.
3604 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3606 /* This is the last chance, in general, before the goto nopage. */
3607 page = get_page_from_freelist(gfp_mask, order,
3608 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3612 /* Allocate without watermarks if the context allows */
3613 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3615 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3616 * the allocation is high priority and these type of
3617 * allocations are system rather than user orientated
3619 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3620 page = get_page_from_freelist(gfp_mask, order,
3621 ALLOC_NO_WATERMARKS, ac);
3626 /* Caller is not willing to reclaim, we can't balance anything */
3627 if (!can_direct_reclaim) {
3629 * All existing users of the __GFP_NOFAIL are blockable, so warn
3630 * of any new users that actually allow this type of allocation
3633 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3637 /* Avoid recursion of direct reclaim */
3638 if (current->flags & PF_MEMALLOC) {
3640 * __GFP_NOFAIL request from this context is rather bizarre
3641 * because we cannot reclaim anything and only can loop waiting
3642 * for somebody to do a work for us.
3644 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3651 /* Avoid allocations with no watermarks from looping endlessly */
3652 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3656 * Try direct compaction. The first pass is asynchronous. Subsequent
3657 * attempts after direct reclaim are synchronous
3659 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3665 /* Checks for THP-specific high-order allocations */
3666 if (is_thp_gfp_mask(gfp_mask)) {
3668 * If compaction is deferred for high-order allocations, it is
3669 * because sync compaction recently failed. If this is the case
3670 * and the caller requested a THP allocation, we do not want
3671 * to heavily disrupt the system, so we fail the allocation
3672 * instead of entering direct reclaim.
3674 if (compact_result == COMPACT_DEFERRED)
3678 * Compaction is contended so rather back off than cause
3681 if(compact_result == COMPACT_CONTENDED)
3685 if (order && compaction_made_progress(compact_result))
3686 compaction_retries++;
3688 /* Try direct reclaim and then allocating */
3689 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3690 &did_some_progress);
3694 /* Do not loop if specifically requested */
3695 if (gfp_mask & __GFP_NORETRY)
3699 * Do not retry costly high order allocations unless they are
3702 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3706 * Costly allocations might have made a progress but this doesn't mean
3707 * their order will become available due to high fragmentation so
3708 * always increment the no progress counter for them
3710 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3711 no_progress_loops = 0;
3713 no_progress_loops++;
3715 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3716 did_some_progress > 0, no_progress_loops))
3720 * It doesn't make any sense to retry for the compaction if the order-0
3721 * reclaim is not able to make any progress because the current
3722 * implementation of the compaction depends on the sufficient amount
3723 * of free memory (see __compaction_suitable)
3725 if (did_some_progress > 0 &&
3726 should_compact_retry(ac, order, alloc_flags,
3727 compact_result, &migration_mode,
3728 compaction_retries))
3731 /* Reclaim has failed us, start killing things */
3732 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3736 /* Retry as long as the OOM killer is making progress */
3737 if (did_some_progress) {
3738 no_progress_loops = 0;
3744 * High-order allocations do not necessarily loop after direct reclaim
3745 * and reclaim/compaction depends on compaction being called after
3746 * reclaim so call directly if necessary.
3747 * It can become very expensive to allocate transparent hugepages at
3748 * fault, so use asynchronous memory compaction for THP unless it is
3749 * khugepaged trying to collapse. All other requests should tolerate
3750 * at least light sync migration.
3752 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3753 migration_mode = MIGRATE_ASYNC;
3755 migration_mode = MIGRATE_SYNC_LIGHT;
3756 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3762 warn_alloc_failed(gfp_mask, order, NULL);
3768 * This is the 'heart' of the zoned buddy allocator.
3771 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3772 struct zonelist *zonelist, nodemask_t *nodemask)
3775 unsigned int cpuset_mems_cookie;
3776 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3777 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3778 struct alloc_context ac = {
3779 .high_zoneidx = gfp_zone(gfp_mask),
3780 .zonelist = zonelist,
3781 .nodemask = nodemask,
3782 .migratetype = gfpflags_to_migratetype(gfp_mask),
3785 if (cpusets_enabled()) {
3786 alloc_mask |= __GFP_HARDWALL;
3787 alloc_flags |= ALLOC_CPUSET;
3789 ac.nodemask = &cpuset_current_mems_allowed;
3792 gfp_mask &= gfp_allowed_mask;
3794 lockdep_trace_alloc(gfp_mask);
3796 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3798 if (should_fail_alloc_page(gfp_mask, order))
3802 * Check the zones suitable for the gfp_mask contain at least one
3803 * valid zone. It's possible to have an empty zonelist as a result
3804 * of __GFP_THISNODE and a memoryless node
3806 if (unlikely(!zonelist->_zonerefs->zone))
3809 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3810 alloc_flags |= ALLOC_CMA;
3813 cpuset_mems_cookie = read_mems_allowed_begin();
3815 /* Dirty zone balancing only done in the fast path */
3816 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3818 /* The preferred zone is used for statistics later */
3819 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3820 ac.high_zoneidx, ac.nodemask);
3821 if (!ac.preferred_zoneref) {
3826 /* First allocation attempt */
3827 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3832 * Runtime PM, block IO and its error handling path can deadlock
3833 * because I/O on the device might not complete.
3835 alloc_mask = memalloc_noio_flags(gfp_mask);
3836 ac.spread_dirty_pages = false;
3839 * Restore the original nodemask if it was potentially replaced with
3840 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3842 if (cpusets_enabled())
3843 ac.nodemask = nodemask;
3844 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3848 * When updating a task's mems_allowed, it is possible to race with
3849 * parallel threads in such a way that an allocation can fail while
3850 * the mask is being updated. If a page allocation is about to fail,
3851 * check if the cpuset changed during allocation and if so, retry.
3853 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3854 alloc_mask = gfp_mask;
3859 if (kmemcheck_enabled && page)
3860 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3862 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3866 EXPORT_SYMBOL(__alloc_pages_nodemask);
3869 * Common helper functions.
3871 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3876 * __get_free_pages() returns a 32-bit address, which cannot represent
3879 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3881 page = alloc_pages(gfp_mask, order);
3884 return (unsigned long) page_address(page);
3886 EXPORT_SYMBOL(__get_free_pages);
3888 unsigned long get_zeroed_page(gfp_t gfp_mask)
3890 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3892 EXPORT_SYMBOL(get_zeroed_page);
3894 void __free_pages(struct page *page, unsigned int order)
3896 if (put_page_testzero(page)) {
3898 free_hot_cold_page(page, false);
3900 __free_pages_ok(page, order);
3904 EXPORT_SYMBOL(__free_pages);
3906 void free_pages(unsigned long addr, unsigned int order)
3909 VM_BUG_ON(!virt_addr_valid((void *)addr));
3910 __free_pages(virt_to_page((void *)addr), order);
3914 EXPORT_SYMBOL(free_pages);
3918 * An arbitrary-length arbitrary-offset area of memory which resides
3919 * within a 0 or higher order page. Multiple fragments within that page
3920 * are individually refcounted, in the page's reference counter.
3922 * The page_frag functions below provide a simple allocation framework for
3923 * page fragments. This is used by the network stack and network device
3924 * drivers to provide a backing region of memory for use as either an
3925 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3927 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3930 struct page *page = NULL;
3931 gfp_t gfp = gfp_mask;
3933 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3934 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3936 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3937 PAGE_FRAG_CACHE_MAX_ORDER);
3938 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3940 if (unlikely(!page))
3941 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3943 nc->va = page ? page_address(page) : NULL;
3948 void *__alloc_page_frag(struct page_frag_cache *nc,
3949 unsigned int fragsz, gfp_t gfp_mask)
3951 unsigned int size = PAGE_SIZE;
3955 if (unlikely(!nc->va)) {
3957 page = __page_frag_refill(nc, gfp_mask);
3961 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3962 /* if size can vary use size else just use PAGE_SIZE */
3965 /* Even if we own the page, we do not use atomic_set().
3966 * This would break get_page_unless_zero() users.
3968 page_ref_add(page, size - 1);
3970 /* reset page count bias and offset to start of new frag */
3971 nc->pfmemalloc = page_is_pfmemalloc(page);
3972 nc->pagecnt_bias = size;
3976 offset = nc->offset - fragsz;
3977 if (unlikely(offset < 0)) {
3978 page = virt_to_page(nc->va);
3980 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3983 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3984 /* if size can vary use size else just use PAGE_SIZE */
3987 /* OK, page count is 0, we can safely set it */
3988 set_page_count(page, size);
3990 /* reset page count bias and offset to start of new frag */
3991 nc->pagecnt_bias = size;
3992 offset = size - fragsz;
3996 nc->offset = offset;
3998 return nc->va + offset;
4000 EXPORT_SYMBOL(__alloc_page_frag);
4003 * Frees a page fragment allocated out of either a compound or order 0 page.
4005 void __free_page_frag(void *addr)
4007 struct page *page = virt_to_head_page(addr);
4009 if (unlikely(put_page_testzero(page)))
4010 __free_pages_ok(page, compound_order(page));
4012 EXPORT_SYMBOL(__free_page_frag);
4015 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
4016 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
4017 * equivalent to alloc_pages.
4019 * It should be used when the caller would like to use kmalloc, but since the
4020 * allocation is large, it has to fall back to the page allocator.
4022 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
4026 page = alloc_pages(gfp_mask, order);
4027 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4028 __free_pages(page, order);
4034 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
4038 page = alloc_pages_node(nid, gfp_mask, order);
4039 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4040 __free_pages(page, order);
4047 * __free_kmem_pages and free_kmem_pages will free pages allocated with
4050 void __free_kmem_pages(struct page *page, unsigned int order)
4052 memcg_kmem_uncharge(page, order);
4053 __free_pages(page, order);
4056 void free_kmem_pages(unsigned long addr, unsigned int order)
4059 VM_BUG_ON(!virt_addr_valid((void *)addr));
4060 __free_kmem_pages(virt_to_page((void *)addr), order);
4064 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4068 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4069 unsigned long used = addr + PAGE_ALIGN(size);
4071 split_page(virt_to_page((void *)addr), order);
4072 while (used < alloc_end) {
4077 return (void *)addr;
4081 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4082 * @size: the number of bytes to allocate
4083 * @gfp_mask: GFP flags for the allocation
4085 * This function is similar to alloc_pages(), except that it allocates the
4086 * minimum number of pages to satisfy the request. alloc_pages() can only
4087 * allocate memory in power-of-two pages.
4089 * This function is also limited by MAX_ORDER.
4091 * Memory allocated by this function must be released by free_pages_exact().
4093 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4095 unsigned int order = get_order(size);
4098 addr = __get_free_pages(gfp_mask, order);
4099 return make_alloc_exact(addr, order, size);
4101 EXPORT_SYMBOL(alloc_pages_exact);
4104 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4106 * @nid: the preferred node ID where memory should be allocated
4107 * @size: the number of bytes to allocate
4108 * @gfp_mask: GFP flags for the allocation
4110 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4113 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4115 unsigned int order = get_order(size);
4116 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4119 return make_alloc_exact((unsigned long)page_address(p), order, size);
4123 * free_pages_exact - release memory allocated via alloc_pages_exact()
4124 * @virt: the value returned by alloc_pages_exact.
4125 * @size: size of allocation, same value as passed to alloc_pages_exact().
4127 * Release the memory allocated by a previous call to alloc_pages_exact.
4129 void free_pages_exact(void *virt, size_t size)
4131 unsigned long addr = (unsigned long)virt;
4132 unsigned long end = addr + PAGE_ALIGN(size);
4134 while (addr < end) {
4139 EXPORT_SYMBOL(free_pages_exact);
4142 * nr_free_zone_pages - count number of pages beyond high watermark
4143 * @offset: The zone index of the highest zone
4145 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4146 * high watermark within all zones at or below a given zone index. For each
4147 * zone, the number of pages is calculated as:
4148 * managed_pages - high_pages
4150 static unsigned long nr_free_zone_pages(int offset)
4155 /* Just pick one node, since fallback list is circular */
4156 unsigned long sum = 0;
4158 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4160 for_each_zone_zonelist(zone, z, zonelist, offset) {
4161 unsigned long size = zone->managed_pages;
4162 unsigned long high = high_wmark_pages(zone);
4171 * nr_free_buffer_pages - count number of pages beyond high watermark
4173 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4174 * watermark within ZONE_DMA and ZONE_NORMAL.
4176 unsigned long nr_free_buffer_pages(void)
4178 return nr_free_zone_pages(gfp_zone(GFP_USER));
4180 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4183 * nr_free_pagecache_pages - count number of pages beyond high watermark
4185 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4186 * high watermark within all zones.
4188 unsigned long nr_free_pagecache_pages(void)
4190 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4193 static inline void show_node(struct zone *zone)
4195 if (IS_ENABLED(CONFIG_NUMA))
4196 printk("Node %d ", zone_to_nid(zone));
4199 long si_mem_available(void)
4202 unsigned long pagecache;
4203 unsigned long wmark_low = 0;
4204 unsigned long pages[NR_LRU_LISTS];
4208 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4209 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4212 wmark_low += zone->watermark[WMARK_LOW];
4215 * Estimate the amount of memory available for userspace allocations,
4216 * without causing swapping.
4218 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4221 * Not all the page cache can be freed, otherwise the system will
4222 * start swapping. Assume at least half of the page cache, or the
4223 * low watermark worth of cache, needs to stay.
4225 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4226 pagecache -= min(pagecache / 2, wmark_low);
4227 available += pagecache;
4230 * Part of the reclaimable slab consists of items that are in use,
4231 * and cannot be freed. Cap this estimate at the low watermark.
4233 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4234 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4240 EXPORT_SYMBOL_GPL(si_mem_available);
4242 void si_meminfo(struct sysinfo *val)
4244 val->totalram = totalram_pages;
4245 val->sharedram = global_page_state(NR_SHMEM);
4246 val->freeram = global_page_state(NR_FREE_PAGES);
4247 val->bufferram = nr_blockdev_pages();
4248 val->totalhigh = totalhigh_pages;
4249 val->freehigh = nr_free_highpages();
4250 val->mem_unit = PAGE_SIZE;
4253 EXPORT_SYMBOL(si_meminfo);
4256 void si_meminfo_node(struct sysinfo *val, int nid)
4258 int zone_type; /* needs to be signed */
4259 unsigned long managed_pages = 0;
4260 unsigned long managed_highpages = 0;
4261 unsigned long free_highpages = 0;
4262 pg_data_t *pgdat = NODE_DATA(nid);
4264 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4265 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4266 val->totalram = managed_pages;
4267 val->sharedram = node_page_state(nid, NR_SHMEM);
4268 val->freeram = node_page_state(nid, NR_FREE_PAGES);
4269 #ifdef CONFIG_HIGHMEM
4270 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4271 struct zone *zone = &pgdat->node_zones[zone_type];
4273 if (is_highmem(zone)) {
4274 managed_highpages += zone->managed_pages;
4275 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4278 val->totalhigh = managed_highpages;
4279 val->freehigh = free_highpages;
4281 val->totalhigh = managed_highpages;
4282 val->freehigh = free_highpages;
4284 val->mem_unit = PAGE_SIZE;
4289 * Determine whether the node should be displayed or not, depending on whether
4290 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4292 bool skip_free_areas_node(unsigned int flags, int nid)
4295 unsigned int cpuset_mems_cookie;
4297 if (!(flags & SHOW_MEM_FILTER_NODES))
4301 cpuset_mems_cookie = read_mems_allowed_begin();
4302 ret = !node_isset(nid, cpuset_current_mems_allowed);
4303 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4308 #define K(x) ((x) << (PAGE_SHIFT-10))
4310 static void show_migration_types(unsigned char type)
4312 static const char types[MIGRATE_TYPES] = {
4313 [MIGRATE_UNMOVABLE] = 'U',
4314 [MIGRATE_MOVABLE] = 'M',
4315 [MIGRATE_RECLAIMABLE] = 'E',
4316 [MIGRATE_HIGHATOMIC] = 'H',
4318 [MIGRATE_CMA] = 'C',
4320 #ifdef CONFIG_MEMORY_ISOLATION
4321 [MIGRATE_ISOLATE] = 'I',
4324 char tmp[MIGRATE_TYPES + 1];
4328 for (i = 0; i < MIGRATE_TYPES; i++) {
4329 if (type & (1 << i))
4334 printk("(%s) ", tmp);
4338 * Show free area list (used inside shift_scroll-lock stuff)
4339 * We also calculate the percentage fragmentation. We do this by counting the
4340 * memory on each free list with the exception of the first item on the list.
4343 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4346 void show_free_areas(unsigned int filter)
4348 unsigned long free_pcp = 0;
4352 for_each_populated_zone(zone) {
4353 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4356 for_each_online_cpu(cpu)
4357 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4360 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4361 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4362 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4363 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4364 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4365 " free:%lu free_pcp:%lu free_cma:%lu\n",
4366 global_page_state(NR_ACTIVE_ANON),
4367 global_page_state(NR_INACTIVE_ANON),
4368 global_page_state(NR_ISOLATED_ANON),
4369 global_page_state(NR_ACTIVE_FILE),
4370 global_page_state(NR_INACTIVE_FILE),
4371 global_page_state(NR_ISOLATED_FILE),
4372 global_page_state(NR_UNEVICTABLE),
4373 global_page_state(NR_FILE_DIRTY),
4374 global_page_state(NR_WRITEBACK),
4375 global_page_state(NR_UNSTABLE_NFS),
4376 global_page_state(NR_SLAB_RECLAIMABLE),
4377 global_page_state(NR_SLAB_UNRECLAIMABLE),
4378 global_page_state(NR_FILE_MAPPED),
4379 global_page_state(NR_SHMEM),
4380 global_page_state(NR_PAGETABLE),
4381 global_page_state(NR_BOUNCE),
4382 global_page_state(NR_FREE_PAGES),
4384 global_page_state(NR_FREE_CMA_PAGES));
4386 for_each_populated_zone(zone) {
4389 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4393 for_each_online_cpu(cpu)
4394 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4402 " active_anon:%lukB"
4403 " inactive_anon:%lukB"
4404 " active_file:%lukB"
4405 " inactive_file:%lukB"
4406 " unevictable:%lukB"
4407 " isolated(anon):%lukB"
4408 " isolated(file):%lukB"
4416 " slab_reclaimable:%lukB"
4417 " slab_unreclaimable:%lukB"
4418 " kernel_stack:%lukB"
4425 " writeback_tmp:%lukB"
4426 " pages_scanned:%lu"
4427 " all_unreclaimable? %s"
4430 K(zone_page_state(zone, NR_FREE_PAGES)),
4431 K(min_wmark_pages(zone)),
4432 K(low_wmark_pages(zone)),
4433 K(high_wmark_pages(zone)),
4434 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4435 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4436 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4437 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4438 K(zone_page_state(zone, NR_UNEVICTABLE)),
4439 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4440 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4441 K(zone->present_pages),
4442 K(zone->managed_pages),
4443 K(zone_page_state(zone, NR_MLOCK)),
4444 K(zone_page_state(zone, NR_FILE_DIRTY)),
4445 K(zone_page_state(zone, NR_WRITEBACK)),
4446 K(zone_page_state(zone, NR_FILE_MAPPED)),
4447 K(zone_page_state(zone, NR_SHMEM)),
4448 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4449 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4450 zone_page_state(zone, NR_KERNEL_STACK) *
4452 K(zone_page_state(zone, NR_PAGETABLE)),
4453 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4454 K(zone_page_state(zone, NR_BOUNCE)),
4456 K(this_cpu_read(zone->pageset->pcp.count)),
4457 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4458 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4459 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4460 (!zone_reclaimable(zone) ? "yes" : "no")
4462 printk("lowmem_reserve[]:");
4463 for (i = 0; i < MAX_NR_ZONES; i++)
4464 printk(" %ld", zone->lowmem_reserve[i]);
4468 for_each_populated_zone(zone) {
4470 unsigned long nr[MAX_ORDER], flags, total = 0;
4471 unsigned char types[MAX_ORDER];
4473 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4476 printk("%s: ", zone->name);
4478 spin_lock_irqsave(&zone->lock, flags);
4479 for (order = 0; order < MAX_ORDER; order++) {
4480 struct free_area *area = &zone->free_area[order];
4483 nr[order] = area->nr_free;
4484 total += nr[order] << order;
4487 for (type = 0; type < MIGRATE_TYPES; type++) {
4488 if (!list_empty(&area->free_list[type]))
4489 types[order] |= 1 << type;
4492 spin_unlock_irqrestore(&zone->lock, flags);
4493 for (order = 0; order < MAX_ORDER; order++) {
4494 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4496 show_migration_types(types[order]);
4498 printk("= %lukB\n", K(total));
4501 hugetlb_show_meminfo();
4503 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4505 show_swap_cache_info();
4508 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4510 zoneref->zone = zone;
4511 zoneref->zone_idx = zone_idx(zone);
4515 * Builds allocation fallback zone lists.
4517 * Add all populated zones of a node to the zonelist.
4519 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4523 enum zone_type zone_type = MAX_NR_ZONES;
4527 zone = pgdat->node_zones + zone_type;
4528 if (populated_zone(zone)) {
4529 zoneref_set_zone(zone,
4530 &zonelist->_zonerefs[nr_zones++]);
4531 check_highest_zone(zone_type);
4533 } while (zone_type);
4541 * 0 = automatic detection of better ordering.
4542 * 1 = order by ([node] distance, -zonetype)
4543 * 2 = order by (-zonetype, [node] distance)
4545 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4546 * the same zonelist. So only NUMA can configure this param.
4548 #define ZONELIST_ORDER_DEFAULT 0
4549 #define ZONELIST_ORDER_NODE 1
4550 #define ZONELIST_ORDER_ZONE 2
4552 /* zonelist order in the kernel.
4553 * set_zonelist_order() will set this to NODE or ZONE.
4555 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4556 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4560 /* The value user specified ....changed by config */
4561 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4562 /* string for sysctl */
4563 #define NUMA_ZONELIST_ORDER_LEN 16
4564 char numa_zonelist_order[16] = "default";
4567 * interface for configure zonelist ordering.
4568 * command line option "numa_zonelist_order"
4569 * = "[dD]efault - default, automatic configuration.
4570 * = "[nN]ode - order by node locality, then by zone within node
4571 * = "[zZ]one - order by zone, then by locality within zone
4574 static int __parse_numa_zonelist_order(char *s)
4576 if (*s == 'd' || *s == 'D') {
4577 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4578 } else if (*s == 'n' || *s == 'N') {
4579 user_zonelist_order = ZONELIST_ORDER_NODE;
4580 } else if (*s == 'z' || *s == 'Z') {
4581 user_zonelist_order = ZONELIST_ORDER_ZONE;
4583 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4589 static __init int setup_numa_zonelist_order(char *s)
4596 ret = __parse_numa_zonelist_order(s);
4598 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4602 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4605 * sysctl handler for numa_zonelist_order
4607 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4608 void __user *buffer, size_t *length,
4611 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4613 static DEFINE_MUTEX(zl_order_mutex);
4615 mutex_lock(&zl_order_mutex);
4617 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4621 strcpy(saved_string, (char *)table->data);
4623 ret = proc_dostring(table, write, buffer, length, ppos);
4627 int oldval = user_zonelist_order;
4629 ret = __parse_numa_zonelist_order((char *)table->data);
4632 * bogus value. restore saved string
4634 strncpy((char *)table->data, saved_string,
4635 NUMA_ZONELIST_ORDER_LEN);
4636 user_zonelist_order = oldval;
4637 } else if (oldval != user_zonelist_order) {
4638 mutex_lock(&zonelists_mutex);
4639 build_all_zonelists(NULL, NULL);
4640 mutex_unlock(&zonelists_mutex);
4644 mutex_unlock(&zl_order_mutex);
4649 #define MAX_NODE_LOAD (nr_online_nodes)
4650 static int node_load[MAX_NUMNODES];
4653 * find_next_best_node - find the next node that should appear in a given node's fallback list
4654 * @node: node whose fallback list we're appending
4655 * @used_node_mask: nodemask_t of already used nodes
4657 * We use a number of factors to determine which is the next node that should
4658 * appear on a given node's fallback list. The node should not have appeared
4659 * already in @node's fallback list, and it should be the next closest node
4660 * according to the distance array (which contains arbitrary distance values
4661 * from each node to each node in the system), and should also prefer nodes
4662 * with no CPUs, since presumably they'll have very little allocation pressure
4663 * on them otherwise.
4664 * It returns -1 if no node is found.
4666 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4669 int min_val = INT_MAX;
4670 int best_node = NUMA_NO_NODE;
4671 const struct cpumask *tmp = cpumask_of_node(0);
4673 /* Use the local node if we haven't already */
4674 if (!node_isset(node, *used_node_mask)) {
4675 node_set(node, *used_node_mask);
4679 for_each_node_state(n, N_MEMORY) {
4681 /* Don't want a node to appear more than once */
4682 if (node_isset(n, *used_node_mask))
4685 /* Use the distance array to find the distance */
4686 val = node_distance(node, n);
4688 /* Penalize nodes under us ("prefer the next node") */
4691 /* Give preference to headless and unused nodes */
4692 tmp = cpumask_of_node(n);
4693 if (!cpumask_empty(tmp))
4694 val += PENALTY_FOR_NODE_WITH_CPUS;
4696 /* Slight preference for less loaded node */
4697 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4698 val += node_load[n];
4700 if (val < min_val) {
4707 node_set(best_node, *used_node_mask);
4714 * Build zonelists ordered by node and zones within node.
4715 * This results in maximum locality--normal zone overflows into local
4716 * DMA zone, if any--but risks exhausting DMA zone.
4718 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4721 struct zonelist *zonelist;
4723 zonelist = &pgdat->node_zonelists[0];
4724 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4726 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4727 zonelist->_zonerefs[j].zone = NULL;
4728 zonelist->_zonerefs[j].zone_idx = 0;
4732 * Build gfp_thisnode zonelists
4734 static void build_thisnode_zonelists(pg_data_t *pgdat)
4737 struct zonelist *zonelist;
4739 zonelist = &pgdat->node_zonelists[1];
4740 j = build_zonelists_node(pgdat, zonelist, 0);
4741 zonelist->_zonerefs[j].zone = NULL;
4742 zonelist->_zonerefs[j].zone_idx = 0;
4746 * Build zonelists ordered by zone and nodes within zones.
4747 * This results in conserving DMA zone[s] until all Normal memory is
4748 * exhausted, but results in overflowing to remote node while memory
4749 * may still exist in local DMA zone.
4751 static int node_order[MAX_NUMNODES];
4753 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4756 int zone_type; /* needs to be signed */
4758 struct zonelist *zonelist;
4760 zonelist = &pgdat->node_zonelists[0];
4762 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4763 for (j = 0; j < nr_nodes; j++) {
4764 node = node_order[j];
4765 z = &NODE_DATA(node)->node_zones[zone_type];
4766 if (populated_zone(z)) {
4768 &zonelist->_zonerefs[pos++]);
4769 check_highest_zone(zone_type);
4773 zonelist->_zonerefs[pos].zone = NULL;
4774 zonelist->_zonerefs[pos].zone_idx = 0;
4777 #if defined(CONFIG_64BIT)
4779 * Devices that require DMA32/DMA are relatively rare and do not justify a
4780 * penalty to every machine in case the specialised case applies. Default
4781 * to Node-ordering on 64-bit NUMA machines
4783 static int default_zonelist_order(void)
4785 return ZONELIST_ORDER_NODE;
4789 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4790 * by the kernel. If processes running on node 0 deplete the low memory zone
4791 * then reclaim will occur more frequency increasing stalls and potentially
4792 * be easier to OOM if a large percentage of the zone is under writeback or
4793 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4794 * Hence, default to zone ordering on 32-bit.
4796 static int default_zonelist_order(void)
4798 return ZONELIST_ORDER_ZONE;
4800 #endif /* CONFIG_64BIT */
4802 static void set_zonelist_order(void)
4804 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4805 current_zonelist_order = default_zonelist_order();
4807 current_zonelist_order = user_zonelist_order;
4810 static void build_zonelists(pg_data_t *pgdat)
4813 nodemask_t used_mask;
4814 int local_node, prev_node;
4815 struct zonelist *zonelist;
4816 unsigned int order = current_zonelist_order;
4818 /* initialize zonelists */
4819 for (i = 0; i < MAX_ZONELISTS; i++) {
4820 zonelist = pgdat->node_zonelists + i;
4821 zonelist->_zonerefs[0].zone = NULL;
4822 zonelist->_zonerefs[0].zone_idx = 0;
4825 /* NUMA-aware ordering of nodes */
4826 local_node = pgdat->node_id;
4827 load = nr_online_nodes;
4828 prev_node = local_node;
4829 nodes_clear(used_mask);
4831 memset(node_order, 0, sizeof(node_order));
4834 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4836 * We don't want to pressure a particular node.
4837 * So adding penalty to the first node in same
4838 * distance group to make it round-robin.
4840 if (node_distance(local_node, node) !=
4841 node_distance(local_node, prev_node))
4842 node_load[node] = load;
4846 if (order == ZONELIST_ORDER_NODE)
4847 build_zonelists_in_node_order(pgdat, node);
4849 node_order[i++] = node; /* remember order */
4852 if (order == ZONELIST_ORDER_ZONE) {
4853 /* calculate node order -- i.e., DMA last! */
4854 build_zonelists_in_zone_order(pgdat, i);
4857 build_thisnode_zonelists(pgdat);
4860 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4862 * Return node id of node used for "local" allocations.
4863 * I.e., first node id of first zone in arg node's generic zonelist.
4864 * Used for initializing percpu 'numa_mem', which is used primarily
4865 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4867 int local_memory_node(int node)
4871 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4872 gfp_zone(GFP_KERNEL),
4874 return z->zone->node;
4878 #else /* CONFIG_NUMA */
4880 static void set_zonelist_order(void)
4882 current_zonelist_order = ZONELIST_ORDER_ZONE;
4885 static void build_zonelists(pg_data_t *pgdat)
4887 int node, local_node;
4889 struct zonelist *zonelist;
4891 local_node = pgdat->node_id;
4893 zonelist = &pgdat->node_zonelists[0];
4894 j = build_zonelists_node(pgdat, zonelist, 0);
4897 * Now we build the zonelist so that it contains the zones
4898 * of all the other nodes.
4899 * We don't want to pressure a particular node, so when
4900 * building the zones for node N, we make sure that the
4901 * zones coming right after the local ones are those from
4902 * node N+1 (modulo N)
4904 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4905 if (!node_online(node))
4907 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4909 for (node = 0; node < local_node; node++) {
4910 if (!node_online(node))
4912 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4915 zonelist->_zonerefs[j].zone = NULL;
4916 zonelist->_zonerefs[j].zone_idx = 0;
4919 #endif /* CONFIG_NUMA */
4922 * Boot pageset table. One per cpu which is going to be used for all
4923 * zones and all nodes. The parameters will be set in such a way
4924 * that an item put on a list will immediately be handed over to
4925 * the buddy list. This is safe since pageset manipulation is done
4926 * with interrupts disabled.
4928 * The boot_pagesets must be kept even after bootup is complete for
4929 * unused processors and/or zones. They do play a role for bootstrapping
4930 * hotplugged processors.
4932 * zoneinfo_show() and maybe other functions do
4933 * not check if the processor is online before following the pageset pointer.
4934 * Other parts of the kernel may not check if the zone is available.
4936 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4937 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4938 static void setup_zone_pageset(struct zone *zone);
4941 * Global mutex to protect against size modification of zonelists
4942 * as well as to serialize pageset setup for the new populated zone.
4944 DEFINE_MUTEX(zonelists_mutex);
4946 /* return values int ....just for stop_machine() */
4947 static int __build_all_zonelists(void *data)
4951 pg_data_t *self = data;
4954 memset(node_load, 0, sizeof(node_load));
4957 if (self && !node_online(self->node_id)) {
4958 build_zonelists(self);
4961 for_each_online_node(nid) {
4962 pg_data_t *pgdat = NODE_DATA(nid);
4964 build_zonelists(pgdat);
4968 * Initialize the boot_pagesets that are going to be used
4969 * for bootstrapping processors. The real pagesets for
4970 * each zone will be allocated later when the per cpu
4971 * allocator is available.
4973 * boot_pagesets are used also for bootstrapping offline
4974 * cpus if the system is already booted because the pagesets
4975 * are needed to initialize allocators on a specific cpu too.
4976 * F.e. the percpu allocator needs the page allocator which
4977 * needs the percpu allocator in order to allocate its pagesets
4978 * (a chicken-egg dilemma).
4980 for_each_possible_cpu(cpu) {
4981 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4983 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4985 * We now know the "local memory node" for each node--
4986 * i.e., the node of the first zone in the generic zonelist.
4987 * Set up numa_mem percpu variable for on-line cpus. During
4988 * boot, only the boot cpu should be on-line; we'll init the
4989 * secondary cpus' numa_mem as they come on-line. During
4990 * node/memory hotplug, we'll fixup all on-line cpus.
4992 if (cpu_online(cpu))
4993 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5000 static noinline void __init
5001 build_all_zonelists_init(void)
5003 __build_all_zonelists(NULL);
5004 mminit_verify_zonelist();
5005 cpuset_init_current_mems_allowed();
5009 * Called with zonelists_mutex held always
5010 * unless system_state == SYSTEM_BOOTING.
5012 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5013 * [we're only called with non-NULL zone through __meminit paths] and
5014 * (2) call of __init annotated helper build_all_zonelists_init
5015 * [protected by SYSTEM_BOOTING].
5017 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5019 set_zonelist_order();
5021 if (system_state == SYSTEM_BOOTING) {
5022 build_all_zonelists_init();
5024 #ifdef CONFIG_MEMORY_HOTPLUG
5026 setup_zone_pageset(zone);
5028 /* we have to stop all cpus to guarantee there is no user
5030 stop_machine(__build_all_zonelists, pgdat, NULL);
5031 /* cpuset refresh routine should be here */
5033 vm_total_pages = nr_free_pagecache_pages();
5035 * Disable grouping by mobility if the number of pages in the
5036 * system is too low to allow the mechanism to work. It would be
5037 * more accurate, but expensive to check per-zone. This check is
5038 * made on memory-hotadd so a system can start with mobility
5039 * disabled and enable it later
5041 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5042 page_group_by_mobility_disabled = 1;
5044 page_group_by_mobility_disabled = 0;
5046 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5048 zonelist_order_name[current_zonelist_order],
5049 page_group_by_mobility_disabled ? "off" : "on",
5052 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5057 * Helper functions to size the waitqueue hash table.
5058 * Essentially these want to choose hash table sizes sufficiently
5059 * large so that collisions trying to wait on pages are rare.
5060 * But in fact, the number of active page waitqueues on typical
5061 * systems is ridiculously low, less than 200. So this is even
5062 * conservative, even though it seems large.
5064 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5065 * waitqueues, i.e. the size of the waitq table given the number of pages.
5067 #define PAGES_PER_WAITQUEUE 256
5069 #ifndef CONFIG_MEMORY_HOTPLUG
5070 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5072 unsigned long size = 1;
5074 pages /= PAGES_PER_WAITQUEUE;
5076 while (size < pages)
5080 * Once we have dozens or even hundreds of threads sleeping
5081 * on IO we've got bigger problems than wait queue collision.
5082 * Limit the size of the wait table to a reasonable size.
5084 size = min(size, 4096UL);
5086 return max(size, 4UL);
5090 * A zone's size might be changed by hot-add, so it is not possible to determine
5091 * a suitable size for its wait_table. So we use the maximum size now.
5093 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5095 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5096 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5097 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5099 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5100 * or more by the traditional way. (See above). It equals:
5102 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5103 * ia64(16K page size) : = ( 8G + 4M)byte.
5104 * powerpc (64K page size) : = (32G +16M)byte.
5106 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5113 * This is an integer logarithm so that shifts can be used later
5114 * to extract the more random high bits from the multiplicative
5115 * hash function before the remainder is taken.
5117 static inline unsigned long wait_table_bits(unsigned long size)
5123 * Initially all pages are reserved - free ones are freed
5124 * up by free_all_bootmem() once the early boot process is
5125 * done. Non-atomic initialization, single-pass.
5127 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5128 unsigned long start_pfn, enum memmap_context context)
5130 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5131 unsigned long end_pfn = start_pfn + size;
5132 pg_data_t *pgdat = NODE_DATA(nid);
5134 unsigned long nr_initialised = 0;
5135 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5136 struct memblock_region *r = NULL, *tmp;
5139 if (highest_memmap_pfn < end_pfn - 1)
5140 highest_memmap_pfn = end_pfn - 1;
5143 * Honor reservation requested by the driver for this ZONE_DEVICE
5146 if (altmap && start_pfn == altmap->base_pfn)
5147 start_pfn += altmap->reserve;
5149 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5151 * There can be holes in boot-time mem_map[]s handed to this
5152 * function. They do not exist on hotplugged memory.
5154 if (context != MEMMAP_EARLY)
5157 if (!early_pfn_valid(pfn))
5159 if (!early_pfn_in_nid(pfn, nid))
5161 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5164 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5166 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5167 * from zone_movable_pfn[nid] to end of each node should be
5168 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5170 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5171 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5175 * Check given memblock attribute by firmware which can affect
5176 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5177 * mirrored, it's an overlapped memmap init. skip it.
5179 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5180 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5181 for_each_memblock(memory, tmp)
5182 if (pfn < memblock_region_memory_end_pfn(tmp))
5186 if (pfn >= memblock_region_memory_base_pfn(r) &&
5187 memblock_is_mirror(r)) {
5188 /* already initialized as NORMAL */
5189 pfn = memblock_region_memory_end_pfn(r);
5197 * Mark the block movable so that blocks are reserved for
5198 * movable at startup. This will force kernel allocations
5199 * to reserve their blocks rather than leaking throughout
5200 * the address space during boot when many long-lived
5201 * kernel allocations are made.
5203 * bitmap is created for zone's valid pfn range. but memmap
5204 * can be created for invalid pages (for alignment)
5205 * check here not to call set_pageblock_migratetype() against
5208 if (!(pfn & (pageblock_nr_pages - 1))) {
5209 struct page *page = pfn_to_page(pfn);
5211 __init_single_page(page, pfn, zone, nid);
5212 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5214 __init_single_pfn(pfn, zone, nid);
5219 static void __meminit zone_init_free_lists(struct zone *zone)
5221 unsigned int order, t;
5222 for_each_migratetype_order(order, t) {
5223 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5224 zone->free_area[order].nr_free = 0;
5228 #ifndef __HAVE_ARCH_MEMMAP_INIT
5229 #define memmap_init(size, nid, zone, start_pfn) \
5230 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5233 static int zone_batchsize(struct zone *zone)
5239 * The per-cpu-pages pools are set to around 1000th of the
5240 * size of the zone. But no more than 1/2 of a meg.
5242 * OK, so we don't know how big the cache is. So guess.
5244 batch = zone->managed_pages / 1024;
5245 if (batch * PAGE_SIZE > 512 * 1024)
5246 batch = (512 * 1024) / PAGE_SIZE;
5247 batch /= 4; /* We effectively *= 4 below */
5252 * Clamp the batch to a 2^n - 1 value. Having a power
5253 * of 2 value was found to be more likely to have
5254 * suboptimal cache aliasing properties in some cases.
5256 * For example if 2 tasks are alternately allocating
5257 * batches of pages, one task can end up with a lot
5258 * of pages of one half of the possible page colors
5259 * and the other with pages of the other colors.
5261 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5266 /* The deferral and batching of frees should be suppressed under NOMMU
5269 * The problem is that NOMMU needs to be able to allocate large chunks
5270 * of contiguous memory as there's no hardware page translation to
5271 * assemble apparent contiguous memory from discontiguous pages.
5273 * Queueing large contiguous runs of pages for batching, however,
5274 * causes the pages to actually be freed in smaller chunks. As there
5275 * can be a significant delay between the individual batches being
5276 * recycled, this leads to the once large chunks of space being
5277 * fragmented and becoming unavailable for high-order allocations.
5284 * pcp->high and pcp->batch values are related and dependent on one another:
5285 * ->batch must never be higher then ->high.
5286 * The following function updates them in a safe manner without read side
5289 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5290 * those fields changing asynchronously (acording the the above rule).
5292 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5293 * outside of boot time (or some other assurance that no concurrent updaters
5296 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5297 unsigned long batch)
5299 /* start with a fail safe value for batch */
5303 /* Update high, then batch, in order */
5310 /* a companion to pageset_set_high() */
5311 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5313 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5316 static void pageset_init(struct per_cpu_pageset *p)
5318 struct per_cpu_pages *pcp;
5321 memset(p, 0, sizeof(*p));
5325 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5326 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5329 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5332 pageset_set_batch(p, batch);
5336 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5337 * to the value high for the pageset p.
5339 static void pageset_set_high(struct per_cpu_pageset *p,
5342 unsigned long batch = max(1UL, high / 4);
5343 if ((high / 4) > (PAGE_SHIFT * 8))
5344 batch = PAGE_SHIFT * 8;
5346 pageset_update(&p->pcp, high, batch);
5349 static void pageset_set_high_and_batch(struct zone *zone,
5350 struct per_cpu_pageset *pcp)
5352 if (percpu_pagelist_fraction)
5353 pageset_set_high(pcp,
5354 (zone->managed_pages /
5355 percpu_pagelist_fraction));
5357 pageset_set_batch(pcp, zone_batchsize(zone));
5360 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5362 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5365 pageset_set_high_and_batch(zone, pcp);
5368 static void __meminit setup_zone_pageset(struct zone *zone)
5371 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5372 for_each_possible_cpu(cpu)
5373 zone_pageset_init(zone, cpu);
5377 * Allocate per cpu pagesets and initialize them.
5378 * Before this call only boot pagesets were available.
5380 void __init setup_per_cpu_pageset(void)
5384 for_each_populated_zone(zone)
5385 setup_zone_pageset(zone);
5388 static noinline __init_refok
5389 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5395 * The per-page waitqueue mechanism uses hashed waitqueues
5398 zone->wait_table_hash_nr_entries =
5399 wait_table_hash_nr_entries(zone_size_pages);
5400 zone->wait_table_bits =
5401 wait_table_bits(zone->wait_table_hash_nr_entries);
5402 alloc_size = zone->wait_table_hash_nr_entries
5403 * sizeof(wait_queue_head_t);
5405 if (!slab_is_available()) {
5406 zone->wait_table = (wait_queue_head_t *)
5407 memblock_virt_alloc_node_nopanic(
5408 alloc_size, zone->zone_pgdat->node_id);
5411 * This case means that a zone whose size was 0 gets new memory
5412 * via memory hot-add.
5413 * But it may be the case that a new node was hot-added. In
5414 * this case vmalloc() will not be able to use this new node's
5415 * memory - this wait_table must be initialized to use this new
5416 * node itself as well.
5417 * To use this new node's memory, further consideration will be
5420 zone->wait_table = vmalloc(alloc_size);
5422 if (!zone->wait_table)
5425 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5426 init_waitqueue_head(zone->wait_table + i);
5431 static __meminit void zone_pcp_init(struct zone *zone)
5434 * per cpu subsystem is not up at this point. The following code
5435 * relies on the ability of the linker to provide the
5436 * offset of a (static) per cpu variable into the per cpu area.
5438 zone->pageset = &boot_pageset;
5440 if (populated_zone(zone))
5441 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5442 zone->name, zone->present_pages,
5443 zone_batchsize(zone));
5446 int __meminit init_currently_empty_zone(struct zone *zone,
5447 unsigned long zone_start_pfn,
5450 struct pglist_data *pgdat = zone->zone_pgdat;
5452 ret = zone_wait_table_init(zone, size);
5455 pgdat->nr_zones = zone_idx(zone) + 1;
5457 zone->zone_start_pfn = zone_start_pfn;
5459 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5460 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5462 (unsigned long)zone_idx(zone),
5463 zone_start_pfn, (zone_start_pfn + size));
5465 zone_init_free_lists(zone);
5470 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5471 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5474 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5476 int __meminit __early_pfn_to_nid(unsigned long pfn,
5477 struct mminit_pfnnid_cache *state)
5479 unsigned long start_pfn, end_pfn;
5482 if (state->last_start <= pfn && pfn < state->last_end)
5483 return state->last_nid;
5485 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5487 state->last_start = start_pfn;
5488 state->last_end = end_pfn;
5489 state->last_nid = nid;
5494 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5497 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5498 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5499 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5501 * If an architecture guarantees that all ranges registered contain no holes
5502 * and may be freed, this this function may be used instead of calling
5503 * memblock_free_early_nid() manually.
5505 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5507 unsigned long start_pfn, end_pfn;
5510 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5511 start_pfn = min(start_pfn, max_low_pfn);
5512 end_pfn = min(end_pfn, max_low_pfn);
5514 if (start_pfn < end_pfn)
5515 memblock_free_early_nid(PFN_PHYS(start_pfn),
5516 (end_pfn - start_pfn) << PAGE_SHIFT,
5522 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5523 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5525 * If an architecture guarantees that all ranges registered contain no holes and may
5526 * be freed, this function may be used instead of calling memory_present() manually.
5528 void __init sparse_memory_present_with_active_regions(int nid)
5530 unsigned long start_pfn, end_pfn;
5533 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5534 memory_present(this_nid, start_pfn, end_pfn);
5538 * get_pfn_range_for_nid - Return the start and end page frames for a node
5539 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5540 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5541 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5543 * It returns the start and end page frame of a node based on information
5544 * provided by memblock_set_node(). If called for a node
5545 * with no available memory, a warning is printed and the start and end
5548 void __meminit get_pfn_range_for_nid(unsigned int nid,
5549 unsigned long *start_pfn, unsigned long *end_pfn)
5551 unsigned long this_start_pfn, this_end_pfn;
5557 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5558 *start_pfn = min(*start_pfn, this_start_pfn);
5559 *end_pfn = max(*end_pfn, this_end_pfn);
5562 if (*start_pfn == -1UL)
5567 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5568 * assumption is made that zones within a node are ordered in monotonic
5569 * increasing memory addresses so that the "highest" populated zone is used
5571 static void __init find_usable_zone_for_movable(void)
5574 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5575 if (zone_index == ZONE_MOVABLE)
5578 if (arch_zone_highest_possible_pfn[zone_index] >
5579 arch_zone_lowest_possible_pfn[zone_index])
5583 VM_BUG_ON(zone_index == -1);
5584 movable_zone = zone_index;
5588 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5589 * because it is sized independent of architecture. Unlike the other zones,
5590 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5591 * in each node depending on the size of each node and how evenly kernelcore
5592 * is distributed. This helper function adjusts the zone ranges
5593 * provided by the architecture for a given node by using the end of the
5594 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5595 * zones within a node are in order of monotonic increases memory addresses
5597 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5598 unsigned long zone_type,
5599 unsigned long node_start_pfn,
5600 unsigned long node_end_pfn,
5601 unsigned long *zone_start_pfn,
5602 unsigned long *zone_end_pfn)
5604 /* Only adjust if ZONE_MOVABLE is on this node */
5605 if (zone_movable_pfn[nid]) {
5606 /* Size ZONE_MOVABLE */
5607 if (zone_type == ZONE_MOVABLE) {
5608 *zone_start_pfn = zone_movable_pfn[nid];
5609 *zone_end_pfn = min(node_end_pfn,
5610 arch_zone_highest_possible_pfn[movable_zone]);
5612 /* Check if this whole range is within ZONE_MOVABLE */
5613 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5614 *zone_start_pfn = *zone_end_pfn;
5619 * Return the number of pages a zone spans in a node, including holes
5620 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5622 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5623 unsigned long zone_type,
5624 unsigned long node_start_pfn,
5625 unsigned long node_end_pfn,
5626 unsigned long *zone_start_pfn,
5627 unsigned long *zone_end_pfn,
5628 unsigned long *ignored)
5630 /* When hotadd a new node from cpu_up(), the node should be empty */
5631 if (!node_start_pfn && !node_end_pfn)
5634 /* Get the start and end of the zone */
5635 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5636 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5637 adjust_zone_range_for_zone_movable(nid, zone_type,
5638 node_start_pfn, node_end_pfn,
5639 zone_start_pfn, zone_end_pfn);
5641 /* Check that this node has pages within the zone's required range */
5642 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5645 /* Move the zone boundaries inside the node if necessary */
5646 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5647 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5649 /* Return the spanned pages */
5650 return *zone_end_pfn - *zone_start_pfn;
5654 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5655 * then all holes in the requested range will be accounted for.
5657 unsigned long __meminit __absent_pages_in_range(int nid,
5658 unsigned long range_start_pfn,
5659 unsigned long range_end_pfn)
5661 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5662 unsigned long start_pfn, end_pfn;
5665 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5666 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5667 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5668 nr_absent -= end_pfn - start_pfn;
5674 * absent_pages_in_range - Return number of page frames in holes within a range
5675 * @start_pfn: The start PFN to start searching for holes
5676 * @end_pfn: The end PFN to stop searching for holes
5678 * It returns the number of pages frames in memory holes within a range.
5680 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5681 unsigned long end_pfn)
5683 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5686 /* Return the number of page frames in holes in a zone on a node */
5687 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5688 unsigned long zone_type,
5689 unsigned long node_start_pfn,
5690 unsigned long node_end_pfn,
5691 unsigned long *ignored)
5693 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5694 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5695 unsigned long zone_start_pfn, zone_end_pfn;
5696 unsigned long nr_absent;
5698 /* When hotadd a new node from cpu_up(), the node should be empty */
5699 if (!node_start_pfn && !node_end_pfn)
5702 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5703 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5705 adjust_zone_range_for_zone_movable(nid, zone_type,
5706 node_start_pfn, node_end_pfn,
5707 &zone_start_pfn, &zone_end_pfn);
5708 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5711 * ZONE_MOVABLE handling.
5712 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5715 if (zone_movable_pfn[nid]) {
5716 if (mirrored_kernelcore) {
5717 unsigned long start_pfn, end_pfn;
5718 struct memblock_region *r;
5720 for_each_memblock(memory, r) {
5721 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5722 zone_start_pfn, zone_end_pfn);
5723 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5724 zone_start_pfn, zone_end_pfn);
5726 if (zone_type == ZONE_MOVABLE &&
5727 memblock_is_mirror(r))
5728 nr_absent += end_pfn - start_pfn;
5730 if (zone_type == ZONE_NORMAL &&
5731 !memblock_is_mirror(r))
5732 nr_absent += end_pfn - start_pfn;
5735 if (zone_type == ZONE_NORMAL)
5736 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5743 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5744 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5745 unsigned long zone_type,
5746 unsigned long node_start_pfn,
5747 unsigned long node_end_pfn,
5748 unsigned long *zone_start_pfn,
5749 unsigned long *zone_end_pfn,
5750 unsigned long *zones_size)
5754 *zone_start_pfn = node_start_pfn;
5755 for (zone = 0; zone < zone_type; zone++)
5756 *zone_start_pfn += zones_size[zone];
5758 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5760 return zones_size[zone_type];
5763 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5764 unsigned long zone_type,
5765 unsigned long node_start_pfn,
5766 unsigned long node_end_pfn,
5767 unsigned long *zholes_size)
5772 return zholes_size[zone_type];
5775 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5777 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5778 unsigned long node_start_pfn,
5779 unsigned long node_end_pfn,
5780 unsigned long *zones_size,
5781 unsigned long *zholes_size)
5783 unsigned long realtotalpages = 0, totalpages = 0;
5786 for (i = 0; i < MAX_NR_ZONES; i++) {
5787 struct zone *zone = pgdat->node_zones + i;
5788 unsigned long zone_start_pfn, zone_end_pfn;
5789 unsigned long size, real_size;
5791 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5797 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5798 node_start_pfn, node_end_pfn,
5801 zone->zone_start_pfn = zone_start_pfn;
5803 zone->zone_start_pfn = 0;
5804 zone->spanned_pages = size;
5805 zone->present_pages = real_size;
5808 realtotalpages += real_size;
5811 pgdat->node_spanned_pages = totalpages;
5812 pgdat->node_present_pages = realtotalpages;
5813 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5817 #ifndef CONFIG_SPARSEMEM
5819 * Calculate the size of the zone->blockflags rounded to an unsigned long
5820 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5821 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5822 * round what is now in bits to nearest long in bits, then return it in
5825 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5827 unsigned long usemapsize;
5829 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5830 usemapsize = roundup(zonesize, pageblock_nr_pages);
5831 usemapsize = usemapsize >> pageblock_order;
5832 usemapsize *= NR_PAGEBLOCK_BITS;
5833 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5835 return usemapsize / 8;
5838 static void __init setup_usemap(struct pglist_data *pgdat,
5840 unsigned long zone_start_pfn,
5841 unsigned long zonesize)
5843 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5844 zone->pageblock_flags = NULL;
5846 zone->pageblock_flags =
5847 memblock_virt_alloc_node_nopanic(usemapsize,
5851 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5852 unsigned long zone_start_pfn, unsigned long zonesize) {}
5853 #endif /* CONFIG_SPARSEMEM */
5855 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5857 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5858 void __paginginit set_pageblock_order(void)
5862 /* Check that pageblock_nr_pages has not already been setup */
5863 if (pageblock_order)
5866 if (HPAGE_SHIFT > PAGE_SHIFT)
5867 order = HUGETLB_PAGE_ORDER;
5869 order = MAX_ORDER - 1;
5872 * Assume the largest contiguous order of interest is a huge page.
5873 * This value may be variable depending on boot parameters on IA64 and
5876 pageblock_order = order;
5878 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5881 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5882 * is unused as pageblock_order is set at compile-time. See
5883 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5886 void __paginginit set_pageblock_order(void)
5890 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5892 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5893 unsigned long present_pages)
5895 unsigned long pages = spanned_pages;
5898 * Provide a more accurate estimation if there are holes within
5899 * the zone and SPARSEMEM is in use. If there are holes within the
5900 * zone, each populated memory region may cost us one or two extra
5901 * memmap pages due to alignment because memmap pages for each
5902 * populated regions may not naturally algined on page boundary.
5903 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5905 if (spanned_pages > present_pages + (present_pages >> 4) &&
5906 IS_ENABLED(CONFIG_SPARSEMEM))
5907 pages = present_pages;
5909 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5913 * Set up the zone data structures:
5914 * - mark all pages reserved
5915 * - mark all memory queues empty
5916 * - clear the memory bitmaps
5918 * NOTE: pgdat should get zeroed by caller.
5920 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5923 int nid = pgdat->node_id;
5926 pgdat_resize_init(pgdat);
5927 #ifdef CONFIG_NUMA_BALANCING
5928 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5929 pgdat->numabalancing_migrate_nr_pages = 0;
5930 pgdat->numabalancing_migrate_next_window = jiffies;
5932 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5933 spin_lock_init(&pgdat->split_queue_lock);
5934 INIT_LIST_HEAD(&pgdat->split_queue);
5935 pgdat->split_queue_len = 0;
5937 init_waitqueue_head(&pgdat->kswapd_wait);
5938 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5939 #ifdef CONFIG_COMPACTION
5940 init_waitqueue_head(&pgdat->kcompactd_wait);
5942 pgdat_page_ext_init(pgdat);
5944 for (j = 0; j < MAX_NR_ZONES; j++) {
5945 struct zone *zone = pgdat->node_zones + j;
5946 unsigned long size, realsize, freesize, memmap_pages;
5947 unsigned long zone_start_pfn = zone->zone_start_pfn;
5949 size = zone->spanned_pages;
5950 realsize = freesize = zone->present_pages;
5953 * Adjust freesize so that it accounts for how much memory
5954 * is used by this zone for memmap. This affects the watermark
5955 * and per-cpu initialisations
5957 memmap_pages = calc_memmap_size(size, realsize);
5958 if (!is_highmem_idx(j)) {
5959 if (freesize >= memmap_pages) {
5960 freesize -= memmap_pages;
5963 " %s zone: %lu pages used for memmap\n",
5964 zone_names[j], memmap_pages);
5966 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5967 zone_names[j], memmap_pages, freesize);
5970 /* Account for reserved pages */
5971 if (j == 0 && freesize > dma_reserve) {
5972 freesize -= dma_reserve;
5973 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5974 zone_names[0], dma_reserve);
5977 if (!is_highmem_idx(j))
5978 nr_kernel_pages += freesize;
5979 /* Charge for highmem memmap if there are enough kernel pages */
5980 else if (nr_kernel_pages > memmap_pages * 2)
5981 nr_kernel_pages -= memmap_pages;
5982 nr_all_pages += freesize;
5985 * Set an approximate value for lowmem here, it will be adjusted
5986 * when the bootmem allocator frees pages into the buddy system.
5987 * And all highmem pages will be managed by the buddy system.
5989 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5992 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5994 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5996 zone->name = zone_names[j];
5997 spin_lock_init(&zone->lock);
5998 spin_lock_init(&zone->lru_lock);
5999 zone_seqlock_init(zone);
6000 zone->zone_pgdat = pgdat;
6001 zone_pcp_init(zone);
6003 /* For bootup, initialized properly in watermark setup */
6004 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
6006 lruvec_init(&zone->lruvec);
6010 set_pageblock_order();
6011 setup_usemap(pgdat, zone, zone_start_pfn, size);
6012 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6014 memmap_init(size, nid, j, zone_start_pfn);
6018 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
6020 unsigned long __maybe_unused start = 0;
6021 unsigned long __maybe_unused offset = 0;
6023 /* Skip empty nodes */
6024 if (!pgdat->node_spanned_pages)
6027 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6028 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6029 offset = pgdat->node_start_pfn - start;
6030 /* ia64 gets its own node_mem_map, before this, without bootmem */
6031 if (!pgdat->node_mem_map) {
6032 unsigned long size, end;
6036 * The zone's endpoints aren't required to be MAX_ORDER
6037 * aligned but the node_mem_map endpoints must be in order
6038 * for the buddy allocator to function correctly.
6040 end = pgdat_end_pfn(pgdat);
6041 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6042 size = (end - start) * sizeof(struct page);
6043 map = alloc_remap(pgdat->node_id, size);
6045 map = memblock_virt_alloc_node_nopanic(size,
6047 pgdat->node_mem_map = map + offset;
6049 #ifndef CONFIG_NEED_MULTIPLE_NODES
6051 * With no DISCONTIG, the global mem_map is just set as node 0's
6053 if (pgdat == NODE_DATA(0)) {
6054 mem_map = NODE_DATA(0)->node_mem_map;
6055 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6056 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6058 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6061 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6064 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6065 unsigned long node_start_pfn, unsigned long *zholes_size)
6067 pg_data_t *pgdat = NODE_DATA(nid);
6068 unsigned long start_pfn = 0;
6069 unsigned long end_pfn = 0;
6071 /* pg_data_t should be reset to zero when it's allocated */
6072 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6074 reset_deferred_meminit(pgdat);
6075 pgdat->node_id = nid;
6076 pgdat->node_start_pfn = node_start_pfn;
6077 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6078 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6079 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6080 (u64)start_pfn << PAGE_SHIFT,
6081 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6083 start_pfn = node_start_pfn;
6085 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6086 zones_size, zholes_size);
6088 alloc_node_mem_map(pgdat);
6089 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6090 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6091 nid, (unsigned long)pgdat,
6092 (unsigned long)pgdat->node_mem_map);
6095 free_area_init_core(pgdat);
6098 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6100 #if MAX_NUMNODES > 1
6102 * Figure out the number of possible node ids.
6104 void __init setup_nr_node_ids(void)
6106 unsigned int highest;
6108 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6109 nr_node_ids = highest + 1;
6114 * node_map_pfn_alignment - determine the maximum internode alignment
6116 * This function should be called after node map is populated and sorted.
6117 * It calculates the maximum power of two alignment which can distinguish
6120 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6121 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6122 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6123 * shifted, 1GiB is enough and this function will indicate so.
6125 * This is used to test whether pfn -> nid mapping of the chosen memory
6126 * model has fine enough granularity to avoid incorrect mapping for the
6127 * populated node map.
6129 * Returns the determined alignment in pfn's. 0 if there is no alignment
6130 * requirement (single node).
6132 unsigned long __init node_map_pfn_alignment(void)
6134 unsigned long accl_mask = 0, last_end = 0;
6135 unsigned long start, end, mask;
6139 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6140 if (!start || last_nid < 0 || last_nid == nid) {
6147 * Start with a mask granular enough to pin-point to the
6148 * start pfn and tick off bits one-by-one until it becomes
6149 * too coarse to separate the current node from the last.
6151 mask = ~((1 << __ffs(start)) - 1);
6152 while (mask && last_end <= (start & (mask << 1)))
6155 /* accumulate all internode masks */
6159 /* convert mask to number of pages */
6160 return ~accl_mask + 1;
6163 /* Find the lowest pfn for a node */
6164 static unsigned long __init find_min_pfn_for_node(int nid)
6166 unsigned long min_pfn = ULONG_MAX;
6167 unsigned long start_pfn;
6170 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6171 min_pfn = min(min_pfn, start_pfn);
6173 if (min_pfn == ULONG_MAX) {
6174 pr_warn("Could not find start_pfn for node %d\n", nid);
6182 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6184 * It returns the minimum PFN based on information provided via
6185 * memblock_set_node().
6187 unsigned long __init find_min_pfn_with_active_regions(void)
6189 return find_min_pfn_for_node(MAX_NUMNODES);
6193 * early_calculate_totalpages()
6194 * Sum pages in active regions for movable zone.
6195 * Populate N_MEMORY for calculating usable_nodes.
6197 static unsigned long __init early_calculate_totalpages(void)
6199 unsigned long totalpages = 0;
6200 unsigned long start_pfn, end_pfn;
6203 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6204 unsigned long pages = end_pfn - start_pfn;
6206 totalpages += pages;
6208 node_set_state(nid, N_MEMORY);
6214 * Find the PFN the Movable zone begins in each node. Kernel memory
6215 * is spread evenly between nodes as long as the nodes have enough
6216 * memory. When they don't, some nodes will have more kernelcore than
6219 static void __init find_zone_movable_pfns_for_nodes(void)
6222 unsigned long usable_startpfn;
6223 unsigned long kernelcore_node, kernelcore_remaining;
6224 /* save the state before borrow the nodemask */
6225 nodemask_t saved_node_state = node_states[N_MEMORY];
6226 unsigned long totalpages = early_calculate_totalpages();
6227 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6228 struct memblock_region *r;
6230 /* Need to find movable_zone earlier when movable_node is specified. */
6231 find_usable_zone_for_movable();
6234 * If movable_node is specified, ignore kernelcore and movablecore
6237 if (movable_node_is_enabled()) {
6238 for_each_memblock(memory, r) {
6239 if (!memblock_is_hotpluggable(r))
6244 usable_startpfn = PFN_DOWN(r->base);
6245 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6246 min(usable_startpfn, zone_movable_pfn[nid]) :
6254 * If kernelcore=mirror is specified, ignore movablecore option
6256 if (mirrored_kernelcore) {
6257 bool mem_below_4gb_not_mirrored = false;
6259 for_each_memblock(memory, r) {
6260 if (memblock_is_mirror(r))
6265 usable_startpfn = memblock_region_memory_base_pfn(r);
6267 if (usable_startpfn < 0x100000) {
6268 mem_below_4gb_not_mirrored = true;
6272 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6273 min(usable_startpfn, zone_movable_pfn[nid]) :
6277 if (mem_below_4gb_not_mirrored)
6278 pr_warn("This configuration results in unmirrored kernel memory.");
6284 * If movablecore=nn[KMG] was specified, calculate what size of
6285 * kernelcore that corresponds so that memory usable for
6286 * any allocation type is evenly spread. If both kernelcore
6287 * and movablecore are specified, then the value of kernelcore
6288 * will be used for required_kernelcore if it's greater than
6289 * what movablecore would have allowed.
6291 if (required_movablecore) {
6292 unsigned long corepages;
6295 * Round-up so that ZONE_MOVABLE is at least as large as what
6296 * was requested by the user
6298 required_movablecore =
6299 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6300 required_movablecore = min(totalpages, required_movablecore);
6301 corepages = totalpages - required_movablecore;
6303 required_kernelcore = max(required_kernelcore, corepages);
6307 * If kernelcore was not specified or kernelcore size is larger
6308 * than totalpages, there is no ZONE_MOVABLE.
6310 if (!required_kernelcore || required_kernelcore >= totalpages)
6313 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6314 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6317 /* Spread kernelcore memory as evenly as possible throughout nodes */
6318 kernelcore_node = required_kernelcore / usable_nodes;
6319 for_each_node_state(nid, N_MEMORY) {
6320 unsigned long start_pfn, end_pfn;
6323 * Recalculate kernelcore_node if the division per node
6324 * now exceeds what is necessary to satisfy the requested
6325 * amount of memory for the kernel
6327 if (required_kernelcore < kernelcore_node)
6328 kernelcore_node = required_kernelcore / usable_nodes;
6331 * As the map is walked, we track how much memory is usable
6332 * by the kernel using kernelcore_remaining. When it is
6333 * 0, the rest of the node is usable by ZONE_MOVABLE
6335 kernelcore_remaining = kernelcore_node;
6337 /* Go through each range of PFNs within this node */
6338 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6339 unsigned long size_pages;
6341 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6342 if (start_pfn >= end_pfn)
6345 /* Account for what is only usable for kernelcore */
6346 if (start_pfn < usable_startpfn) {
6347 unsigned long kernel_pages;
6348 kernel_pages = min(end_pfn, usable_startpfn)
6351 kernelcore_remaining -= min(kernel_pages,
6352 kernelcore_remaining);
6353 required_kernelcore -= min(kernel_pages,
6354 required_kernelcore);
6356 /* Continue if range is now fully accounted */
6357 if (end_pfn <= usable_startpfn) {
6360 * Push zone_movable_pfn to the end so
6361 * that if we have to rebalance
6362 * kernelcore across nodes, we will
6363 * not double account here
6365 zone_movable_pfn[nid] = end_pfn;
6368 start_pfn = usable_startpfn;
6372 * The usable PFN range for ZONE_MOVABLE is from
6373 * start_pfn->end_pfn. Calculate size_pages as the
6374 * number of pages used as kernelcore
6376 size_pages = end_pfn - start_pfn;
6377 if (size_pages > kernelcore_remaining)
6378 size_pages = kernelcore_remaining;
6379 zone_movable_pfn[nid] = start_pfn + size_pages;
6382 * Some kernelcore has been met, update counts and
6383 * break if the kernelcore for this node has been
6386 required_kernelcore -= min(required_kernelcore,
6388 kernelcore_remaining -= size_pages;
6389 if (!kernelcore_remaining)
6395 * If there is still required_kernelcore, we do another pass with one
6396 * less node in the count. This will push zone_movable_pfn[nid] further
6397 * along on the nodes that still have memory until kernelcore is
6401 if (usable_nodes && required_kernelcore > usable_nodes)
6405 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6406 for (nid = 0; nid < MAX_NUMNODES; nid++)
6407 zone_movable_pfn[nid] =
6408 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6411 /* restore the node_state */
6412 node_states[N_MEMORY] = saved_node_state;
6415 /* Any regular or high memory on that node ? */
6416 static void check_for_memory(pg_data_t *pgdat, int nid)
6418 enum zone_type zone_type;
6420 if (N_MEMORY == N_NORMAL_MEMORY)
6423 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6424 struct zone *zone = &pgdat->node_zones[zone_type];
6425 if (populated_zone(zone)) {
6426 node_set_state(nid, N_HIGH_MEMORY);
6427 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6428 zone_type <= ZONE_NORMAL)
6429 node_set_state(nid, N_NORMAL_MEMORY);
6436 * free_area_init_nodes - Initialise all pg_data_t and zone data
6437 * @max_zone_pfn: an array of max PFNs for each zone
6439 * This will call free_area_init_node() for each active node in the system.
6440 * Using the page ranges provided by memblock_set_node(), the size of each
6441 * zone in each node and their holes is calculated. If the maximum PFN
6442 * between two adjacent zones match, it is assumed that the zone is empty.
6443 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6444 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6445 * starts where the previous one ended. For example, ZONE_DMA32 starts
6446 * at arch_max_dma_pfn.
6448 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6450 unsigned long start_pfn, end_pfn;
6453 /* Record where the zone boundaries are */
6454 memset(arch_zone_lowest_possible_pfn, 0,
6455 sizeof(arch_zone_lowest_possible_pfn));
6456 memset(arch_zone_highest_possible_pfn, 0,
6457 sizeof(arch_zone_highest_possible_pfn));
6458 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6459 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6460 for (i = 1; i < MAX_NR_ZONES; i++) {
6461 if (i == ZONE_MOVABLE)
6463 arch_zone_lowest_possible_pfn[i] =
6464 arch_zone_highest_possible_pfn[i-1];
6465 arch_zone_highest_possible_pfn[i] =
6466 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6468 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6469 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6471 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6472 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6473 find_zone_movable_pfns_for_nodes();
6475 /* Print out the zone ranges */
6476 pr_info("Zone ranges:\n");
6477 for (i = 0; i < MAX_NR_ZONES; i++) {
6478 if (i == ZONE_MOVABLE)
6480 pr_info(" %-8s ", zone_names[i]);
6481 if (arch_zone_lowest_possible_pfn[i] ==
6482 arch_zone_highest_possible_pfn[i])
6485 pr_cont("[mem %#018Lx-%#018Lx]\n",
6486 (u64)arch_zone_lowest_possible_pfn[i]
6488 ((u64)arch_zone_highest_possible_pfn[i]
6489 << PAGE_SHIFT) - 1);
6492 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6493 pr_info("Movable zone start for each node\n");
6494 for (i = 0; i < MAX_NUMNODES; i++) {
6495 if (zone_movable_pfn[i])
6496 pr_info(" Node %d: %#018Lx\n", i,
6497 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6500 /* Print out the early node map */
6501 pr_info("Early memory node ranges\n");
6502 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6503 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6504 (u64)start_pfn << PAGE_SHIFT,
6505 ((u64)end_pfn << PAGE_SHIFT) - 1);
6507 /* Initialise every node */
6508 mminit_verify_pageflags_layout();
6509 setup_nr_node_ids();
6510 for_each_online_node(nid) {
6511 pg_data_t *pgdat = NODE_DATA(nid);
6512 free_area_init_node(nid, NULL,
6513 find_min_pfn_for_node(nid), NULL);
6515 /* Any memory on that node */
6516 if (pgdat->node_present_pages)
6517 node_set_state(nid, N_MEMORY);
6518 check_for_memory(pgdat, nid);
6522 static int __init cmdline_parse_core(char *p, unsigned long *core)
6524 unsigned long long coremem;
6528 coremem = memparse(p, &p);
6529 *core = coremem >> PAGE_SHIFT;
6531 /* Paranoid check that UL is enough for the coremem value */
6532 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6538 * kernelcore=size sets the amount of memory for use for allocations that
6539 * cannot be reclaimed or migrated.
6541 static int __init cmdline_parse_kernelcore(char *p)
6543 /* parse kernelcore=mirror */
6544 if (parse_option_str(p, "mirror")) {
6545 mirrored_kernelcore = true;
6549 return cmdline_parse_core(p, &required_kernelcore);
6553 * movablecore=size sets the amount of memory for use for allocations that
6554 * can be reclaimed or migrated.
6556 static int __init cmdline_parse_movablecore(char *p)
6558 return cmdline_parse_core(p, &required_movablecore);
6561 early_param("kernelcore", cmdline_parse_kernelcore);
6562 early_param("movablecore", cmdline_parse_movablecore);
6564 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6566 void adjust_managed_page_count(struct page *page, long count)
6568 spin_lock(&managed_page_count_lock);
6569 page_zone(page)->managed_pages += count;
6570 totalram_pages += count;
6571 #ifdef CONFIG_HIGHMEM
6572 if (PageHighMem(page))
6573 totalhigh_pages += count;
6575 spin_unlock(&managed_page_count_lock);
6577 EXPORT_SYMBOL(adjust_managed_page_count);
6579 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6582 unsigned long pages = 0;
6584 start = (void *)PAGE_ALIGN((unsigned long)start);
6585 end = (void *)((unsigned long)end & PAGE_MASK);
6586 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6587 if ((unsigned int)poison <= 0xFF)
6588 memset(pos, poison, PAGE_SIZE);
6589 free_reserved_page(virt_to_page(pos));
6593 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6594 s, pages << (PAGE_SHIFT - 10), start, end);
6598 EXPORT_SYMBOL(free_reserved_area);
6600 #ifdef CONFIG_HIGHMEM
6601 void free_highmem_page(struct page *page)
6603 __free_reserved_page(page);
6605 page_zone(page)->managed_pages++;
6611 void __init mem_init_print_info(const char *str)
6613 unsigned long physpages, codesize, datasize, rosize, bss_size;
6614 unsigned long init_code_size, init_data_size;
6616 physpages = get_num_physpages();
6617 codesize = _etext - _stext;
6618 datasize = _edata - _sdata;
6619 rosize = __end_rodata - __start_rodata;
6620 bss_size = __bss_stop - __bss_start;
6621 init_data_size = __init_end - __init_begin;
6622 init_code_size = _einittext - _sinittext;
6625 * Detect special cases and adjust section sizes accordingly:
6626 * 1) .init.* may be embedded into .data sections
6627 * 2) .init.text.* may be out of [__init_begin, __init_end],
6628 * please refer to arch/tile/kernel/vmlinux.lds.S.
6629 * 3) .rodata.* may be embedded into .text or .data sections.
6631 #define adj_init_size(start, end, size, pos, adj) \
6633 if (start <= pos && pos < end && size > adj) \
6637 adj_init_size(__init_begin, __init_end, init_data_size,
6638 _sinittext, init_code_size);
6639 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6640 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6641 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6642 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6644 #undef adj_init_size
6646 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6647 #ifdef CONFIG_HIGHMEM
6651 nr_free_pages() << (PAGE_SHIFT - 10),
6652 physpages << (PAGE_SHIFT - 10),
6653 codesize >> 10, datasize >> 10, rosize >> 10,
6654 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6655 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6656 totalcma_pages << (PAGE_SHIFT - 10),
6657 #ifdef CONFIG_HIGHMEM
6658 totalhigh_pages << (PAGE_SHIFT - 10),
6660 str ? ", " : "", str ? str : "");
6664 * set_dma_reserve - set the specified number of pages reserved in the first zone
6665 * @new_dma_reserve: The number of pages to mark reserved
6667 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6668 * In the DMA zone, a significant percentage may be consumed by kernel image
6669 * and other unfreeable allocations which can skew the watermarks badly. This
6670 * function may optionally be used to account for unfreeable pages in the
6671 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6672 * smaller per-cpu batchsize.
6674 void __init set_dma_reserve(unsigned long new_dma_reserve)
6676 dma_reserve = new_dma_reserve;
6679 void __init free_area_init(unsigned long *zones_size)
6681 free_area_init_node(0, zones_size,
6682 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6685 static int page_alloc_cpu_notify(struct notifier_block *self,
6686 unsigned long action, void *hcpu)
6688 int cpu = (unsigned long)hcpu;
6690 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6691 lru_add_drain_cpu(cpu);
6695 * Spill the event counters of the dead processor
6696 * into the current processors event counters.
6697 * This artificially elevates the count of the current
6700 vm_events_fold_cpu(cpu);
6703 * Zero the differential counters of the dead processor
6704 * so that the vm statistics are consistent.
6706 * This is only okay since the processor is dead and cannot
6707 * race with what we are doing.
6709 cpu_vm_stats_fold(cpu);
6714 void __init page_alloc_init(void)
6716 hotcpu_notifier(page_alloc_cpu_notify, 0);
6720 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6721 * or min_free_kbytes changes.
6723 static void calculate_totalreserve_pages(void)
6725 struct pglist_data *pgdat;
6726 unsigned long reserve_pages = 0;
6727 enum zone_type i, j;
6729 for_each_online_pgdat(pgdat) {
6730 for (i = 0; i < MAX_NR_ZONES; i++) {
6731 struct zone *zone = pgdat->node_zones + i;
6734 /* Find valid and maximum lowmem_reserve in the zone */
6735 for (j = i; j < MAX_NR_ZONES; j++) {
6736 if (zone->lowmem_reserve[j] > max)
6737 max = zone->lowmem_reserve[j];
6740 /* we treat the high watermark as reserved pages. */
6741 max += high_wmark_pages(zone);
6743 if (max > zone->managed_pages)
6744 max = zone->managed_pages;
6746 zone->totalreserve_pages = max;
6748 reserve_pages += max;
6751 totalreserve_pages = reserve_pages;
6755 * setup_per_zone_lowmem_reserve - called whenever
6756 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6757 * has a correct pages reserved value, so an adequate number of
6758 * pages are left in the zone after a successful __alloc_pages().
6760 static void setup_per_zone_lowmem_reserve(void)
6762 struct pglist_data *pgdat;
6763 enum zone_type j, idx;
6765 for_each_online_pgdat(pgdat) {
6766 for (j = 0; j < MAX_NR_ZONES; j++) {
6767 struct zone *zone = pgdat->node_zones + j;
6768 unsigned long managed_pages = zone->managed_pages;
6770 zone->lowmem_reserve[j] = 0;
6774 struct zone *lower_zone;
6778 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6779 sysctl_lowmem_reserve_ratio[idx] = 1;
6781 lower_zone = pgdat->node_zones + idx;
6782 lower_zone->lowmem_reserve[j] = managed_pages /
6783 sysctl_lowmem_reserve_ratio[idx];
6784 managed_pages += lower_zone->managed_pages;
6789 /* update totalreserve_pages */
6790 calculate_totalreserve_pages();
6793 static void __setup_per_zone_wmarks(void)
6795 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6796 unsigned long lowmem_pages = 0;
6798 unsigned long flags;
6800 /* Calculate total number of !ZONE_HIGHMEM pages */
6801 for_each_zone(zone) {
6802 if (!is_highmem(zone))
6803 lowmem_pages += zone->managed_pages;
6806 for_each_zone(zone) {
6809 spin_lock_irqsave(&zone->lock, flags);
6810 tmp = (u64)pages_min * zone->managed_pages;
6811 do_div(tmp, lowmem_pages);
6812 if (is_highmem(zone)) {
6814 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6815 * need highmem pages, so cap pages_min to a small
6818 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6819 * deltas control asynch page reclaim, and so should
6820 * not be capped for highmem.
6822 unsigned long min_pages;
6824 min_pages = zone->managed_pages / 1024;
6825 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6826 zone->watermark[WMARK_MIN] = min_pages;
6829 * If it's a lowmem zone, reserve a number of pages
6830 * proportionate to the zone's size.
6832 zone->watermark[WMARK_MIN] = tmp;
6836 * Set the kswapd watermarks distance according to the
6837 * scale factor in proportion to available memory, but
6838 * ensure a minimum size on small systems.
6840 tmp = max_t(u64, tmp >> 2,
6841 mult_frac(zone->managed_pages,
6842 watermark_scale_factor, 10000));
6844 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6845 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6847 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6848 high_wmark_pages(zone) - low_wmark_pages(zone) -
6849 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6851 spin_unlock_irqrestore(&zone->lock, flags);
6854 /* update totalreserve_pages */
6855 calculate_totalreserve_pages();
6859 * setup_per_zone_wmarks - called when min_free_kbytes changes
6860 * or when memory is hot-{added|removed}
6862 * Ensures that the watermark[min,low,high] values for each zone are set
6863 * correctly with respect to min_free_kbytes.
6865 void setup_per_zone_wmarks(void)
6867 mutex_lock(&zonelists_mutex);
6868 __setup_per_zone_wmarks();
6869 mutex_unlock(&zonelists_mutex);
6873 * Initialise min_free_kbytes.
6875 * For small machines we want it small (128k min). For large machines
6876 * we want it large (64MB max). But it is not linear, because network
6877 * bandwidth does not increase linearly with machine size. We use
6879 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6880 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6896 int __meminit init_per_zone_wmark_min(void)
6898 unsigned long lowmem_kbytes;
6899 int new_min_free_kbytes;
6901 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6902 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6904 if (new_min_free_kbytes > user_min_free_kbytes) {
6905 min_free_kbytes = new_min_free_kbytes;
6906 if (min_free_kbytes < 128)
6907 min_free_kbytes = 128;
6908 if (min_free_kbytes > 65536)
6909 min_free_kbytes = 65536;
6911 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6912 new_min_free_kbytes, user_min_free_kbytes);
6914 setup_per_zone_wmarks();
6915 refresh_zone_stat_thresholds();
6916 setup_per_zone_lowmem_reserve();
6919 core_initcall(init_per_zone_wmark_min)
6922 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6923 * that we can call two helper functions whenever min_free_kbytes
6926 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6927 void __user *buffer, size_t *length, loff_t *ppos)
6931 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6936 user_min_free_kbytes = min_free_kbytes;
6937 setup_per_zone_wmarks();
6942 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6943 void __user *buffer, size_t *length, loff_t *ppos)
6947 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6952 setup_per_zone_wmarks();
6958 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6959 void __user *buffer, size_t *length, loff_t *ppos)
6964 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6969 zone->min_unmapped_pages = (zone->managed_pages *
6970 sysctl_min_unmapped_ratio) / 100;
6974 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6975 void __user *buffer, size_t *length, loff_t *ppos)
6980 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6985 zone->min_slab_pages = (zone->managed_pages *
6986 sysctl_min_slab_ratio) / 100;
6992 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6993 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6994 * whenever sysctl_lowmem_reserve_ratio changes.
6996 * The reserve ratio obviously has absolutely no relation with the
6997 * minimum watermarks. The lowmem reserve ratio can only make sense
6998 * if in function of the boot time zone sizes.
7000 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7001 void __user *buffer, size_t *length, loff_t *ppos)
7003 proc_dointvec_minmax(table, write, buffer, length, ppos);
7004 setup_per_zone_lowmem_reserve();
7009 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7010 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7011 * pagelist can have before it gets flushed back to buddy allocator.
7013 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7014 void __user *buffer, size_t *length, loff_t *ppos)
7017 int old_percpu_pagelist_fraction;
7020 mutex_lock(&pcp_batch_high_lock);
7021 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7023 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7024 if (!write || ret < 0)
7027 /* Sanity checking to avoid pcp imbalance */
7028 if (percpu_pagelist_fraction &&
7029 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7030 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7036 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7039 for_each_populated_zone(zone) {
7042 for_each_possible_cpu(cpu)
7043 pageset_set_high_and_batch(zone,
7044 per_cpu_ptr(zone->pageset, cpu));
7047 mutex_unlock(&pcp_batch_high_lock);
7052 int hashdist = HASHDIST_DEFAULT;
7054 static int __init set_hashdist(char *str)
7058 hashdist = simple_strtoul(str, &str, 0);
7061 __setup("hashdist=", set_hashdist);
7065 * allocate a large system hash table from bootmem
7066 * - it is assumed that the hash table must contain an exact power-of-2
7067 * quantity of entries
7068 * - limit is the number of hash buckets, not the total allocation size
7070 void *__init alloc_large_system_hash(const char *tablename,
7071 unsigned long bucketsize,
7072 unsigned long numentries,
7075 unsigned int *_hash_shift,
7076 unsigned int *_hash_mask,
7077 unsigned long low_limit,
7078 unsigned long high_limit)
7080 unsigned long long max = high_limit;
7081 unsigned long log2qty, size;
7084 /* allow the kernel cmdline to have a say */
7086 /* round applicable memory size up to nearest megabyte */
7087 numentries = nr_kernel_pages;
7089 /* It isn't necessary when PAGE_SIZE >= 1MB */
7090 if (PAGE_SHIFT < 20)
7091 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7093 /* limit to 1 bucket per 2^scale bytes of low memory */
7094 if (scale > PAGE_SHIFT)
7095 numentries >>= (scale - PAGE_SHIFT);
7097 numentries <<= (PAGE_SHIFT - scale);
7099 /* Make sure we've got at least a 0-order allocation.. */
7100 if (unlikely(flags & HASH_SMALL)) {
7101 /* Makes no sense without HASH_EARLY */
7102 WARN_ON(!(flags & HASH_EARLY));
7103 if (!(numentries >> *_hash_shift)) {
7104 numentries = 1UL << *_hash_shift;
7105 BUG_ON(!numentries);
7107 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7108 numentries = PAGE_SIZE / bucketsize;
7110 numentries = roundup_pow_of_two(numentries);
7112 /* limit allocation size to 1/16 total memory by default */
7114 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7115 do_div(max, bucketsize);
7117 max = min(max, 0x80000000ULL);
7119 if (numentries < low_limit)
7120 numentries = low_limit;
7121 if (numentries > max)
7124 log2qty = ilog2(numentries);
7127 size = bucketsize << log2qty;
7128 if (flags & HASH_EARLY)
7129 table = memblock_virt_alloc_nopanic(size, 0);
7131 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7134 * If bucketsize is not a power-of-two, we may free
7135 * some pages at the end of hash table which
7136 * alloc_pages_exact() automatically does
7138 if (get_order(size) < MAX_ORDER) {
7139 table = alloc_pages_exact(size, GFP_ATOMIC);
7140 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7143 } while (!table && size > PAGE_SIZE && --log2qty);
7146 panic("Failed to allocate %s hash table\n", tablename);
7148 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7149 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7152 *_hash_shift = log2qty;
7154 *_hash_mask = (1 << log2qty) - 1;
7160 * This function checks whether pageblock includes unmovable pages or not.
7161 * If @count is not zero, it is okay to include less @count unmovable pages
7163 * PageLRU check without isolation or lru_lock could race so that
7164 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7165 * expect this function should be exact.
7167 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7168 bool skip_hwpoisoned_pages)
7170 unsigned long pfn, iter, found;
7174 * For avoiding noise data, lru_add_drain_all() should be called
7175 * If ZONE_MOVABLE, the zone never contains unmovable pages
7177 if (zone_idx(zone) == ZONE_MOVABLE)
7179 mt = get_pageblock_migratetype(page);
7180 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7183 pfn = page_to_pfn(page);
7184 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7185 unsigned long check = pfn + iter;
7187 if (!pfn_valid_within(check))
7190 page = pfn_to_page(check);
7193 * Hugepages are not in LRU lists, but they're movable.
7194 * We need not scan over tail pages bacause we don't
7195 * handle each tail page individually in migration.
7197 if (PageHuge(page)) {
7198 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7203 * We can't use page_count without pin a page
7204 * because another CPU can free compound page.
7205 * This check already skips compound tails of THP
7206 * because their page->_refcount is zero at all time.
7208 if (!page_ref_count(page)) {
7209 if (PageBuddy(page))
7210 iter += (1 << page_order(page)) - 1;
7215 * The HWPoisoned page may be not in buddy system, and
7216 * page_count() is not 0.
7218 if (skip_hwpoisoned_pages && PageHWPoison(page))
7224 * If there are RECLAIMABLE pages, we need to check
7225 * it. But now, memory offline itself doesn't call
7226 * shrink_node_slabs() and it still to be fixed.
7229 * If the page is not RAM, page_count()should be 0.
7230 * we don't need more check. This is an _used_ not-movable page.
7232 * The problematic thing here is PG_reserved pages. PG_reserved
7233 * is set to both of a memory hole page and a _used_ kernel
7242 bool is_pageblock_removable_nolock(struct page *page)
7248 * We have to be careful here because we are iterating over memory
7249 * sections which are not zone aware so we might end up outside of
7250 * the zone but still within the section.
7251 * We have to take care about the node as well. If the node is offline
7252 * its NODE_DATA will be NULL - see page_zone.
7254 if (!node_online(page_to_nid(page)))
7257 zone = page_zone(page);
7258 pfn = page_to_pfn(page);
7259 if (!zone_spans_pfn(zone, pfn))
7262 return !has_unmovable_pages(zone, page, 0, true);
7265 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7267 static unsigned long pfn_max_align_down(unsigned long pfn)
7269 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7270 pageblock_nr_pages) - 1);
7273 static unsigned long pfn_max_align_up(unsigned long pfn)
7275 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7276 pageblock_nr_pages));
7279 /* [start, end) must belong to a single zone. */
7280 static int __alloc_contig_migrate_range(struct compact_control *cc,
7281 unsigned long start, unsigned long end)
7283 /* This function is based on compact_zone() from compaction.c. */
7284 unsigned long nr_reclaimed;
7285 unsigned long pfn = start;
7286 unsigned int tries = 0;
7291 while (pfn < end || !list_empty(&cc->migratepages)) {
7292 if (fatal_signal_pending(current)) {
7297 if (list_empty(&cc->migratepages)) {
7298 cc->nr_migratepages = 0;
7299 pfn = isolate_migratepages_range(cc, pfn, end);
7305 } else if (++tries == 5) {
7306 ret = ret < 0 ? ret : -EBUSY;
7310 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7312 cc->nr_migratepages -= nr_reclaimed;
7314 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7315 NULL, 0, cc->mode, MR_CMA);
7318 putback_movable_pages(&cc->migratepages);
7325 * alloc_contig_range() -- tries to allocate given range of pages
7326 * @start: start PFN to allocate
7327 * @end: one-past-the-last PFN to allocate
7328 * @migratetype: migratetype of the underlaying pageblocks (either
7329 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7330 * in range must have the same migratetype and it must
7331 * be either of the two.
7333 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7334 * aligned, however it's the caller's responsibility to guarantee that
7335 * we are the only thread that changes migrate type of pageblocks the
7338 * The PFN range must belong to a single zone.
7340 * Returns zero on success or negative error code. On success all
7341 * pages which PFN is in [start, end) are allocated for the caller and
7342 * need to be freed with free_contig_range().
7344 int alloc_contig_range(unsigned long start, unsigned long end,
7345 unsigned migratetype)
7347 unsigned long outer_start, outer_end;
7351 struct compact_control cc = {
7352 .nr_migratepages = 0,
7354 .zone = page_zone(pfn_to_page(start)),
7355 .mode = MIGRATE_SYNC,
7356 .ignore_skip_hint = true,
7358 INIT_LIST_HEAD(&cc.migratepages);
7361 * What we do here is we mark all pageblocks in range as
7362 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7363 * have different sizes, and due to the way page allocator
7364 * work, we align the range to biggest of the two pages so
7365 * that page allocator won't try to merge buddies from
7366 * different pageblocks and change MIGRATE_ISOLATE to some
7367 * other migration type.
7369 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7370 * migrate the pages from an unaligned range (ie. pages that
7371 * we are interested in). This will put all the pages in
7372 * range back to page allocator as MIGRATE_ISOLATE.
7374 * When this is done, we take the pages in range from page
7375 * allocator removing them from the buddy system. This way
7376 * page allocator will never consider using them.
7378 * This lets us mark the pageblocks back as
7379 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7380 * aligned range but not in the unaligned, original range are
7381 * put back to page allocator so that buddy can use them.
7384 ret = start_isolate_page_range(pfn_max_align_down(start),
7385 pfn_max_align_up(end), migratetype,
7391 * In case of -EBUSY, we'd like to know which page causes problem.
7392 * So, just fall through. We will check it in test_pages_isolated().
7394 ret = __alloc_contig_migrate_range(&cc, start, end);
7395 if (ret && ret != -EBUSY)
7399 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7400 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7401 * more, all pages in [start, end) are free in page allocator.
7402 * What we are going to do is to allocate all pages from
7403 * [start, end) (that is remove them from page allocator).
7405 * The only problem is that pages at the beginning and at the
7406 * end of interesting range may be not aligned with pages that
7407 * page allocator holds, ie. they can be part of higher order
7408 * pages. Because of this, we reserve the bigger range and
7409 * once this is done free the pages we are not interested in.
7411 * We don't have to hold zone->lock here because the pages are
7412 * isolated thus they won't get removed from buddy.
7415 lru_add_drain_all();
7416 drain_all_pages(cc.zone);
7419 outer_start = start;
7420 while (!PageBuddy(pfn_to_page(outer_start))) {
7421 if (++order >= MAX_ORDER) {
7422 outer_start = start;
7425 outer_start &= ~0UL << order;
7428 if (outer_start != start) {
7429 order = page_order(pfn_to_page(outer_start));
7432 * outer_start page could be small order buddy page and
7433 * it doesn't include start page. Adjust outer_start
7434 * in this case to report failed page properly
7435 * on tracepoint in test_pages_isolated()
7437 if (outer_start + (1UL << order) <= start)
7438 outer_start = start;
7441 /* Make sure the range is really isolated. */
7442 if (test_pages_isolated(outer_start, end, false)) {
7443 pr_info("%s: [%lx, %lx) PFNs busy\n",
7444 __func__, outer_start, end);
7449 /* Grab isolated pages from freelists. */
7450 outer_end = isolate_freepages_range(&cc, outer_start, end);
7456 /* Free head and tail (if any) */
7457 if (start != outer_start)
7458 free_contig_range(outer_start, start - outer_start);
7459 if (end != outer_end)
7460 free_contig_range(end, outer_end - end);
7463 undo_isolate_page_range(pfn_max_align_down(start),
7464 pfn_max_align_up(end), migratetype);
7468 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7470 unsigned int count = 0;
7472 for (; nr_pages--; pfn++) {
7473 struct page *page = pfn_to_page(pfn);
7475 count += page_count(page) != 1;
7478 WARN(count != 0, "%d pages are still in use!\n", count);
7482 #ifdef CONFIG_MEMORY_HOTPLUG
7484 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7485 * page high values need to be recalulated.
7487 void __meminit zone_pcp_update(struct zone *zone)
7490 mutex_lock(&pcp_batch_high_lock);
7491 for_each_possible_cpu(cpu)
7492 pageset_set_high_and_batch(zone,
7493 per_cpu_ptr(zone->pageset, cpu));
7494 mutex_unlock(&pcp_batch_high_lock);
7498 void zone_pcp_reset(struct zone *zone)
7500 unsigned long flags;
7502 struct per_cpu_pageset *pset;
7504 /* avoid races with drain_pages() */
7505 local_irq_save(flags);
7506 if (zone->pageset != &boot_pageset) {
7507 for_each_online_cpu(cpu) {
7508 pset = per_cpu_ptr(zone->pageset, cpu);
7509 drain_zonestat(zone, pset);
7511 free_percpu(zone->pageset);
7512 zone->pageset = &boot_pageset;
7514 local_irq_restore(flags);
7517 #ifdef CONFIG_MEMORY_HOTREMOVE
7519 * All pages in the range must be in a single zone and isolated
7520 * before calling this.
7523 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7527 unsigned int order, i;
7529 unsigned long flags;
7530 /* find the first valid pfn */
7531 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7536 zone = page_zone(pfn_to_page(pfn));
7537 spin_lock_irqsave(&zone->lock, flags);
7539 while (pfn < end_pfn) {
7540 if (!pfn_valid(pfn)) {
7544 page = pfn_to_page(pfn);
7546 * The HWPoisoned page may be not in buddy system, and
7547 * page_count() is not 0.
7549 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7551 SetPageReserved(page);
7555 BUG_ON(page_count(page));
7556 BUG_ON(!PageBuddy(page));
7557 order = page_order(page);
7558 #ifdef CONFIG_DEBUG_VM
7559 pr_info("remove from free list %lx %d %lx\n",
7560 pfn, 1 << order, end_pfn);
7562 list_del(&page->lru);
7563 rmv_page_order(page);
7564 zone->free_area[order].nr_free--;
7565 for (i = 0; i < (1 << order); i++)
7566 SetPageReserved((page+i));
7567 pfn += (1 << order);
7569 spin_unlock_irqrestore(&zone->lock, flags);
7573 bool is_free_buddy_page(struct page *page)
7575 struct zone *zone = page_zone(page);
7576 unsigned long pfn = page_to_pfn(page);
7577 unsigned long flags;
7580 spin_lock_irqsave(&zone->lock, flags);
7581 for (order = 0; order < MAX_ORDER; order++) {
7582 struct page *page_head = page - (pfn & ((1 << order) - 1));
7584 if (PageBuddy(page_head) && page_order(page_head) >= order)
7587 spin_unlock_irqrestore(&zone->lock, flags);
7589 return order < MAX_ORDER;