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);
2618 } while (page && check_new_pcp(page));
2620 __dec_zone_state(zone, NR_ALLOC_BATCH);
2621 list_del(&page->lru);
2625 * We most definitely don't want callers attempting to
2626 * allocate greater than order-1 page units with __GFP_NOFAIL.
2628 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2629 spin_lock_irqsave(&zone->lock, flags);
2633 if (alloc_flags & ALLOC_HARDER) {
2634 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2636 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2639 page = __rmqueue(zone, order, migratetype);
2640 } while (page && check_new_pages(page, order));
2641 spin_unlock(&zone->lock);
2644 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2645 __mod_zone_freepage_state(zone, -(1 << order),
2646 get_pcppage_migratetype(page));
2649 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2650 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2651 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2653 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2654 zone_statistics(preferred_zone, zone, gfp_flags);
2655 local_irq_restore(flags);
2657 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2661 local_irq_restore(flags);
2665 #ifdef CONFIG_FAIL_PAGE_ALLOC
2668 struct fault_attr attr;
2670 bool ignore_gfp_highmem;
2671 bool ignore_gfp_reclaim;
2673 } fail_page_alloc = {
2674 .attr = FAULT_ATTR_INITIALIZER,
2675 .ignore_gfp_reclaim = true,
2676 .ignore_gfp_highmem = true,
2680 static int __init setup_fail_page_alloc(char *str)
2682 return setup_fault_attr(&fail_page_alloc.attr, str);
2684 __setup("fail_page_alloc=", setup_fail_page_alloc);
2686 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2688 if (order < fail_page_alloc.min_order)
2690 if (gfp_mask & __GFP_NOFAIL)
2692 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2694 if (fail_page_alloc.ignore_gfp_reclaim &&
2695 (gfp_mask & __GFP_DIRECT_RECLAIM))
2698 return should_fail(&fail_page_alloc.attr, 1 << order);
2701 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2703 static int __init fail_page_alloc_debugfs(void)
2705 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2708 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2709 &fail_page_alloc.attr);
2711 return PTR_ERR(dir);
2713 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2714 &fail_page_alloc.ignore_gfp_reclaim))
2716 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2717 &fail_page_alloc.ignore_gfp_highmem))
2719 if (!debugfs_create_u32("min-order", mode, dir,
2720 &fail_page_alloc.min_order))
2725 debugfs_remove_recursive(dir);
2730 late_initcall(fail_page_alloc_debugfs);
2732 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2734 #else /* CONFIG_FAIL_PAGE_ALLOC */
2736 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2741 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2744 * Return true if free base pages are above 'mark'. For high-order checks it
2745 * will return true of the order-0 watermark is reached and there is at least
2746 * one free page of a suitable size. Checking now avoids taking the zone lock
2747 * to check in the allocation paths if no pages are free.
2749 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2750 int classzone_idx, unsigned int alloc_flags,
2755 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2757 /* free_pages may go negative - that's OK */
2758 free_pages -= (1 << order) - 1;
2760 if (alloc_flags & ALLOC_HIGH)
2764 * If the caller does not have rights to ALLOC_HARDER then subtract
2765 * the high-atomic reserves. This will over-estimate the size of the
2766 * atomic reserve but it avoids a search.
2768 if (likely(!alloc_harder))
2769 free_pages -= z->nr_reserved_highatomic;
2774 /* If allocation can't use CMA areas don't use free CMA pages */
2775 if (!(alloc_flags & ALLOC_CMA))
2776 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2780 * Check watermarks for an order-0 allocation request. If these
2781 * are not met, then a high-order request also cannot go ahead
2782 * even if a suitable page happened to be free.
2784 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2787 /* If this is an order-0 request then the watermark is fine */
2791 /* For a high-order request, check at least one suitable page is free */
2792 for (o = order; o < MAX_ORDER; o++) {
2793 struct free_area *area = &z->free_area[o];
2802 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2803 if (!list_empty(&area->free_list[mt]))
2808 if ((alloc_flags & ALLOC_CMA) &&
2809 !list_empty(&area->free_list[MIGRATE_CMA])) {
2817 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2818 int classzone_idx, unsigned int alloc_flags)
2820 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2821 zone_page_state(z, NR_FREE_PAGES));
2824 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2825 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2827 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2831 /* If allocation can't use CMA areas don't use free CMA pages */
2832 if (!(alloc_flags & ALLOC_CMA))
2833 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2837 * Fast check for order-0 only. If this fails then the reserves
2838 * need to be calculated. There is a corner case where the check
2839 * passes but only the high-order atomic reserve are free. If
2840 * the caller is !atomic then it'll uselessly search the free
2841 * list. That corner case is then slower but it is harmless.
2843 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2846 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2850 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2851 unsigned long mark, int classzone_idx)
2853 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2855 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2856 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2858 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2863 static bool zone_local(struct zone *local_zone, struct zone *zone)
2865 return local_zone->node == zone->node;
2868 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2870 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2873 #else /* CONFIG_NUMA */
2874 static bool zone_local(struct zone *local_zone, struct zone *zone)
2879 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2883 #endif /* CONFIG_NUMA */
2885 static void reset_alloc_batches(struct zone *preferred_zone)
2887 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2890 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2891 high_wmark_pages(zone) - low_wmark_pages(zone) -
2892 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2893 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2894 } while (zone++ != preferred_zone);
2898 * get_page_from_freelist goes through the zonelist trying to allocate
2901 static struct page *
2902 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2903 const struct alloc_context *ac)
2905 struct zoneref *z = ac->preferred_zoneref;
2907 bool fair_skipped = false;
2908 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2912 * Scan zonelist, looking for a zone with enough free.
2913 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2915 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2920 if (cpusets_enabled() &&
2921 (alloc_flags & ALLOC_CPUSET) &&
2922 !__cpuset_zone_allowed(zone, gfp_mask))
2925 * Distribute pages in proportion to the individual
2926 * zone size to ensure fair page aging. The zone a
2927 * page was allocated in should have no effect on the
2928 * time the page has in memory before being reclaimed.
2931 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2932 fair_skipped = true;
2935 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2942 * When allocating a page cache page for writing, we
2943 * want to get it from a zone that is within its dirty
2944 * limit, such that no single zone holds more than its
2945 * proportional share of globally allowed dirty pages.
2946 * The dirty limits take into account the zone's
2947 * lowmem reserves and high watermark so that kswapd
2948 * should be able to balance it without having to
2949 * write pages from its LRU list.
2951 * This may look like it could increase pressure on
2952 * lower zones by failing allocations in higher zones
2953 * before they are full. But the pages that do spill
2954 * over are limited as the lower zones are protected
2955 * by this very same mechanism. It should not become
2956 * a practical burden to them.
2958 * XXX: For now, allow allocations to potentially
2959 * exceed the per-zone dirty limit in the slowpath
2960 * (spread_dirty_pages unset) before going into reclaim,
2961 * which is important when on a NUMA setup the allowed
2962 * zones are together not big enough to reach the
2963 * global limit. The proper fix for these situations
2964 * will require awareness of zones in the
2965 * dirty-throttling and the flusher threads.
2967 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2970 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2971 if (!zone_watermark_fast(zone, order, mark,
2972 ac_classzone_idx(ac), alloc_flags)) {
2975 /* Checked here to keep the fast path fast */
2976 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2977 if (alloc_flags & ALLOC_NO_WATERMARKS)
2980 if (zone_reclaim_mode == 0 ||
2981 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2984 ret = zone_reclaim(zone, gfp_mask, order);
2986 case ZONE_RECLAIM_NOSCAN:
2989 case ZONE_RECLAIM_FULL:
2990 /* scanned but unreclaimable */
2993 /* did we reclaim enough */
2994 if (zone_watermark_ok(zone, order, mark,
2995 ac_classzone_idx(ac), alloc_flags))
3003 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
3004 gfp_mask, alloc_flags, ac->migratetype);
3006 prep_new_page(page, order, gfp_mask, alloc_flags);
3009 * If this is a high-order atomic allocation then check
3010 * if the pageblock should be reserved for the future
3012 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3013 reserve_highatomic_pageblock(page, zone, order);
3020 * The first pass makes sure allocations are spread fairly within the
3021 * local node. However, the local node might have free pages left
3022 * after the fairness batches are exhausted, and remote zones haven't
3023 * even been considered yet. Try once more without fairness, and
3024 * include remote zones now, before entering the slowpath and waking
3025 * kswapd: prefer spilling to a remote zone over swapping locally.
3030 fair_skipped = false;
3031 reset_alloc_batches(ac->preferred_zoneref->zone);
3039 * Large machines with many possible nodes should not always dump per-node
3040 * meminfo in irq context.
3042 static inline bool should_suppress_show_mem(void)
3047 ret = in_interrupt();
3052 static DEFINE_RATELIMIT_STATE(nopage_rs,
3053 DEFAULT_RATELIMIT_INTERVAL,
3054 DEFAULT_RATELIMIT_BURST);
3056 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3058 unsigned int filter = SHOW_MEM_FILTER_NODES;
3060 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3061 debug_guardpage_minorder() > 0)
3065 * This documents exceptions given to allocations in certain
3066 * contexts that are allowed to allocate outside current's set
3069 if (!(gfp_mask & __GFP_NOMEMALLOC))
3070 if (test_thread_flag(TIF_MEMDIE) ||
3071 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3072 filter &= ~SHOW_MEM_FILTER_NODES;
3073 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3074 filter &= ~SHOW_MEM_FILTER_NODES;
3077 struct va_format vaf;
3080 va_start(args, fmt);
3085 pr_warn("%pV", &vaf);
3090 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3091 current->comm, order, gfp_mask, &gfp_mask);
3093 if (!should_suppress_show_mem())
3097 static inline struct page *
3098 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3099 const struct alloc_context *ac, unsigned long *did_some_progress)
3101 struct oom_control oc = {
3102 .zonelist = ac->zonelist,
3103 .nodemask = ac->nodemask,
3104 .gfp_mask = gfp_mask,
3109 *did_some_progress = 0;
3112 * Acquire the oom lock. If that fails, somebody else is
3113 * making progress for us.
3115 if (!mutex_trylock(&oom_lock)) {
3116 *did_some_progress = 1;
3117 schedule_timeout_uninterruptible(1);
3122 * Go through the zonelist yet one more time, keep very high watermark
3123 * here, this is only to catch a parallel oom killing, we must fail if
3124 * we're still under heavy pressure.
3126 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3127 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3131 if (!(gfp_mask & __GFP_NOFAIL)) {
3132 /* Coredumps can quickly deplete all memory reserves */
3133 if (current->flags & PF_DUMPCORE)
3135 /* The OOM killer will not help higher order allocs */
3136 if (order > PAGE_ALLOC_COSTLY_ORDER)
3138 /* The OOM killer does not needlessly kill tasks for lowmem */
3139 if (ac->high_zoneidx < ZONE_NORMAL)
3141 if (pm_suspended_storage())
3144 * XXX: GFP_NOFS allocations should rather fail than rely on
3145 * other request to make a forward progress.
3146 * We are in an unfortunate situation where out_of_memory cannot
3147 * do much for this context but let's try it to at least get
3148 * access to memory reserved if the current task is killed (see
3149 * out_of_memory). Once filesystems are ready to handle allocation
3150 * failures more gracefully we should just bail out here.
3153 /* The OOM killer may not free memory on a specific node */
3154 if (gfp_mask & __GFP_THISNODE)
3157 /* Exhausted what can be done so it's blamo time */
3158 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3159 *did_some_progress = 1;
3161 if (gfp_mask & __GFP_NOFAIL) {
3162 page = get_page_from_freelist(gfp_mask, order,
3163 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3165 * fallback to ignore cpuset restriction if our nodes
3169 page = get_page_from_freelist(gfp_mask, order,
3170 ALLOC_NO_WATERMARKS, ac);
3174 mutex_unlock(&oom_lock);
3180 * Maximum number of compaction retries wit a progress before OOM
3181 * killer is consider as the only way to move forward.
3183 #define MAX_COMPACT_RETRIES 16
3185 #ifdef CONFIG_COMPACTION
3186 /* Try memory compaction for high-order allocations before reclaim */
3187 static struct page *
3188 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3189 unsigned int alloc_flags, const struct alloc_context *ac,
3190 enum migrate_mode mode, enum compact_result *compact_result)
3193 int contended_compaction;
3198 current->flags |= PF_MEMALLOC;
3199 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3200 mode, &contended_compaction);
3201 current->flags &= ~PF_MEMALLOC;
3203 if (*compact_result <= COMPACT_INACTIVE)
3207 * At least in one zone compaction wasn't deferred or skipped, so let's
3208 * count a compaction stall
3210 count_vm_event(COMPACTSTALL);
3212 page = get_page_from_freelist(gfp_mask, order,
3213 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3216 struct zone *zone = page_zone(page);
3218 zone->compact_blockskip_flush = false;
3219 compaction_defer_reset(zone, order, true);
3220 count_vm_event(COMPACTSUCCESS);
3225 * It's bad if compaction run occurs and fails. The most likely reason
3226 * is that pages exist, but not enough to satisfy watermarks.
3228 count_vm_event(COMPACTFAIL);
3231 * In all zones where compaction was attempted (and not
3232 * deferred or skipped), lock contention has been detected.
3233 * For THP allocation we do not want to disrupt the others
3234 * so we fallback to base pages instead.
3236 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3237 *compact_result = COMPACT_CONTENDED;
3240 * If compaction was aborted due to need_resched(), we do not
3241 * want to further increase allocation latency, unless it is
3242 * khugepaged trying to collapse.
3244 if (contended_compaction == COMPACT_CONTENDED_SCHED
3245 && !(current->flags & PF_KTHREAD))
3246 *compact_result = COMPACT_CONTENDED;
3254 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3255 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3256 int compaction_retries)
3258 int max_retries = MAX_COMPACT_RETRIES;
3264 * compaction considers all the zone as desperately out of memory
3265 * so it doesn't really make much sense to retry except when the
3266 * failure could be caused by weak migration mode.
3268 if (compaction_failed(compact_result)) {
3269 if (*migrate_mode == MIGRATE_ASYNC) {
3270 *migrate_mode = MIGRATE_SYNC_LIGHT;
3277 * make sure the compaction wasn't deferred or didn't bail out early
3278 * due to locks contention before we declare that we should give up.
3279 * But do not retry if the given zonelist is not suitable for
3282 if (compaction_withdrawn(compact_result))
3283 return compaction_zonelist_suitable(ac, order, alloc_flags);
3286 * !costly requests are much more important than __GFP_REPEAT
3287 * costly ones because they are de facto nofail and invoke OOM
3288 * killer to move on while costly can fail and users are ready
3289 * to cope with that. 1/4 retries is rather arbitrary but we
3290 * would need much more detailed feedback from compaction to
3291 * make a better decision.
3293 if (order > PAGE_ALLOC_COSTLY_ORDER)
3295 if (compaction_retries <= max_retries)
3301 static inline struct page *
3302 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3303 unsigned int alloc_flags, const struct alloc_context *ac,
3304 enum migrate_mode mode, enum compact_result *compact_result)
3306 *compact_result = COMPACT_SKIPPED;
3311 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3312 enum compact_result compact_result,
3313 enum migrate_mode *migrate_mode,
3314 int compaction_retries)
3319 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3323 * There are setups with compaction disabled which would prefer to loop
3324 * inside the allocator rather than hit the oom killer prematurely.
3325 * Let's give them a good hope and keep retrying while the order-0
3326 * watermarks are OK.
3328 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3330 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3331 ac_classzone_idx(ac), alloc_flags))
3336 #endif /* CONFIG_COMPACTION */
3338 /* Perform direct synchronous page reclaim */
3340 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3341 const struct alloc_context *ac)
3343 struct reclaim_state reclaim_state;
3348 /* We now go into synchronous reclaim */
3349 cpuset_memory_pressure_bump();
3350 current->flags |= PF_MEMALLOC;
3351 lockdep_set_current_reclaim_state(gfp_mask);
3352 reclaim_state.reclaimed_slab = 0;
3353 current->reclaim_state = &reclaim_state;
3355 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3358 current->reclaim_state = NULL;
3359 lockdep_clear_current_reclaim_state();
3360 current->flags &= ~PF_MEMALLOC;
3367 /* The really slow allocator path where we enter direct reclaim */
3368 static inline struct page *
3369 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3370 unsigned int alloc_flags, const struct alloc_context *ac,
3371 unsigned long *did_some_progress)
3373 struct page *page = NULL;
3374 bool drained = false;
3376 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3377 if (unlikely(!(*did_some_progress)))
3381 page = get_page_from_freelist(gfp_mask, order,
3382 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3385 * If an allocation failed after direct reclaim, it could be because
3386 * pages are pinned on the per-cpu lists or in high alloc reserves.
3387 * Shrink them them and try again
3389 if (!page && !drained) {
3390 unreserve_highatomic_pageblock(ac);
3391 drain_all_pages(NULL);
3399 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3404 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3405 ac->high_zoneidx, ac->nodemask)
3406 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3409 static inline unsigned int
3410 gfp_to_alloc_flags(gfp_t gfp_mask)
3412 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3414 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3415 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3418 * The caller may dip into page reserves a bit more if the caller
3419 * cannot run direct reclaim, or if the caller has realtime scheduling
3420 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3421 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3423 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3425 if (gfp_mask & __GFP_ATOMIC) {
3427 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3428 * if it can't schedule.
3430 if (!(gfp_mask & __GFP_NOMEMALLOC))
3431 alloc_flags |= ALLOC_HARDER;
3433 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3434 * comment for __cpuset_node_allowed().
3436 alloc_flags &= ~ALLOC_CPUSET;
3437 } else if (unlikely(rt_task(current)) && !in_interrupt())
3438 alloc_flags |= ALLOC_HARDER;
3440 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3441 if (gfp_mask & __GFP_MEMALLOC)
3442 alloc_flags |= ALLOC_NO_WATERMARKS;
3443 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3444 alloc_flags |= ALLOC_NO_WATERMARKS;
3445 else if (!in_interrupt() &&
3446 ((current->flags & PF_MEMALLOC) ||
3447 unlikely(test_thread_flag(TIF_MEMDIE))))
3448 alloc_flags |= ALLOC_NO_WATERMARKS;
3451 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3452 alloc_flags |= ALLOC_CMA;
3457 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3459 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3462 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3464 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3468 * Maximum number of reclaim retries without any progress before OOM killer
3469 * is consider as the only way to move forward.
3471 #define MAX_RECLAIM_RETRIES 16
3474 * Checks whether it makes sense to retry the reclaim to make a forward progress
3475 * for the given allocation request.
3476 * The reclaim feedback represented by did_some_progress (any progress during
3477 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3478 * any progress in a row) is considered as well as the reclaimable pages on the
3479 * applicable zone list (with a backoff mechanism which is a function of
3480 * no_progress_loops).
3482 * Returns true if a retry is viable or false to enter the oom path.
3485 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3486 struct alloc_context *ac, int alloc_flags,
3487 bool did_some_progress, int no_progress_loops)
3493 * Make sure we converge to OOM if we cannot make any progress
3494 * several times in the row.
3496 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3500 * Keep reclaiming pages while there is a chance this will lead somewhere.
3501 * If none of the target zones can satisfy our allocation request even
3502 * if all reclaimable pages are considered then we are screwed and have
3505 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3507 unsigned long available;
3508 unsigned long reclaimable;
3510 available = reclaimable = zone_reclaimable_pages(zone);
3511 available -= DIV_ROUND_UP(no_progress_loops * available,
3512 MAX_RECLAIM_RETRIES);
3513 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3516 * Would the allocation succeed if we reclaimed the whole
3519 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3520 ac_classzone_idx(ac), alloc_flags, available)) {
3522 * If we didn't make any progress and have a lot of
3523 * dirty + writeback pages then we should wait for
3524 * an IO to complete to slow down the reclaim and
3525 * prevent from pre mature OOM
3527 if (!did_some_progress) {
3528 unsigned long writeback;
3529 unsigned long dirty;
3531 writeback = zone_page_state_snapshot(zone,
3533 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3535 if (2*(writeback + dirty) > reclaimable) {
3536 congestion_wait(BLK_RW_ASYNC, HZ/10);
3542 * Memory allocation/reclaim might be called from a WQ
3543 * context and the current implementation of the WQ
3544 * concurrency control doesn't recognize that
3545 * a particular WQ is congested if the worker thread is
3546 * looping without ever sleeping. Therefore we have to
3547 * do a short sleep here rather than calling
3550 if (current->flags & PF_WQ_WORKER)
3551 schedule_timeout_uninterruptible(1);
3562 static inline struct page *
3563 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3564 struct alloc_context *ac)
3566 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3567 struct page *page = NULL;
3568 unsigned int alloc_flags;
3569 unsigned long did_some_progress;
3570 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3571 enum compact_result compact_result;
3572 int compaction_retries = 0;
3573 int no_progress_loops = 0;
3576 * In the slowpath, we sanity check order to avoid ever trying to
3577 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3578 * be using allocators in order of preference for an area that is
3581 if (order >= MAX_ORDER) {
3582 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3587 * We also sanity check to catch abuse of atomic reserves being used by
3588 * callers that are not in atomic context.
3590 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3591 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3592 gfp_mask &= ~__GFP_ATOMIC;
3595 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3596 wake_all_kswapds(order, ac);
3599 * OK, we're below the kswapd watermark and have kicked background
3600 * reclaim. Now things get more complex, so set up alloc_flags according
3601 * to how we want to proceed.
3603 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3605 /* This is the last chance, in general, before the goto nopage. */
3606 page = get_page_from_freelist(gfp_mask, order,
3607 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3611 /* Allocate without watermarks if the context allows */
3612 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3614 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3615 * the allocation is high priority and these type of
3616 * allocations are system rather than user orientated
3618 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3619 page = get_page_from_freelist(gfp_mask, order,
3620 ALLOC_NO_WATERMARKS, ac);
3625 /* Caller is not willing to reclaim, we can't balance anything */
3626 if (!can_direct_reclaim) {
3628 * All existing users of the __GFP_NOFAIL are blockable, so warn
3629 * of any new users that actually allow this type of allocation
3632 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3636 /* Avoid recursion of direct reclaim */
3637 if (current->flags & PF_MEMALLOC) {
3639 * __GFP_NOFAIL request from this context is rather bizarre
3640 * because we cannot reclaim anything and only can loop waiting
3641 * for somebody to do a work for us.
3643 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3650 /* Avoid allocations with no watermarks from looping endlessly */
3651 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3655 * Try direct compaction. The first pass is asynchronous. Subsequent
3656 * attempts after direct reclaim are synchronous
3658 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3664 /* Checks for THP-specific high-order allocations */
3665 if (is_thp_gfp_mask(gfp_mask)) {
3667 * If compaction is deferred for high-order allocations, it is
3668 * because sync compaction recently failed. If this is the case
3669 * and the caller requested a THP allocation, we do not want
3670 * to heavily disrupt the system, so we fail the allocation
3671 * instead of entering direct reclaim.
3673 if (compact_result == COMPACT_DEFERRED)
3677 * Compaction is contended so rather back off than cause
3680 if(compact_result == COMPACT_CONTENDED)
3684 if (order && compaction_made_progress(compact_result))
3685 compaction_retries++;
3687 /* Try direct reclaim and then allocating */
3688 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3689 &did_some_progress);
3693 /* Do not loop if specifically requested */
3694 if (gfp_mask & __GFP_NORETRY)
3698 * Do not retry costly high order allocations unless they are
3701 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3705 * Costly allocations might have made a progress but this doesn't mean
3706 * their order will become available due to high fragmentation so
3707 * always increment the no progress counter for them
3709 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3710 no_progress_loops = 0;
3712 no_progress_loops++;
3714 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3715 did_some_progress > 0, no_progress_loops))
3719 * It doesn't make any sense to retry for the compaction if the order-0
3720 * reclaim is not able to make any progress because the current
3721 * implementation of the compaction depends on the sufficient amount
3722 * of free memory (see __compaction_suitable)
3724 if (did_some_progress > 0 &&
3725 should_compact_retry(ac, order, alloc_flags,
3726 compact_result, &migration_mode,
3727 compaction_retries))
3730 /* Reclaim has failed us, start killing things */
3731 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3735 /* Retry as long as the OOM killer is making progress */
3736 if (did_some_progress) {
3737 no_progress_loops = 0;
3743 * High-order allocations do not necessarily loop after direct reclaim
3744 * and reclaim/compaction depends on compaction being called after
3745 * reclaim so call directly if necessary.
3746 * It can become very expensive to allocate transparent hugepages at
3747 * fault, so use asynchronous memory compaction for THP unless it is
3748 * khugepaged trying to collapse. All other requests should tolerate
3749 * at least light sync migration.
3751 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3752 migration_mode = MIGRATE_ASYNC;
3754 migration_mode = MIGRATE_SYNC_LIGHT;
3755 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3761 warn_alloc_failed(gfp_mask, order, NULL);
3767 * This is the 'heart' of the zoned buddy allocator.
3770 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3771 struct zonelist *zonelist, nodemask_t *nodemask)
3774 unsigned int cpuset_mems_cookie;
3775 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3776 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3777 struct alloc_context ac = {
3778 .high_zoneidx = gfp_zone(gfp_mask),
3779 .zonelist = zonelist,
3780 .nodemask = nodemask,
3781 .migratetype = gfpflags_to_migratetype(gfp_mask),
3784 if (cpusets_enabled()) {
3785 alloc_mask |= __GFP_HARDWALL;
3786 alloc_flags |= ALLOC_CPUSET;
3788 ac.nodemask = &cpuset_current_mems_allowed;
3791 gfp_mask &= gfp_allowed_mask;
3793 lockdep_trace_alloc(gfp_mask);
3795 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3797 if (should_fail_alloc_page(gfp_mask, order))
3801 * Check the zones suitable for the gfp_mask contain at least one
3802 * valid zone. It's possible to have an empty zonelist as a result
3803 * of __GFP_THISNODE and a memoryless node
3805 if (unlikely(!zonelist->_zonerefs->zone))
3808 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3809 alloc_flags |= ALLOC_CMA;
3812 cpuset_mems_cookie = read_mems_allowed_begin();
3814 /* Dirty zone balancing only done in the fast path */
3815 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3817 /* The preferred zone is used for statistics later */
3818 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3819 ac.high_zoneidx, ac.nodemask);
3820 if (!ac.preferred_zoneref) {
3825 /* First allocation attempt */
3826 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3831 * Runtime PM, block IO and its error handling path can deadlock
3832 * because I/O on the device might not complete.
3834 alloc_mask = memalloc_noio_flags(gfp_mask);
3835 ac.spread_dirty_pages = false;
3838 * Restore the original nodemask if it was potentially replaced with
3839 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3841 if (cpusets_enabled())
3842 ac.nodemask = nodemask;
3843 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3847 * When updating a task's mems_allowed, it is possible to race with
3848 * parallel threads in such a way that an allocation can fail while
3849 * the mask is being updated. If a page allocation is about to fail,
3850 * check if the cpuset changed during allocation and if so, retry.
3852 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3853 alloc_mask = gfp_mask;
3858 if (kmemcheck_enabled && page)
3859 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3861 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3865 EXPORT_SYMBOL(__alloc_pages_nodemask);
3868 * Common helper functions.
3870 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3875 * __get_free_pages() returns a 32-bit address, which cannot represent
3878 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3880 page = alloc_pages(gfp_mask, order);
3883 return (unsigned long) page_address(page);
3885 EXPORT_SYMBOL(__get_free_pages);
3887 unsigned long get_zeroed_page(gfp_t gfp_mask)
3889 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3891 EXPORT_SYMBOL(get_zeroed_page);
3893 void __free_pages(struct page *page, unsigned int order)
3895 if (put_page_testzero(page)) {
3897 free_hot_cold_page(page, false);
3899 __free_pages_ok(page, order);
3903 EXPORT_SYMBOL(__free_pages);
3905 void free_pages(unsigned long addr, unsigned int order)
3908 VM_BUG_ON(!virt_addr_valid((void *)addr));
3909 __free_pages(virt_to_page((void *)addr), order);
3913 EXPORT_SYMBOL(free_pages);
3917 * An arbitrary-length arbitrary-offset area of memory which resides
3918 * within a 0 or higher order page. Multiple fragments within that page
3919 * are individually refcounted, in the page's reference counter.
3921 * The page_frag functions below provide a simple allocation framework for
3922 * page fragments. This is used by the network stack and network device
3923 * drivers to provide a backing region of memory for use as either an
3924 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3926 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3929 struct page *page = NULL;
3930 gfp_t gfp = gfp_mask;
3932 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3933 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3935 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3936 PAGE_FRAG_CACHE_MAX_ORDER);
3937 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3939 if (unlikely(!page))
3940 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3942 nc->va = page ? page_address(page) : NULL;
3947 void *__alloc_page_frag(struct page_frag_cache *nc,
3948 unsigned int fragsz, gfp_t gfp_mask)
3950 unsigned int size = PAGE_SIZE;
3954 if (unlikely(!nc->va)) {
3956 page = __page_frag_refill(nc, gfp_mask);
3960 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3961 /* if size can vary use size else just use PAGE_SIZE */
3964 /* Even if we own the page, we do not use atomic_set().
3965 * This would break get_page_unless_zero() users.
3967 page_ref_add(page, size - 1);
3969 /* reset page count bias and offset to start of new frag */
3970 nc->pfmemalloc = page_is_pfmemalloc(page);
3971 nc->pagecnt_bias = size;
3975 offset = nc->offset - fragsz;
3976 if (unlikely(offset < 0)) {
3977 page = virt_to_page(nc->va);
3979 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3982 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3983 /* if size can vary use size else just use PAGE_SIZE */
3986 /* OK, page count is 0, we can safely set it */
3987 set_page_count(page, size);
3989 /* reset page count bias and offset to start of new frag */
3990 nc->pagecnt_bias = size;
3991 offset = size - fragsz;
3995 nc->offset = offset;
3997 return nc->va + offset;
3999 EXPORT_SYMBOL(__alloc_page_frag);
4002 * Frees a page fragment allocated out of either a compound or order 0 page.
4004 void __free_page_frag(void *addr)
4006 struct page *page = virt_to_head_page(addr);
4008 if (unlikely(put_page_testzero(page)))
4009 __free_pages_ok(page, compound_order(page));
4011 EXPORT_SYMBOL(__free_page_frag);
4014 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
4015 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
4016 * equivalent to alloc_pages.
4018 * It should be used when the caller would like to use kmalloc, but since the
4019 * allocation is large, it has to fall back to the page allocator.
4021 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
4025 page = alloc_pages(gfp_mask, order);
4026 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4027 __free_pages(page, order);
4033 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
4037 page = alloc_pages_node(nid, gfp_mask, order);
4038 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4039 __free_pages(page, order);
4046 * __free_kmem_pages and free_kmem_pages will free pages allocated with
4049 void __free_kmem_pages(struct page *page, unsigned int order)
4051 memcg_kmem_uncharge(page, order);
4052 __free_pages(page, order);
4055 void free_kmem_pages(unsigned long addr, unsigned int order)
4058 VM_BUG_ON(!virt_addr_valid((void *)addr));
4059 __free_kmem_pages(virt_to_page((void *)addr), order);
4063 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4067 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4068 unsigned long used = addr + PAGE_ALIGN(size);
4070 split_page(virt_to_page((void *)addr), order);
4071 while (used < alloc_end) {
4076 return (void *)addr;
4080 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4081 * @size: the number of bytes to allocate
4082 * @gfp_mask: GFP flags for the allocation
4084 * This function is similar to alloc_pages(), except that it allocates the
4085 * minimum number of pages to satisfy the request. alloc_pages() can only
4086 * allocate memory in power-of-two pages.
4088 * This function is also limited by MAX_ORDER.
4090 * Memory allocated by this function must be released by free_pages_exact().
4092 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4094 unsigned int order = get_order(size);
4097 addr = __get_free_pages(gfp_mask, order);
4098 return make_alloc_exact(addr, order, size);
4100 EXPORT_SYMBOL(alloc_pages_exact);
4103 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4105 * @nid: the preferred node ID where memory should be allocated
4106 * @size: the number of bytes to allocate
4107 * @gfp_mask: GFP flags for the allocation
4109 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4112 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4114 unsigned int order = get_order(size);
4115 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4118 return make_alloc_exact((unsigned long)page_address(p), order, size);
4122 * free_pages_exact - release memory allocated via alloc_pages_exact()
4123 * @virt: the value returned by alloc_pages_exact.
4124 * @size: size of allocation, same value as passed to alloc_pages_exact().
4126 * Release the memory allocated by a previous call to alloc_pages_exact.
4128 void free_pages_exact(void *virt, size_t size)
4130 unsigned long addr = (unsigned long)virt;
4131 unsigned long end = addr + PAGE_ALIGN(size);
4133 while (addr < end) {
4138 EXPORT_SYMBOL(free_pages_exact);
4141 * nr_free_zone_pages - count number of pages beyond high watermark
4142 * @offset: The zone index of the highest zone
4144 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4145 * high watermark within all zones at or below a given zone index. For each
4146 * zone, the number of pages is calculated as:
4147 * managed_pages - high_pages
4149 static unsigned long nr_free_zone_pages(int offset)
4154 /* Just pick one node, since fallback list is circular */
4155 unsigned long sum = 0;
4157 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4159 for_each_zone_zonelist(zone, z, zonelist, offset) {
4160 unsigned long size = zone->managed_pages;
4161 unsigned long high = high_wmark_pages(zone);
4170 * nr_free_buffer_pages - count number of pages beyond high watermark
4172 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4173 * watermark within ZONE_DMA and ZONE_NORMAL.
4175 unsigned long nr_free_buffer_pages(void)
4177 return nr_free_zone_pages(gfp_zone(GFP_USER));
4179 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4182 * nr_free_pagecache_pages - count number of pages beyond high watermark
4184 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4185 * high watermark within all zones.
4187 unsigned long nr_free_pagecache_pages(void)
4189 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4192 static inline void show_node(struct zone *zone)
4194 if (IS_ENABLED(CONFIG_NUMA))
4195 printk("Node %d ", zone_to_nid(zone));
4198 long si_mem_available(void)
4201 unsigned long pagecache;
4202 unsigned long wmark_low = 0;
4203 unsigned long pages[NR_LRU_LISTS];
4207 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4208 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4211 wmark_low += zone->watermark[WMARK_LOW];
4214 * Estimate the amount of memory available for userspace allocations,
4215 * without causing swapping.
4217 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4220 * Not all the page cache can be freed, otherwise the system will
4221 * start swapping. Assume at least half of the page cache, or the
4222 * low watermark worth of cache, needs to stay.
4224 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4225 pagecache -= min(pagecache / 2, wmark_low);
4226 available += pagecache;
4229 * Part of the reclaimable slab consists of items that are in use,
4230 * and cannot be freed. Cap this estimate at the low watermark.
4232 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4233 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4239 EXPORT_SYMBOL_GPL(si_mem_available);
4241 void si_meminfo(struct sysinfo *val)
4243 val->totalram = totalram_pages;
4244 val->sharedram = global_page_state(NR_SHMEM);
4245 val->freeram = global_page_state(NR_FREE_PAGES);
4246 val->bufferram = nr_blockdev_pages();
4247 val->totalhigh = totalhigh_pages;
4248 val->freehigh = nr_free_highpages();
4249 val->mem_unit = PAGE_SIZE;
4252 EXPORT_SYMBOL(si_meminfo);
4255 void si_meminfo_node(struct sysinfo *val, int nid)
4257 int zone_type; /* needs to be signed */
4258 unsigned long managed_pages = 0;
4259 unsigned long managed_highpages = 0;
4260 unsigned long free_highpages = 0;
4261 pg_data_t *pgdat = NODE_DATA(nid);
4263 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4264 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4265 val->totalram = managed_pages;
4266 val->sharedram = node_page_state(nid, NR_SHMEM);
4267 val->freeram = node_page_state(nid, NR_FREE_PAGES);
4268 #ifdef CONFIG_HIGHMEM
4269 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4270 struct zone *zone = &pgdat->node_zones[zone_type];
4272 if (is_highmem(zone)) {
4273 managed_highpages += zone->managed_pages;
4274 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4277 val->totalhigh = managed_highpages;
4278 val->freehigh = free_highpages;
4280 val->totalhigh = managed_highpages;
4281 val->freehigh = free_highpages;
4283 val->mem_unit = PAGE_SIZE;
4288 * Determine whether the node should be displayed or not, depending on whether
4289 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4291 bool skip_free_areas_node(unsigned int flags, int nid)
4294 unsigned int cpuset_mems_cookie;
4296 if (!(flags & SHOW_MEM_FILTER_NODES))
4300 cpuset_mems_cookie = read_mems_allowed_begin();
4301 ret = !node_isset(nid, cpuset_current_mems_allowed);
4302 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4307 #define K(x) ((x) << (PAGE_SHIFT-10))
4309 static void show_migration_types(unsigned char type)
4311 static const char types[MIGRATE_TYPES] = {
4312 [MIGRATE_UNMOVABLE] = 'U',
4313 [MIGRATE_MOVABLE] = 'M',
4314 [MIGRATE_RECLAIMABLE] = 'E',
4315 [MIGRATE_HIGHATOMIC] = 'H',
4317 [MIGRATE_CMA] = 'C',
4319 #ifdef CONFIG_MEMORY_ISOLATION
4320 [MIGRATE_ISOLATE] = 'I',
4323 char tmp[MIGRATE_TYPES + 1];
4327 for (i = 0; i < MIGRATE_TYPES; i++) {
4328 if (type & (1 << i))
4333 printk("(%s) ", tmp);
4337 * Show free area list (used inside shift_scroll-lock stuff)
4338 * We also calculate the percentage fragmentation. We do this by counting the
4339 * memory on each free list with the exception of the first item on the list.
4342 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4345 void show_free_areas(unsigned int filter)
4347 unsigned long free_pcp = 0;
4351 for_each_populated_zone(zone) {
4352 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4355 for_each_online_cpu(cpu)
4356 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4359 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4360 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4361 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4362 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4363 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4364 " free:%lu free_pcp:%lu free_cma:%lu\n",
4365 global_page_state(NR_ACTIVE_ANON),
4366 global_page_state(NR_INACTIVE_ANON),
4367 global_page_state(NR_ISOLATED_ANON),
4368 global_page_state(NR_ACTIVE_FILE),
4369 global_page_state(NR_INACTIVE_FILE),
4370 global_page_state(NR_ISOLATED_FILE),
4371 global_page_state(NR_UNEVICTABLE),
4372 global_page_state(NR_FILE_DIRTY),
4373 global_page_state(NR_WRITEBACK),
4374 global_page_state(NR_UNSTABLE_NFS),
4375 global_page_state(NR_SLAB_RECLAIMABLE),
4376 global_page_state(NR_SLAB_UNRECLAIMABLE),
4377 global_page_state(NR_FILE_MAPPED),
4378 global_page_state(NR_SHMEM),
4379 global_page_state(NR_PAGETABLE),
4380 global_page_state(NR_BOUNCE),
4381 global_page_state(NR_FREE_PAGES),
4383 global_page_state(NR_FREE_CMA_PAGES));
4385 for_each_populated_zone(zone) {
4388 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4392 for_each_online_cpu(cpu)
4393 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4401 " active_anon:%lukB"
4402 " inactive_anon:%lukB"
4403 " active_file:%lukB"
4404 " inactive_file:%lukB"
4405 " unevictable:%lukB"
4406 " isolated(anon):%lukB"
4407 " isolated(file):%lukB"
4415 " slab_reclaimable:%lukB"
4416 " slab_unreclaimable:%lukB"
4417 " kernel_stack:%lukB"
4424 " writeback_tmp:%lukB"
4425 " pages_scanned:%lu"
4426 " all_unreclaimable? %s"
4429 K(zone_page_state(zone, NR_FREE_PAGES)),
4430 K(min_wmark_pages(zone)),
4431 K(low_wmark_pages(zone)),
4432 K(high_wmark_pages(zone)),
4433 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4434 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4435 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4436 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4437 K(zone_page_state(zone, NR_UNEVICTABLE)),
4438 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4439 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4440 K(zone->present_pages),
4441 K(zone->managed_pages),
4442 K(zone_page_state(zone, NR_MLOCK)),
4443 K(zone_page_state(zone, NR_FILE_DIRTY)),
4444 K(zone_page_state(zone, NR_WRITEBACK)),
4445 K(zone_page_state(zone, NR_FILE_MAPPED)),
4446 K(zone_page_state(zone, NR_SHMEM)),
4447 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4448 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4449 zone_page_state(zone, NR_KERNEL_STACK) *
4451 K(zone_page_state(zone, NR_PAGETABLE)),
4452 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4453 K(zone_page_state(zone, NR_BOUNCE)),
4455 K(this_cpu_read(zone->pageset->pcp.count)),
4456 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4457 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4458 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4459 (!zone_reclaimable(zone) ? "yes" : "no")
4461 printk("lowmem_reserve[]:");
4462 for (i = 0; i < MAX_NR_ZONES; i++)
4463 printk(" %ld", zone->lowmem_reserve[i]);
4467 for_each_populated_zone(zone) {
4469 unsigned long nr[MAX_ORDER], flags, total = 0;
4470 unsigned char types[MAX_ORDER];
4472 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4475 printk("%s: ", zone->name);
4477 spin_lock_irqsave(&zone->lock, flags);
4478 for (order = 0; order < MAX_ORDER; order++) {
4479 struct free_area *area = &zone->free_area[order];
4482 nr[order] = area->nr_free;
4483 total += nr[order] << order;
4486 for (type = 0; type < MIGRATE_TYPES; type++) {
4487 if (!list_empty(&area->free_list[type]))
4488 types[order] |= 1 << type;
4491 spin_unlock_irqrestore(&zone->lock, flags);
4492 for (order = 0; order < MAX_ORDER; order++) {
4493 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4495 show_migration_types(types[order]);
4497 printk("= %lukB\n", K(total));
4500 hugetlb_show_meminfo();
4502 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4504 show_swap_cache_info();
4507 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4509 zoneref->zone = zone;
4510 zoneref->zone_idx = zone_idx(zone);
4514 * Builds allocation fallback zone lists.
4516 * Add all populated zones of a node to the zonelist.
4518 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4522 enum zone_type zone_type = MAX_NR_ZONES;
4526 zone = pgdat->node_zones + zone_type;
4527 if (populated_zone(zone)) {
4528 zoneref_set_zone(zone,
4529 &zonelist->_zonerefs[nr_zones++]);
4530 check_highest_zone(zone_type);
4532 } while (zone_type);
4540 * 0 = automatic detection of better ordering.
4541 * 1 = order by ([node] distance, -zonetype)
4542 * 2 = order by (-zonetype, [node] distance)
4544 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4545 * the same zonelist. So only NUMA can configure this param.
4547 #define ZONELIST_ORDER_DEFAULT 0
4548 #define ZONELIST_ORDER_NODE 1
4549 #define ZONELIST_ORDER_ZONE 2
4551 /* zonelist order in the kernel.
4552 * set_zonelist_order() will set this to NODE or ZONE.
4554 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4555 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4559 /* The value user specified ....changed by config */
4560 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4561 /* string for sysctl */
4562 #define NUMA_ZONELIST_ORDER_LEN 16
4563 char numa_zonelist_order[16] = "default";
4566 * interface for configure zonelist ordering.
4567 * command line option "numa_zonelist_order"
4568 * = "[dD]efault - default, automatic configuration.
4569 * = "[nN]ode - order by node locality, then by zone within node
4570 * = "[zZ]one - order by zone, then by locality within zone
4573 static int __parse_numa_zonelist_order(char *s)
4575 if (*s == 'd' || *s == 'D') {
4576 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4577 } else if (*s == 'n' || *s == 'N') {
4578 user_zonelist_order = ZONELIST_ORDER_NODE;
4579 } else if (*s == 'z' || *s == 'Z') {
4580 user_zonelist_order = ZONELIST_ORDER_ZONE;
4582 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4588 static __init int setup_numa_zonelist_order(char *s)
4595 ret = __parse_numa_zonelist_order(s);
4597 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4601 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4604 * sysctl handler for numa_zonelist_order
4606 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4607 void __user *buffer, size_t *length,
4610 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4612 static DEFINE_MUTEX(zl_order_mutex);
4614 mutex_lock(&zl_order_mutex);
4616 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4620 strcpy(saved_string, (char *)table->data);
4622 ret = proc_dostring(table, write, buffer, length, ppos);
4626 int oldval = user_zonelist_order;
4628 ret = __parse_numa_zonelist_order((char *)table->data);
4631 * bogus value. restore saved string
4633 strncpy((char *)table->data, saved_string,
4634 NUMA_ZONELIST_ORDER_LEN);
4635 user_zonelist_order = oldval;
4636 } else if (oldval != user_zonelist_order) {
4637 mutex_lock(&zonelists_mutex);
4638 build_all_zonelists(NULL, NULL);
4639 mutex_unlock(&zonelists_mutex);
4643 mutex_unlock(&zl_order_mutex);
4648 #define MAX_NODE_LOAD (nr_online_nodes)
4649 static int node_load[MAX_NUMNODES];
4652 * find_next_best_node - find the next node that should appear in a given node's fallback list
4653 * @node: node whose fallback list we're appending
4654 * @used_node_mask: nodemask_t of already used nodes
4656 * We use a number of factors to determine which is the next node that should
4657 * appear on a given node's fallback list. The node should not have appeared
4658 * already in @node's fallback list, and it should be the next closest node
4659 * according to the distance array (which contains arbitrary distance values
4660 * from each node to each node in the system), and should also prefer nodes
4661 * with no CPUs, since presumably they'll have very little allocation pressure
4662 * on them otherwise.
4663 * It returns -1 if no node is found.
4665 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4668 int min_val = INT_MAX;
4669 int best_node = NUMA_NO_NODE;
4670 const struct cpumask *tmp = cpumask_of_node(0);
4672 /* Use the local node if we haven't already */
4673 if (!node_isset(node, *used_node_mask)) {
4674 node_set(node, *used_node_mask);
4678 for_each_node_state(n, N_MEMORY) {
4680 /* Don't want a node to appear more than once */
4681 if (node_isset(n, *used_node_mask))
4684 /* Use the distance array to find the distance */
4685 val = node_distance(node, n);
4687 /* Penalize nodes under us ("prefer the next node") */
4690 /* Give preference to headless and unused nodes */
4691 tmp = cpumask_of_node(n);
4692 if (!cpumask_empty(tmp))
4693 val += PENALTY_FOR_NODE_WITH_CPUS;
4695 /* Slight preference for less loaded node */
4696 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4697 val += node_load[n];
4699 if (val < min_val) {
4706 node_set(best_node, *used_node_mask);
4713 * Build zonelists ordered by node and zones within node.
4714 * This results in maximum locality--normal zone overflows into local
4715 * DMA zone, if any--but risks exhausting DMA zone.
4717 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4720 struct zonelist *zonelist;
4722 zonelist = &pgdat->node_zonelists[0];
4723 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4725 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4726 zonelist->_zonerefs[j].zone = NULL;
4727 zonelist->_zonerefs[j].zone_idx = 0;
4731 * Build gfp_thisnode zonelists
4733 static void build_thisnode_zonelists(pg_data_t *pgdat)
4736 struct zonelist *zonelist;
4738 zonelist = &pgdat->node_zonelists[1];
4739 j = build_zonelists_node(pgdat, zonelist, 0);
4740 zonelist->_zonerefs[j].zone = NULL;
4741 zonelist->_zonerefs[j].zone_idx = 0;
4745 * Build zonelists ordered by zone and nodes within zones.
4746 * This results in conserving DMA zone[s] until all Normal memory is
4747 * exhausted, but results in overflowing to remote node while memory
4748 * may still exist in local DMA zone.
4750 static int node_order[MAX_NUMNODES];
4752 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4755 int zone_type; /* needs to be signed */
4757 struct zonelist *zonelist;
4759 zonelist = &pgdat->node_zonelists[0];
4761 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4762 for (j = 0; j < nr_nodes; j++) {
4763 node = node_order[j];
4764 z = &NODE_DATA(node)->node_zones[zone_type];
4765 if (populated_zone(z)) {
4767 &zonelist->_zonerefs[pos++]);
4768 check_highest_zone(zone_type);
4772 zonelist->_zonerefs[pos].zone = NULL;
4773 zonelist->_zonerefs[pos].zone_idx = 0;
4776 #if defined(CONFIG_64BIT)
4778 * Devices that require DMA32/DMA are relatively rare and do not justify a
4779 * penalty to every machine in case the specialised case applies. Default
4780 * to Node-ordering on 64-bit NUMA machines
4782 static int default_zonelist_order(void)
4784 return ZONELIST_ORDER_NODE;
4788 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4789 * by the kernel. If processes running on node 0 deplete the low memory zone
4790 * then reclaim will occur more frequency increasing stalls and potentially
4791 * be easier to OOM if a large percentage of the zone is under writeback or
4792 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4793 * Hence, default to zone ordering on 32-bit.
4795 static int default_zonelist_order(void)
4797 return ZONELIST_ORDER_ZONE;
4799 #endif /* CONFIG_64BIT */
4801 static void set_zonelist_order(void)
4803 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4804 current_zonelist_order = default_zonelist_order();
4806 current_zonelist_order = user_zonelist_order;
4809 static void build_zonelists(pg_data_t *pgdat)
4812 nodemask_t used_mask;
4813 int local_node, prev_node;
4814 struct zonelist *zonelist;
4815 unsigned int order = current_zonelist_order;
4817 /* initialize zonelists */
4818 for (i = 0; i < MAX_ZONELISTS; i++) {
4819 zonelist = pgdat->node_zonelists + i;
4820 zonelist->_zonerefs[0].zone = NULL;
4821 zonelist->_zonerefs[0].zone_idx = 0;
4824 /* NUMA-aware ordering of nodes */
4825 local_node = pgdat->node_id;
4826 load = nr_online_nodes;
4827 prev_node = local_node;
4828 nodes_clear(used_mask);
4830 memset(node_order, 0, sizeof(node_order));
4833 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4835 * We don't want to pressure a particular node.
4836 * So adding penalty to the first node in same
4837 * distance group to make it round-robin.
4839 if (node_distance(local_node, node) !=
4840 node_distance(local_node, prev_node))
4841 node_load[node] = load;
4845 if (order == ZONELIST_ORDER_NODE)
4846 build_zonelists_in_node_order(pgdat, node);
4848 node_order[i++] = node; /* remember order */
4851 if (order == ZONELIST_ORDER_ZONE) {
4852 /* calculate node order -- i.e., DMA last! */
4853 build_zonelists_in_zone_order(pgdat, i);
4856 build_thisnode_zonelists(pgdat);
4859 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4861 * Return node id of node used for "local" allocations.
4862 * I.e., first node id of first zone in arg node's generic zonelist.
4863 * Used for initializing percpu 'numa_mem', which is used primarily
4864 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4866 int local_memory_node(int node)
4870 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4871 gfp_zone(GFP_KERNEL),
4873 return z->zone->node;
4877 #else /* CONFIG_NUMA */
4879 static void set_zonelist_order(void)
4881 current_zonelist_order = ZONELIST_ORDER_ZONE;
4884 static void build_zonelists(pg_data_t *pgdat)
4886 int node, local_node;
4888 struct zonelist *zonelist;
4890 local_node = pgdat->node_id;
4892 zonelist = &pgdat->node_zonelists[0];
4893 j = build_zonelists_node(pgdat, zonelist, 0);
4896 * Now we build the zonelist so that it contains the zones
4897 * of all the other nodes.
4898 * We don't want to pressure a particular node, so when
4899 * building the zones for node N, we make sure that the
4900 * zones coming right after the local ones are those from
4901 * node N+1 (modulo N)
4903 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4904 if (!node_online(node))
4906 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4908 for (node = 0; node < local_node; node++) {
4909 if (!node_online(node))
4911 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4914 zonelist->_zonerefs[j].zone = NULL;
4915 zonelist->_zonerefs[j].zone_idx = 0;
4918 #endif /* CONFIG_NUMA */
4921 * Boot pageset table. One per cpu which is going to be used for all
4922 * zones and all nodes. The parameters will be set in such a way
4923 * that an item put on a list will immediately be handed over to
4924 * the buddy list. This is safe since pageset manipulation is done
4925 * with interrupts disabled.
4927 * The boot_pagesets must be kept even after bootup is complete for
4928 * unused processors and/or zones. They do play a role for bootstrapping
4929 * hotplugged processors.
4931 * zoneinfo_show() and maybe other functions do
4932 * not check if the processor is online before following the pageset pointer.
4933 * Other parts of the kernel may not check if the zone is available.
4935 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4936 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4937 static void setup_zone_pageset(struct zone *zone);
4940 * Global mutex to protect against size modification of zonelists
4941 * as well as to serialize pageset setup for the new populated zone.
4943 DEFINE_MUTEX(zonelists_mutex);
4945 /* return values int ....just for stop_machine() */
4946 static int __build_all_zonelists(void *data)
4950 pg_data_t *self = data;
4953 memset(node_load, 0, sizeof(node_load));
4956 if (self && !node_online(self->node_id)) {
4957 build_zonelists(self);
4960 for_each_online_node(nid) {
4961 pg_data_t *pgdat = NODE_DATA(nid);
4963 build_zonelists(pgdat);
4967 * Initialize the boot_pagesets that are going to be used
4968 * for bootstrapping processors. The real pagesets for
4969 * each zone will be allocated later when the per cpu
4970 * allocator is available.
4972 * boot_pagesets are used also for bootstrapping offline
4973 * cpus if the system is already booted because the pagesets
4974 * are needed to initialize allocators on a specific cpu too.
4975 * F.e. the percpu allocator needs the page allocator which
4976 * needs the percpu allocator in order to allocate its pagesets
4977 * (a chicken-egg dilemma).
4979 for_each_possible_cpu(cpu) {
4980 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4982 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4984 * We now know the "local memory node" for each node--
4985 * i.e., the node of the first zone in the generic zonelist.
4986 * Set up numa_mem percpu variable for on-line cpus. During
4987 * boot, only the boot cpu should be on-line; we'll init the
4988 * secondary cpus' numa_mem as they come on-line. During
4989 * node/memory hotplug, we'll fixup all on-line cpus.
4991 if (cpu_online(cpu))
4992 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4999 static noinline void __init
5000 build_all_zonelists_init(void)
5002 __build_all_zonelists(NULL);
5003 mminit_verify_zonelist();
5004 cpuset_init_current_mems_allowed();
5008 * Called with zonelists_mutex held always
5009 * unless system_state == SYSTEM_BOOTING.
5011 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5012 * [we're only called with non-NULL zone through __meminit paths] and
5013 * (2) call of __init annotated helper build_all_zonelists_init
5014 * [protected by SYSTEM_BOOTING].
5016 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5018 set_zonelist_order();
5020 if (system_state == SYSTEM_BOOTING) {
5021 build_all_zonelists_init();
5023 #ifdef CONFIG_MEMORY_HOTPLUG
5025 setup_zone_pageset(zone);
5027 /* we have to stop all cpus to guarantee there is no user
5029 stop_machine(__build_all_zonelists, pgdat, NULL);
5030 /* cpuset refresh routine should be here */
5032 vm_total_pages = nr_free_pagecache_pages();
5034 * Disable grouping by mobility if the number of pages in the
5035 * system is too low to allow the mechanism to work. It would be
5036 * more accurate, but expensive to check per-zone. This check is
5037 * made on memory-hotadd so a system can start with mobility
5038 * disabled and enable it later
5040 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5041 page_group_by_mobility_disabled = 1;
5043 page_group_by_mobility_disabled = 0;
5045 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5047 zonelist_order_name[current_zonelist_order],
5048 page_group_by_mobility_disabled ? "off" : "on",
5051 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5056 * Helper functions to size the waitqueue hash table.
5057 * Essentially these want to choose hash table sizes sufficiently
5058 * large so that collisions trying to wait on pages are rare.
5059 * But in fact, the number of active page waitqueues on typical
5060 * systems is ridiculously low, less than 200. So this is even
5061 * conservative, even though it seems large.
5063 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5064 * waitqueues, i.e. the size of the waitq table given the number of pages.
5066 #define PAGES_PER_WAITQUEUE 256
5068 #ifndef CONFIG_MEMORY_HOTPLUG
5069 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5071 unsigned long size = 1;
5073 pages /= PAGES_PER_WAITQUEUE;
5075 while (size < pages)
5079 * Once we have dozens or even hundreds of threads sleeping
5080 * on IO we've got bigger problems than wait queue collision.
5081 * Limit the size of the wait table to a reasonable size.
5083 size = min(size, 4096UL);
5085 return max(size, 4UL);
5089 * A zone's size might be changed by hot-add, so it is not possible to determine
5090 * a suitable size for its wait_table. So we use the maximum size now.
5092 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5094 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5095 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5096 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5098 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5099 * or more by the traditional way. (See above). It equals:
5101 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5102 * ia64(16K page size) : = ( 8G + 4M)byte.
5103 * powerpc (64K page size) : = (32G +16M)byte.
5105 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5112 * This is an integer logarithm so that shifts can be used later
5113 * to extract the more random high bits from the multiplicative
5114 * hash function before the remainder is taken.
5116 static inline unsigned long wait_table_bits(unsigned long size)
5122 * Initially all pages are reserved - free ones are freed
5123 * up by free_all_bootmem() once the early boot process is
5124 * done. Non-atomic initialization, single-pass.
5126 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5127 unsigned long start_pfn, enum memmap_context context)
5129 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5130 unsigned long end_pfn = start_pfn + size;
5131 pg_data_t *pgdat = NODE_DATA(nid);
5133 unsigned long nr_initialised = 0;
5134 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5135 struct memblock_region *r = NULL, *tmp;
5138 if (highest_memmap_pfn < end_pfn - 1)
5139 highest_memmap_pfn = end_pfn - 1;
5142 * Honor reservation requested by the driver for this ZONE_DEVICE
5145 if (altmap && start_pfn == altmap->base_pfn)
5146 start_pfn += altmap->reserve;
5148 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5150 * There can be holes in boot-time mem_map[]s handed to this
5151 * function. They do not exist on hotplugged memory.
5153 if (context != MEMMAP_EARLY)
5156 if (!early_pfn_valid(pfn))
5158 if (!early_pfn_in_nid(pfn, nid))
5160 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5163 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5165 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5166 * from zone_movable_pfn[nid] to end of each node should be
5167 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5169 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5170 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5174 * Check given memblock attribute by firmware which can affect
5175 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5176 * mirrored, it's an overlapped memmap init. skip it.
5178 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5179 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5180 for_each_memblock(memory, tmp)
5181 if (pfn < memblock_region_memory_end_pfn(tmp))
5185 if (pfn >= memblock_region_memory_base_pfn(r) &&
5186 memblock_is_mirror(r)) {
5187 /* already initialized as NORMAL */
5188 pfn = memblock_region_memory_end_pfn(r);
5196 * Mark the block movable so that blocks are reserved for
5197 * movable at startup. This will force kernel allocations
5198 * to reserve their blocks rather than leaking throughout
5199 * the address space during boot when many long-lived
5200 * kernel allocations are made.
5202 * bitmap is created for zone's valid pfn range. but memmap
5203 * can be created for invalid pages (for alignment)
5204 * check here not to call set_pageblock_migratetype() against
5207 if (!(pfn & (pageblock_nr_pages - 1))) {
5208 struct page *page = pfn_to_page(pfn);
5210 __init_single_page(page, pfn, zone, nid);
5211 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5213 __init_single_pfn(pfn, zone, nid);
5218 static void __meminit zone_init_free_lists(struct zone *zone)
5220 unsigned int order, t;
5221 for_each_migratetype_order(order, t) {
5222 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5223 zone->free_area[order].nr_free = 0;
5227 #ifndef __HAVE_ARCH_MEMMAP_INIT
5228 #define memmap_init(size, nid, zone, start_pfn) \
5229 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5232 static int zone_batchsize(struct zone *zone)
5238 * The per-cpu-pages pools are set to around 1000th of the
5239 * size of the zone. But no more than 1/2 of a meg.
5241 * OK, so we don't know how big the cache is. So guess.
5243 batch = zone->managed_pages / 1024;
5244 if (batch * PAGE_SIZE > 512 * 1024)
5245 batch = (512 * 1024) / PAGE_SIZE;
5246 batch /= 4; /* We effectively *= 4 below */
5251 * Clamp the batch to a 2^n - 1 value. Having a power
5252 * of 2 value was found to be more likely to have
5253 * suboptimal cache aliasing properties in some cases.
5255 * For example if 2 tasks are alternately allocating
5256 * batches of pages, one task can end up with a lot
5257 * of pages of one half of the possible page colors
5258 * and the other with pages of the other colors.
5260 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5265 /* The deferral and batching of frees should be suppressed under NOMMU
5268 * The problem is that NOMMU needs to be able to allocate large chunks
5269 * of contiguous memory as there's no hardware page translation to
5270 * assemble apparent contiguous memory from discontiguous pages.
5272 * Queueing large contiguous runs of pages for batching, however,
5273 * causes the pages to actually be freed in smaller chunks. As there
5274 * can be a significant delay between the individual batches being
5275 * recycled, this leads to the once large chunks of space being
5276 * fragmented and becoming unavailable for high-order allocations.
5283 * pcp->high and pcp->batch values are related and dependent on one another:
5284 * ->batch must never be higher then ->high.
5285 * The following function updates them in a safe manner without read side
5288 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5289 * those fields changing asynchronously (acording the the above rule).
5291 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5292 * outside of boot time (or some other assurance that no concurrent updaters
5295 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5296 unsigned long batch)
5298 /* start with a fail safe value for batch */
5302 /* Update high, then batch, in order */
5309 /* a companion to pageset_set_high() */
5310 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5312 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5315 static void pageset_init(struct per_cpu_pageset *p)
5317 struct per_cpu_pages *pcp;
5320 memset(p, 0, sizeof(*p));
5324 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5325 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5328 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5331 pageset_set_batch(p, batch);
5335 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5336 * to the value high for the pageset p.
5338 static void pageset_set_high(struct per_cpu_pageset *p,
5341 unsigned long batch = max(1UL, high / 4);
5342 if ((high / 4) > (PAGE_SHIFT * 8))
5343 batch = PAGE_SHIFT * 8;
5345 pageset_update(&p->pcp, high, batch);
5348 static void pageset_set_high_and_batch(struct zone *zone,
5349 struct per_cpu_pageset *pcp)
5351 if (percpu_pagelist_fraction)
5352 pageset_set_high(pcp,
5353 (zone->managed_pages /
5354 percpu_pagelist_fraction));
5356 pageset_set_batch(pcp, zone_batchsize(zone));
5359 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5361 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5364 pageset_set_high_and_batch(zone, pcp);
5367 static void __meminit setup_zone_pageset(struct zone *zone)
5370 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5371 for_each_possible_cpu(cpu)
5372 zone_pageset_init(zone, cpu);
5376 * Allocate per cpu pagesets and initialize them.
5377 * Before this call only boot pagesets were available.
5379 void __init setup_per_cpu_pageset(void)
5383 for_each_populated_zone(zone)
5384 setup_zone_pageset(zone);
5387 static noinline __init_refok
5388 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5394 * The per-page waitqueue mechanism uses hashed waitqueues
5397 zone->wait_table_hash_nr_entries =
5398 wait_table_hash_nr_entries(zone_size_pages);
5399 zone->wait_table_bits =
5400 wait_table_bits(zone->wait_table_hash_nr_entries);
5401 alloc_size = zone->wait_table_hash_nr_entries
5402 * sizeof(wait_queue_head_t);
5404 if (!slab_is_available()) {
5405 zone->wait_table = (wait_queue_head_t *)
5406 memblock_virt_alloc_node_nopanic(
5407 alloc_size, zone->zone_pgdat->node_id);
5410 * This case means that a zone whose size was 0 gets new memory
5411 * via memory hot-add.
5412 * But it may be the case that a new node was hot-added. In
5413 * this case vmalloc() will not be able to use this new node's
5414 * memory - this wait_table must be initialized to use this new
5415 * node itself as well.
5416 * To use this new node's memory, further consideration will be
5419 zone->wait_table = vmalloc(alloc_size);
5421 if (!zone->wait_table)
5424 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5425 init_waitqueue_head(zone->wait_table + i);
5430 static __meminit void zone_pcp_init(struct zone *zone)
5433 * per cpu subsystem is not up at this point. The following code
5434 * relies on the ability of the linker to provide the
5435 * offset of a (static) per cpu variable into the per cpu area.
5437 zone->pageset = &boot_pageset;
5439 if (populated_zone(zone))
5440 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5441 zone->name, zone->present_pages,
5442 zone_batchsize(zone));
5445 int __meminit init_currently_empty_zone(struct zone *zone,
5446 unsigned long zone_start_pfn,
5449 struct pglist_data *pgdat = zone->zone_pgdat;
5451 ret = zone_wait_table_init(zone, size);
5454 pgdat->nr_zones = zone_idx(zone) + 1;
5456 zone->zone_start_pfn = zone_start_pfn;
5458 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5459 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5461 (unsigned long)zone_idx(zone),
5462 zone_start_pfn, (zone_start_pfn + size));
5464 zone_init_free_lists(zone);
5469 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5470 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5473 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5475 int __meminit __early_pfn_to_nid(unsigned long pfn,
5476 struct mminit_pfnnid_cache *state)
5478 unsigned long start_pfn, end_pfn;
5481 if (state->last_start <= pfn && pfn < state->last_end)
5482 return state->last_nid;
5484 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5486 state->last_start = start_pfn;
5487 state->last_end = end_pfn;
5488 state->last_nid = nid;
5493 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5496 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5497 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5498 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5500 * If an architecture guarantees that all ranges registered contain no holes
5501 * and may be freed, this this function may be used instead of calling
5502 * memblock_free_early_nid() manually.
5504 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5506 unsigned long start_pfn, end_pfn;
5509 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5510 start_pfn = min(start_pfn, max_low_pfn);
5511 end_pfn = min(end_pfn, max_low_pfn);
5513 if (start_pfn < end_pfn)
5514 memblock_free_early_nid(PFN_PHYS(start_pfn),
5515 (end_pfn - start_pfn) << PAGE_SHIFT,
5521 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5522 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5524 * If an architecture guarantees that all ranges registered contain no holes and may
5525 * be freed, this function may be used instead of calling memory_present() manually.
5527 void __init sparse_memory_present_with_active_regions(int nid)
5529 unsigned long start_pfn, end_pfn;
5532 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5533 memory_present(this_nid, start_pfn, end_pfn);
5537 * get_pfn_range_for_nid - Return the start and end page frames for a node
5538 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5539 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5540 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5542 * It returns the start and end page frame of a node based on information
5543 * provided by memblock_set_node(). If called for a node
5544 * with no available memory, a warning is printed and the start and end
5547 void __meminit get_pfn_range_for_nid(unsigned int nid,
5548 unsigned long *start_pfn, unsigned long *end_pfn)
5550 unsigned long this_start_pfn, this_end_pfn;
5556 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5557 *start_pfn = min(*start_pfn, this_start_pfn);
5558 *end_pfn = max(*end_pfn, this_end_pfn);
5561 if (*start_pfn == -1UL)
5566 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5567 * assumption is made that zones within a node are ordered in monotonic
5568 * increasing memory addresses so that the "highest" populated zone is used
5570 static void __init find_usable_zone_for_movable(void)
5573 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5574 if (zone_index == ZONE_MOVABLE)
5577 if (arch_zone_highest_possible_pfn[zone_index] >
5578 arch_zone_lowest_possible_pfn[zone_index])
5582 VM_BUG_ON(zone_index == -1);
5583 movable_zone = zone_index;
5587 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5588 * because it is sized independent of architecture. Unlike the other zones,
5589 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5590 * in each node depending on the size of each node and how evenly kernelcore
5591 * is distributed. This helper function adjusts the zone ranges
5592 * provided by the architecture for a given node by using the end of the
5593 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5594 * zones within a node are in order of monotonic increases memory addresses
5596 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5597 unsigned long zone_type,
5598 unsigned long node_start_pfn,
5599 unsigned long node_end_pfn,
5600 unsigned long *zone_start_pfn,
5601 unsigned long *zone_end_pfn)
5603 /* Only adjust if ZONE_MOVABLE is on this node */
5604 if (zone_movable_pfn[nid]) {
5605 /* Size ZONE_MOVABLE */
5606 if (zone_type == ZONE_MOVABLE) {
5607 *zone_start_pfn = zone_movable_pfn[nid];
5608 *zone_end_pfn = min(node_end_pfn,
5609 arch_zone_highest_possible_pfn[movable_zone]);
5611 /* Check if this whole range is within ZONE_MOVABLE */
5612 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5613 *zone_start_pfn = *zone_end_pfn;
5618 * Return the number of pages a zone spans in a node, including holes
5619 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5621 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5622 unsigned long zone_type,
5623 unsigned long node_start_pfn,
5624 unsigned long node_end_pfn,
5625 unsigned long *zone_start_pfn,
5626 unsigned long *zone_end_pfn,
5627 unsigned long *ignored)
5629 /* When hotadd a new node from cpu_up(), the node should be empty */
5630 if (!node_start_pfn && !node_end_pfn)
5633 /* Get the start and end of the zone */
5634 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5635 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5636 adjust_zone_range_for_zone_movable(nid, zone_type,
5637 node_start_pfn, node_end_pfn,
5638 zone_start_pfn, zone_end_pfn);
5640 /* Check that this node has pages within the zone's required range */
5641 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5644 /* Move the zone boundaries inside the node if necessary */
5645 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5646 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5648 /* Return the spanned pages */
5649 return *zone_end_pfn - *zone_start_pfn;
5653 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5654 * then all holes in the requested range will be accounted for.
5656 unsigned long __meminit __absent_pages_in_range(int nid,
5657 unsigned long range_start_pfn,
5658 unsigned long range_end_pfn)
5660 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5661 unsigned long start_pfn, end_pfn;
5664 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5665 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5666 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5667 nr_absent -= end_pfn - start_pfn;
5673 * absent_pages_in_range - Return number of page frames in holes within a range
5674 * @start_pfn: The start PFN to start searching for holes
5675 * @end_pfn: The end PFN to stop searching for holes
5677 * It returns the number of pages frames in memory holes within a range.
5679 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5680 unsigned long end_pfn)
5682 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5685 /* Return the number of page frames in holes in a zone on a node */
5686 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5687 unsigned long zone_type,
5688 unsigned long node_start_pfn,
5689 unsigned long node_end_pfn,
5690 unsigned long *ignored)
5692 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5693 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5694 unsigned long zone_start_pfn, zone_end_pfn;
5695 unsigned long nr_absent;
5697 /* When hotadd a new node from cpu_up(), the node should be empty */
5698 if (!node_start_pfn && !node_end_pfn)
5701 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5702 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5704 adjust_zone_range_for_zone_movable(nid, zone_type,
5705 node_start_pfn, node_end_pfn,
5706 &zone_start_pfn, &zone_end_pfn);
5707 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5710 * ZONE_MOVABLE handling.
5711 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5714 if (zone_movable_pfn[nid]) {
5715 if (mirrored_kernelcore) {
5716 unsigned long start_pfn, end_pfn;
5717 struct memblock_region *r;
5719 for_each_memblock(memory, r) {
5720 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5721 zone_start_pfn, zone_end_pfn);
5722 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5723 zone_start_pfn, zone_end_pfn);
5725 if (zone_type == ZONE_MOVABLE &&
5726 memblock_is_mirror(r))
5727 nr_absent += end_pfn - start_pfn;
5729 if (zone_type == ZONE_NORMAL &&
5730 !memblock_is_mirror(r))
5731 nr_absent += end_pfn - start_pfn;
5734 if (zone_type == ZONE_NORMAL)
5735 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5742 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5743 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5744 unsigned long zone_type,
5745 unsigned long node_start_pfn,
5746 unsigned long node_end_pfn,
5747 unsigned long *zone_start_pfn,
5748 unsigned long *zone_end_pfn,
5749 unsigned long *zones_size)
5753 *zone_start_pfn = node_start_pfn;
5754 for (zone = 0; zone < zone_type; zone++)
5755 *zone_start_pfn += zones_size[zone];
5757 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5759 return zones_size[zone_type];
5762 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5763 unsigned long zone_type,
5764 unsigned long node_start_pfn,
5765 unsigned long node_end_pfn,
5766 unsigned long *zholes_size)
5771 return zholes_size[zone_type];
5774 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5776 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5777 unsigned long node_start_pfn,
5778 unsigned long node_end_pfn,
5779 unsigned long *zones_size,
5780 unsigned long *zholes_size)
5782 unsigned long realtotalpages = 0, totalpages = 0;
5785 for (i = 0; i < MAX_NR_ZONES; i++) {
5786 struct zone *zone = pgdat->node_zones + i;
5787 unsigned long zone_start_pfn, zone_end_pfn;
5788 unsigned long size, real_size;
5790 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5796 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5797 node_start_pfn, node_end_pfn,
5800 zone->zone_start_pfn = zone_start_pfn;
5802 zone->zone_start_pfn = 0;
5803 zone->spanned_pages = size;
5804 zone->present_pages = real_size;
5807 realtotalpages += real_size;
5810 pgdat->node_spanned_pages = totalpages;
5811 pgdat->node_present_pages = realtotalpages;
5812 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5816 #ifndef CONFIG_SPARSEMEM
5818 * Calculate the size of the zone->blockflags rounded to an unsigned long
5819 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5820 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5821 * round what is now in bits to nearest long in bits, then return it in
5824 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5826 unsigned long usemapsize;
5828 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5829 usemapsize = roundup(zonesize, pageblock_nr_pages);
5830 usemapsize = usemapsize >> pageblock_order;
5831 usemapsize *= NR_PAGEBLOCK_BITS;
5832 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5834 return usemapsize / 8;
5837 static void __init setup_usemap(struct pglist_data *pgdat,
5839 unsigned long zone_start_pfn,
5840 unsigned long zonesize)
5842 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5843 zone->pageblock_flags = NULL;
5845 zone->pageblock_flags =
5846 memblock_virt_alloc_node_nopanic(usemapsize,
5850 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5851 unsigned long zone_start_pfn, unsigned long zonesize) {}
5852 #endif /* CONFIG_SPARSEMEM */
5854 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5856 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5857 void __paginginit set_pageblock_order(void)
5861 /* Check that pageblock_nr_pages has not already been setup */
5862 if (pageblock_order)
5865 if (HPAGE_SHIFT > PAGE_SHIFT)
5866 order = HUGETLB_PAGE_ORDER;
5868 order = MAX_ORDER - 1;
5871 * Assume the largest contiguous order of interest is a huge page.
5872 * This value may be variable depending on boot parameters on IA64 and
5875 pageblock_order = order;
5877 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5880 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5881 * is unused as pageblock_order is set at compile-time. See
5882 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5885 void __paginginit set_pageblock_order(void)
5889 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5891 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5892 unsigned long present_pages)
5894 unsigned long pages = spanned_pages;
5897 * Provide a more accurate estimation if there are holes within
5898 * the zone and SPARSEMEM is in use. If there are holes within the
5899 * zone, each populated memory region may cost us one or two extra
5900 * memmap pages due to alignment because memmap pages for each
5901 * populated regions may not naturally algined on page boundary.
5902 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5904 if (spanned_pages > present_pages + (present_pages >> 4) &&
5905 IS_ENABLED(CONFIG_SPARSEMEM))
5906 pages = present_pages;
5908 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5912 * Set up the zone data structures:
5913 * - mark all pages reserved
5914 * - mark all memory queues empty
5915 * - clear the memory bitmaps
5917 * NOTE: pgdat should get zeroed by caller.
5919 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5922 int nid = pgdat->node_id;
5925 pgdat_resize_init(pgdat);
5926 #ifdef CONFIG_NUMA_BALANCING
5927 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5928 pgdat->numabalancing_migrate_nr_pages = 0;
5929 pgdat->numabalancing_migrate_next_window = jiffies;
5931 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5932 spin_lock_init(&pgdat->split_queue_lock);
5933 INIT_LIST_HEAD(&pgdat->split_queue);
5934 pgdat->split_queue_len = 0;
5936 init_waitqueue_head(&pgdat->kswapd_wait);
5937 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5938 #ifdef CONFIG_COMPACTION
5939 init_waitqueue_head(&pgdat->kcompactd_wait);
5941 pgdat_page_ext_init(pgdat);
5943 for (j = 0; j < MAX_NR_ZONES; j++) {
5944 struct zone *zone = pgdat->node_zones + j;
5945 unsigned long size, realsize, freesize, memmap_pages;
5946 unsigned long zone_start_pfn = zone->zone_start_pfn;
5948 size = zone->spanned_pages;
5949 realsize = freesize = zone->present_pages;
5952 * Adjust freesize so that it accounts for how much memory
5953 * is used by this zone for memmap. This affects the watermark
5954 * and per-cpu initialisations
5956 memmap_pages = calc_memmap_size(size, realsize);
5957 if (!is_highmem_idx(j)) {
5958 if (freesize >= memmap_pages) {
5959 freesize -= memmap_pages;
5962 " %s zone: %lu pages used for memmap\n",
5963 zone_names[j], memmap_pages);
5965 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5966 zone_names[j], memmap_pages, freesize);
5969 /* Account for reserved pages */
5970 if (j == 0 && freesize > dma_reserve) {
5971 freesize -= dma_reserve;
5972 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5973 zone_names[0], dma_reserve);
5976 if (!is_highmem_idx(j))
5977 nr_kernel_pages += freesize;
5978 /* Charge for highmem memmap if there are enough kernel pages */
5979 else if (nr_kernel_pages > memmap_pages * 2)
5980 nr_kernel_pages -= memmap_pages;
5981 nr_all_pages += freesize;
5984 * Set an approximate value for lowmem here, it will be adjusted
5985 * when the bootmem allocator frees pages into the buddy system.
5986 * And all highmem pages will be managed by the buddy system.
5988 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5991 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5993 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5995 zone->name = zone_names[j];
5996 spin_lock_init(&zone->lock);
5997 spin_lock_init(&zone->lru_lock);
5998 zone_seqlock_init(zone);
5999 zone->zone_pgdat = pgdat;
6000 zone_pcp_init(zone);
6002 /* For bootup, initialized properly in watermark setup */
6003 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
6005 lruvec_init(&zone->lruvec);
6009 set_pageblock_order();
6010 setup_usemap(pgdat, zone, zone_start_pfn, size);
6011 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6013 memmap_init(size, nid, j, zone_start_pfn);
6017 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
6019 unsigned long __maybe_unused start = 0;
6020 unsigned long __maybe_unused offset = 0;
6022 /* Skip empty nodes */
6023 if (!pgdat->node_spanned_pages)
6026 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6027 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6028 offset = pgdat->node_start_pfn - start;
6029 /* ia64 gets its own node_mem_map, before this, without bootmem */
6030 if (!pgdat->node_mem_map) {
6031 unsigned long size, end;
6035 * The zone's endpoints aren't required to be MAX_ORDER
6036 * aligned but the node_mem_map endpoints must be in order
6037 * for the buddy allocator to function correctly.
6039 end = pgdat_end_pfn(pgdat);
6040 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6041 size = (end - start) * sizeof(struct page);
6042 map = alloc_remap(pgdat->node_id, size);
6044 map = memblock_virt_alloc_node_nopanic(size,
6046 pgdat->node_mem_map = map + offset;
6048 #ifndef CONFIG_NEED_MULTIPLE_NODES
6050 * With no DISCONTIG, the global mem_map is just set as node 0's
6052 if (pgdat == NODE_DATA(0)) {
6053 mem_map = NODE_DATA(0)->node_mem_map;
6054 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6055 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6057 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6060 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6063 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6064 unsigned long node_start_pfn, unsigned long *zholes_size)
6066 pg_data_t *pgdat = NODE_DATA(nid);
6067 unsigned long start_pfn = 0;
6068 unsigned long end_pfn = 0;
6070 /* pg_data_t should be reset to zero when it's allocated */
6071 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6073 reset_deferred_meminit(pgdat);
6074 pgdat->node_id = nid;
6075 pgdat->node_start_pfn = node_start_pfn;
6076 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6077 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6078 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6079 (u64)start_pfn << PAGE_SHIFT,
6080 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6082 start_pfn = node_start_pfn;
6084 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6085 zones_size, zholes_size);
6087 alloc_node_mem_map(pgdat);
6088 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6089 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6090 nid, (unsigned long)pgdat,
6091 (unsigned long)pgdat->node_mem_map);
6094 free_area_init_core(pgdat);
6097 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6099 #if MAX_NUMNODES > 1
6101 * Figure out the number of possible node ids.
6103 void __init setup_nr_node_ids(void)
6105 unsigned int highest;
6107 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6108 nr_node_ids = highest + 1;
6113 * node_map_pfn_alignment - determine the maximum internode alignment
6115 * This function should be called after node map is populated and sorted.
6116 * It calculates the maximum power of two alignment which can distinguish
6119 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6120 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6121 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6122 * shifted, 1GiB is enough and this function will indicate so.
6124 * This is used to test whether pfn -> nid mapping of the chosen memory
6125 * model has fine enough granularity to avoid incorrect mapping for the
6126 * populated node map.
6128 * Returns the determined alignment in pfn's. 0 if there is no alignment
6129 * requirement (single node).
6131 unsigned long __init node_map_pfn_alignment(void)
6133 unsigned long accl_mask = 0, last_end = 0;
6134 unsigned long start, end, mask;
6138 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6139 if (!start || last_nid < 0 || last_nid == nid) {
6146 * Start with a mask granular enough to pin-point to the
6147 * start pfn and tick off bits one-by-one until it becomes
6148 * too coarse to separate the current node from the last.
6150 mask = ~((1 << __ffs(start)) - 1);
6151 while (mask && last_end <= (start & (mask << 1)))
6154 /* accumulate all internode masks */
6158 /* convert mask to number of pages */
6159 return ~accl_mask + 1;
6162 /* Find the lowest pfn for a node */
6163 static unsigned long __init find_min_pfn_for_node(int nid)
6165 unsigned long min_pfn = ULONG_MAX;
6166 unsigned long start_pfn;
6169 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6170 min_pfn = min(min_pfn, start_pfn);
6172 if (min_pfn == ULONG_MAX) {
6173 pr_warn("Could not find start_pfn for node %d\n", nid);
6181 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6183 * It returns the minimum PFN based on information provided via
6184 * memblock_set_node().
6186 unsigned long __init find_min_pfn_with_active_regions(void)
6188 return find_min_pfn_for_node(MAX_NUMNODES);
6192 * early_calculate_totalpages()
6193 * Sum pages in active regions for movable zone.
6194 * Populate N_MEMORY for calculating usable_nodes.
6196 static unsigned long __init early_calculate_totalpages(void)
6198 unsigned long totalpages = 0;
6199 unsigned long start_pfn, end_pfn;
6202 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6203 unsigned long pages = end_pfn - start_pfn;
6205 totalpages += pages;
6207 node_set_state(nid, N_MEMORY);
6213 * Find the PFN the Movable zone begins in each node. Kernel memory
6214 * is spread evenly between nodes as long as the nodes have enough
6215 * memory. When they don't, some nodes will have more kernelcore than
6218 static void __init find_zone_movable_pfns_for_nodes(void)
6221 unsigned long usable_startpfn;
6222 unsigned long kernelcore_node, kernelcore_remaining;
6223 /* save the state before borrow the nodemask */
6224 nodemask_t saved_node_state = node_states[N_MEMORY];
6225 unsigned long totalpages = early_calculate_totalpages();
6226 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6227 struct memblock_region *r;
6229 /* Need to find movable_zone earlier when movable_node is specified. */
6230 find_usable_zone_for_movable();
6233 * If movable_node is specified, ignore kernelcore and movablecore
6236 if (movable_node_is_enabled()) {
6237 for_each_memblock(memory, r) {
6238 if (!memblock_is_hotpluggable(r))
6243 usable_startpfn = PFN_DOWN(r->base);
6244 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6245 min(usable_startpfn, zone_movable_pfn[nid]) :
6253 * If kernelcore=mirror is specified, ignore movablecore option
6255 if (mirrored_kernelcore) {
6256 bool mem_below_4gb_not_mirrored = false;
6258 for_each_memblock(memory, r) {
6259 if (memblock_is_mirror(r))
6264 usable_startpfn = memblock_region_memory_base_pfn(r);
6266 if (usable_startpfn < 0x100000) {
6267 mem_below_4gb_not_mirrored = true;
6271 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6272 min(usable_startpfn, zone_movable_pfn[nid]) :
6276 if (mem_below_4gb_not_mirrored)
6277 pr_warn("This configuration results in unmirrored kernel memory.");
6283 * If movablecore=nn[KMG] was specified, calculate what size of
6284 * kernelcore that corresponds so that memory usable for
6285 * any allocation type is evenly spread. If both kernelcore
6286 * and movablecore are specified, then the value of kernelcore
6287 * will be used for required_kernelcore if it's greater than
6288 * what movablecore would have allowed.
6290 if (required_movablecore) {
6291 unsigned long corepages;
6294 * Round-up so that ZONE_MOVABLE is at least as large as what
6295 * was requested by the user
6297 required_movablecore =
6298 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6299 required_movablecore = min(totalpages, required_movablecore);
6300 corepages = totalpages - required_movablecore;
6302 required_kernelcore = max(required_kernelcore, corepages);
6306 * If kernelcore was not specified or kernelcore size is larger
6307 * than totalpages, there is no ZONE_MOVABLE.
6309 if (!required_kernelcore || required_kernelcore >= totalpages)
6312 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6313 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6316 /* Spread kernelcore memory as evenly as possible throughout nodes */
6317 kernelcore_node = required_kernelcore / usable_nodes;
6318 for_each_node_state(nid, N_MEMORY) {
6319 unsigned long start_pfn, end_pfn;
6322 * Recalculate kernelcore_node if the division per node
6323 * now exceeds what is necessary to satisfy the requested
6324 * amount of memory for the kernel
6326 if (required_kernelcore < kernelcore_node)
6327 kernelcore_node = required_kernelcore / usable_nodes;
6330 * As the map is walked, we track how much memory is usable
6331 * by the kernel using kernelcore_remaining. When it is
6332 * 0, the rest of the node is usable by ZONE_MOVABLE
6334 kernelcore_remaining = kernelcore_node;
6336 /* Go through each range of PFNs within this node */
6337 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6338 unsigned long size_pages;
6340 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6341 if (start_pfn >= end_pfn)
6344 /* Account for what is only usable for kernelcore */
6345 if (start_pfn < usable_startpfn) {
6346 unsigned long kernel_pages;
6347 kernel_pages = min(end_pfn, usable_startpfn)
6350 kernelcore_remaining -= min(kernel_pages,
6351 kernelcore_remaining);
6352 required_kernelcore -= min(kernel_pages,
6353 required_kernelcore);
6355 /* Continue if range is now fully accounted */
6356 if (end_pfn <= usable_startpfn) {
6359 * Push zone_movable_pfn to the end so
6360 * that if we have to rebalance
6361 * kernelcore across nodes, we will
6362 * not double account here
6364 zone_movable_pfn[nid] = end_pfn;
6367 start_pfn = usable_startpfn;
6371 * The usable PFN range for ZONE_MOVABLE is from
6372 * start_pfn->end_pfn. Calculate size_pages as the
6373 * number of pages used as kernelcore
6375 size_pages = end_pfn - start_pfn;
6376 if (size_pages > kernelcore_remaining)
6377 size_pages = kernelcore_remaining;
6378 zone_movable_pfn[nid] = start_pfn + size_pages;
6381 * Some kernelcore has been met, update counts and
6382 * break if the kernelcore for this node has been
6385 required_kernelcore -= min(required_kernelcore,
6387 kernelcore_remaining -= size_pages;
6388 if (!kernelcore_remaining)
6394 * If there is still required_kernelcore, we do another pass with one
6395 * less node in the count. This will push zone_movable_pfn[nid] further
6396 * along on the nodes that still have memory until kernelcore is
6400 if (usable_nodes && required_kernelcore > usable_nodes)
6404 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6405 for (nid = 0; nid < MAX_NUMNODES; nid++)
6406 zone_movable_pfn[nid] =
6407 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6410 /* restore the node_state */
6411 node_states[N_MEMORY] = saved_node_state;
6414 /* Any regular or high memory on that node ? */
6415 static void check_for_memory(pg_data_t *pgdat, int nid)
6417 enum zone_type zone_type;
6419 if (N_MEMORY == N_NORMAL_MEMORY)
6422 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6423 struct zone *zone = &pgdat->node_zones[zone_type];
6424 if (populated_zone(zone)) {
6425 node_set_state(nid, N_HIGH_MEMORY);
6426 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6427 zone_type <= ZONE_NORMAL)
6428 node_set_state(nid, N_NORMAL_MEMORY);
6435 * free_area_init_nodes - Initialise all pg_data_t and zone data
6436 * @max_zone_pfn: an array of max PFNs for each zone
6438 * This will call free_area_init_node() for each active node in the system.
6439 * Using the page ranges provided by memblock_set_node(), the size of each
6440 * zone in each node and their holes is calculated. If the maximum PFN
6441 * between two adjacent zones match, it is assumed that the zone is empty.
6442 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6443 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6444 * starts where the previous one ended. For example, ZONE_DMA32 starts
6445 * at arch_max_dma_pfn.
6447 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6449 unsigned long start_pfn, end_pfn;
6452 /* Record where the zone boundaries are */
6453 memset(arch_zone_lowest_possible_pfn, 0,
6454 sizeof(arch_zone_lowest_possible_pfn));
6455 memset(arch_zone_highest_possible_pfn, 0,
6456 sizeof(arch_zone_highest_possible_pfn));
6457 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6458 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6459 for (i = 1; i < MAX_NR_ZONES; i++) {
6460 if (i == ZONE_MOVABLE)
6462 arch_zone_lowest_possible_pfn[i] =
6463 arch_zone_highest_possible_pfn[i-1];
6464 arch_zone_highest_possible_pfn[i] =
6465 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6467 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6468 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6470 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6471 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6472 find_zone_movable_pfns_for_nodes();
6474 /* Print out the zone ranges */
6475 pr_info("Zone ranges:\n");
6476 for (i = 0; i < MAX_NR_ZONES; i++) {
6477 if (i == ZONE_MOVABLE)
6479 pr_info(" %-8s ", zone_names[i]);
6480 if (arch_zone_lowest_possible_pfn[i] ==
6481 arch_zone_highest_possible_pfn[i])
6484 pr_cont("[mem %#018Lx-%#018Lx]\n",
6485 (u64)arch_zone_lowest_possible_pfn[i]
6487 ((u64)arch_zone_highest_possible_pfn[i]
6488 << PAGE_SHIFT) - 1);
6491 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6492 pr_info("Movable zone start for each node\n");
6493 for (i = 0; i < MAX_NUMNODES; i++) {
6494 if (zone_movable_pfn[i])
6495 pr_info(" Node %d: %#018Lx\n", i,
6496 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6499 /* Print out the early node map */
6500 pr_info("Early memory node ranges\n");
6501 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6502 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6503 (u64)start_pfn << PAGE_SHIFT,
6504 ((u64)end_pfn << PAGE_SHIFT) - 1);
6506 /* Initialise every node */
6507 mminit_verify_pageflags_layout();
6508 setup_nr_node_ids();
6509 for_each_online_node(nid) {
6510 pg_data_t *pgdat = NODE_DATA(nid);
6511 free_area_init_node(nid, NULL,
6512 find_min_pfn_for_node(nid), NULL);
6514 /* Any memory on that node */
6515 if (pgdat->node_present_pages)
6516 node_set_state(nid, N_MEMORY);
6517 check_for_memory(pgdat, nid);
6521 static int __init cmdline_parse_core(char *p, unsigned long *core)
6523 unsigned long long coremem;
6527 coremem = memparse(p, &p);
6528 *core = coremem >> PAGE_SHIFT;
6530 /* Paranoid check that UL is enough for the coremem value */
6531 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6537 * kernelcore=size sets the amount of memory for use for allocations that
6538 * cannot be reclaimed or migrated.
6540 static int __init cmdline_parse_kernelcore(char *p)
6542 /* parse kernelcore=mirror */
6543 if (parse_option_str(p, "mirror")) {
6544 mirrored_kernelcore = true;
6548 return cmdline_parse_core(p, &required_kernelcore);
6552 * movablecore=size sets the amount of memory for use for allocations that
6553 * can be reclaimed or migrated.
6555 static int __init cmdline_parse_movablecore(char *p)
6557 return cmdline_parse_core(p, &required_movablecore);
6560 early_param("kernelcore", cmdline_parse_kernelcore);
6561 early_param("movablecore", cmdline_parse_movablecore);
6563 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6565 void adjust_managed_page_count(struct page *page, long count)
6567 spin_lock(&managed_page_count_lock);
6568 page_zone(page)->managed_pages += count;
6569 totalram_pages += count;
6570 #ifdef CONFIG_HIGHMEM
6571 if (PageHighMem(page))
6572 totalhigh_pages += count;
6574 spin_unlock(&managed_page_count_lock);
6576 EXPORT_SYMBOL(adjust_managed_page_count);
6578 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6581 unsigned long pages = 0;
6583 start = (void *)PAGE_ALIGN((unsigned long)start);
6584 end = (void *)((unsigned long)end & PAGE_MASK);
6585 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6586 if ((unsigned int)poison <= 0xFF)
6587 memset(pos, poison, PAGE_SIZE);
6588 free_reserved_page(virt_to_page(pos));
6592 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6593 s, pages << (PAGE_SHIFT - 10), start, end);
6597 EXPORT_SYMBOL(free_reserved_area);
6599 #ifdef CONFIG_HIGHMEM
6600 void free_highmem_page(struct page *page)
6602 __free_reserved_page(page);
6604 page_zone(page)->managed_pages++;
6610 void __init mem_init_print_info(const char *str)
6612 unsigned long physpages, codesize, datasize, rosize, bss_size;
6613 unsigned long init_code_size, init_data_size;
6615 physpages = get_num_physpages();
6616 codesize = _etext - _stext;
6617 datasize = _edata - _sdata;
6618 rosize = __end_rodata - __start_rodata;
6619 bss_size = __bss_stop - __bss_start;
6620 init_data_size = __init_end - __init_begin;
6621 init_code_size = _einittext - _sinittext;
6624 * Detect special cases and adjust section sizes accordingly:
6625 * 1) .init.* may be embedded into .data sections
6626 * 2) .init.text.* may be out of [__init_begin, __init_end],
6627 * please refer to arch/tile/kernel/vmlinux.lds.S.
6628 * 3) .rodata.* may be embedded into .text or .data sections.
6630 #define adj_init_size(start, end, size, pos, adj) \
6632 if (start <= pos && pos < end && size > adj) \
6636 adj_init_size(__init_begin, __init_end, init_data_size,
6637 _sinittext, init_code_size);
6638 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6639 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6640 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6641 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6643 #undef adj_init_size
6645 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6646 #ifdef CONFIG_HIGHMEM
6650 nr_free_pages() << (PAGE_SHIFT - 10),
6651 physpages << (PAGE_SHIFT - 10),
6652 codesize >> 10, datasize >> 10, rosize >> 10,
6653 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6654 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6655 totalcma_pages << (PAGE_SHIFT - 10),
6656 #ifdef CONFIG_HIGHMEM
6657 totalhigh_pages << (PAGE_SHIFT - 10),
6659 str ? ", " : "", str ? str : "");
6663 * set_dma_reserve - set the specified number of pages reserved in the first zone
6664 * @new_dma_reserve: The number of pages to mark reserved
6666 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6667 * In the DMA zone, a significant percentage may be consumed by kernel image
6668 * and other unfreeable allocations which can skew the watermarks badly. This
6669 * function may optionally be used to account for unfreeable pages in the
6670 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6671 * smaller per-cpu batchsize.
6673 void __init set_dma_reserve(unsigned long new_dma_reserve)
6675 dma_reserve = new_dma_reserve;
6678 void __init free_area_init(unsigned long *zones_size)
6680 free_area_init_node(0, zones_size,
6681 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6684 static int page_alloc_cpu_notify(struct notifier_block *self,
6685 unsigned long action, void *hcpu)
6687 int cpu = (unsigned long)hcpu;
6689 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6690 lru_add_drain_cpu(cpu);
6694 * Spill the event counters of the dead processor
6695 * into the current processors event counters.
6696 * This artificially elevates the count of the current
6699 vm_events_fold_cpu(cpu);
6702 * Zero the differential counters of the dead processor
6703 * so that the vm statistics are consistent.
6705 * This is only okay since the processor is dead and cannot
6706 * race with what we are doing.
6708 cpu_vm_stats_fold(cpu);
6713 void __init page_alloc_init(void)
6715 hotcpu_notifier(page_alloc_cpu_notify, 0);
6719 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6720 * or min_free_kbytes changes.
6722 static void calculate_totalreserve_pages(void)
6724 struct pglist_data *pgdat;
6725 unsigned long reserve_pages = 0;
6726 enum zone_type i, j;
6728 for_each_online_pgdat(pgdat) {
6729 for (i = 0; i < MAX_NR_ZONES; i++) {
6730 struct zone *zone = pgdat->node_zones + i;
6733 /* Find valid and maximum lowmem_reserve in the zone */
6734 for (j = i; j < MAX_NR_ZONES; j++) {
6735 if (zone->lowmem_reserve[j] > max)
6736 max = zone->lowmem_reserve[j];
6739 /* we treat the high watermark as reserved pages. */
6740 max += high_wmark_pages(zone);
6742 if (max > zone->managed_pages)
6743 max = zone->managed_pages;
6745 zone->totalreserve_pages = max;
6747 reserve_pages += max;
6750 totalreserve_pages = reserve_pages;
6754 * setup_per_zone_lowmem_reserve - called whenever
6755 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6756 * has a correct pages reserved value, so an adequate number of
6757 * pages are left in the zone after a successful __alloc_pages().
6759 static void setup_per_zone_lowmem_reserve(void)
6761 struct pglist_data *pgdat;
6762 enum zone_type j, idx;
6764 for_each_online_pgdat(pgdat) {
6765 for (j = 0; j < MAX_NR_ZONES; j++) {
6766 struct zone *zone = pgdat->node_zones + j;
6767 unsigned long managed_pages = zone->managed_pages;
6769 zone->lowmem_reserve[j] = 0;
6773 struct zone *lower_zone;
6777 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6778 sysctl_lowmem_reserve_ratio[idx] = 1;
6780 lower_zone = pgdat->node_zones + idx;
6781 lower_zone->lowmem_reserve[j] = managed_pages /
6782 sysctl_lowmem_reserve_ratio[idx];
6783 managed_pages += lower_zone->managed_pages;
6788 /* update totalreserve_pages */
6789 calculate_totalreserve_pages();
6792 static void __setup_per_zone_wmarks(void)
6794 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6795 unsigned long lowmem_pages = 0;
6797 unsigned long flags;
6799 /* Calculate total number of !ZONE_HIGHMEM pages */
6800 for_each_zone(zone) {
6801 if (!is_highmem(zone))
6802 lowmem_pages += zone->managed_pages;
6805 for_each_zone(zone) {
6808 spin_lock_irqsave(&zone->lock, flags);
6809 tmp = (u64)pages_min * zone->managed_pages;
6810 do_div(tmp, lowmem_pages);
6811 if (is_highmem(zone)) {
6813 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6814 * need highmem pages, so cap pages_min to a small
6817 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6818 * deltas control asynch page reclaim, and so should
6819 * not be capped for highmem.
6821 unsigned long min_pages;
6823 min_pages = zone->managed_pages / 1024;
6824 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6825 zone->watermark[WMARK_MIN] = min_pages;
6828 * If it's a lowmem zone, reserve a number of pages
6829 * proportionate to the zone's size.
6831 zone->watermark[WMARK_MIN] = tmp;
6835 * Set the kswapd watermarks distance according to the
6836 * scale factor in proportion to available memory, but
6837 * ensure a minimum size on small systems.
6839 tmp = max_t(u64, tmp >> 2,
6840 mult_frac(zone->managed_pages,
6841 watermark_scale_factor, 10000));
6843 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6844 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6846 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6847 high_wmark_pages(zone) - low_wmark_pages(zone) -
6848 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6850 spin_unlock_irqrestore(&zone->lock, flags);
6853 /* update totalreserve_pages */
6854 calculate_totalreserve_pages();
6858 * setup_per_zone_wmarks - called when min_free_kbytes changes
6859 * or when memory is hot-{added|removed}
6861 * Ensures that the watermark[min,low,high] values for each zone are set
6862 * correctly with respect to min_free_kbytes.
6864 void setup_per_zone_wmarks(void)
6866 mutex_lock(&zonelists_mutex);
6867 __setup_per_zone_wmarks();
6868 mutex_unlock(&zonelists_mutex);
6872 * Initialise min_free_kbytes.
6874 * For small machines we want it small (128k min). For large machines
6875 * we want it large (64MB max). But it is not linear, because network
6876 * bandwidth does not increase linearly with machine size. We use
6878 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6879 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6895 int __meminit init_per_zone_wmark_min(void)
6897 unsigned long lowmem_kbytes;
6898 int new_min_free_kbytes;
6900 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6901 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6903 if (new_min_free_kbytes > user_min_free_kbytes) {
6904 min_free_kbytes = new_min_free_kbytes;
6905 if (min_free_kbytes < 128)
6906 min_free_kbytes = 128;
6907 if (min_free_kbytes > 65536)
6908 min_free_kbytes = 65536;
6910 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6911 new_min_free_kbytes, user_min_free_kbytes);
6913 setup_per_zone_wmarks();
6914 refresh_zone_stat_thresholds();
6915 setup_per_zone_lowmem_reserve();
6918 core_initcall(init_per_zone_wmark_min)
6921 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6922 * that we can call two helper functions whenever min_free_kbytes
6925 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6926 void __user *buffer, size_t *length, loff_t *ppos)
6930 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6935 user_min_free_kbytes = min_free_kbytes;
6936 setup_per_zone_wmarks();
6941 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6942 void __user *buffer, size_t *length, loff_t *ppos)
6946 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6951 setup_per_zone_wmarks();
6957 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6958 void __user *buffer, size_t *length, loff_t *ppos)
6963 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6968 zone->min_unmapped_pages = (zone->managed_pages *
6969 sysctl_min_unmapped_ratio) / 100;
6973 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6974 void __user *buffer, size_t *length, loff_t *ppos)
6979 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6984 zone->min_slab_pages = (zone->managed_pages *
6985 sysctl_min_slab_ratio) / 100;
6991 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6992 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6993 * whenever sysctl_lowmem_reserve_ratio changes.
6995 * The reserve ratio obviously has absolutely no relation with the
6996 * minimum watermarks. The lowmem reserve ratio can only make sense
6997 * if in function of the boot time zone sizes.
6999 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7000 void __user *buffer, size_t *length, loff_t *ppos)
7002 proc_dointvec_minmax(table, write, buffer, length, ppos);
7003 setup_per_zone_lowmem_reserve();
7008 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7009 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7010 * pagelist can have before it gets flushed back to buddy allocator.
7012 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7013 void __user *buffer, size_t *length, loff_t *ppos)
7016 int old_percpu_pagelist_fraction;
7019 mutex_lock(&pcp_batch_high_lock);
7020 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7022 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7023 if (!write || ret < 0)
7026 /* Sanity checking to avoid pcp imbalance */
7027 if (percpu_pagelist_fraction &&
7028 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7029 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7035 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7038 for_each_populated_zone(zone) {
7041 for_each_possible_cpu(cpu)
7042 pageset_set_high_and_batch(zone,
7043 per_cpu_ptr(zone->pageset, cpu));
7046 mutex_unlock(&pcp_batch_high_lock);
7051 int hashdist = HASHDIST_DEFAULT;
7053 static int __init set_hashdist(char *str)
7057 hashdist = simple_strtoul(str, &str, 0);
7060 __setup("hashdist=", set_hashdist);
7064 * allocate a large system hash table from bootmem
7065 * - it is assumed that the hash table must contain an exact power-of-2
7066 * quantity of entries
7067 * - limit is the number of hash buckets, not the total allocation size
7069 void *__init alloc_large_system_hash(const char *tablename,
7070 unsigned long bucketsize,
7071 unsigned long numentries,
7074 unsigned int *_hash_shift,
7075 unsigned int *_hash_mask,
7076 unsigned long low_limit,
7077 unsigned long high_limit)
7079 unsigned long long max = high_limit;
7080 unsigned long log2qty, size;
7083 /* allow the kernel cmdline to have a say */
7085 /* round applicable memory size up to nearest megabyte */
7086 numentries = nr_kernel_pages;
7088 /* It isn't necessary when PAGE_SIZE >= 1MB */
7089 if (PAGE_SHIFT < 20)
7090 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7092 /* limit to 1 bucket per 2^scale bytes of low memory */
7093 if (scale > PAGE_SHIFT)
7094 numentries >>= (scale - PAGE_SHIFT);
7096 numentries <<= (PAGE_SHIFT - scale);
7098 /* Make sure we've got at least a 0-order allocation.. */
7099 if (unlikely(flags & HASH_SMALL)) {
7100 /* Makes no sense without HASH_EARLY */
7101 WARN_ON(!(flags & HASH_EARLY));
7102 if (!(numentries >> *_hash_shift)) {
7103 numentries = 1UL << *_hash_shift;
7104 BUG_ON(!numentries);
7106 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7107 numentries = PAGE_SIZE / bucketsize;
7109 numentries = roundup_pow_of_two(numentries);
7111 /* limit allocation size to 1/16 total memory by default */
7113 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7114 do_div(max, bucketsize);
7116 max = min(max, 0x80000000ULL);
7118 if (numentries < low_limit)
7119 numentries = low_limit;
7120 if (numentries > max)
7123 log2qty = ilog2(numentries);
7126 size = bucketsize << log2qty;
7127 if (flags & HASH_EARLY)
7128 table = memblock_virt_alloc_nopanic(size, 0);
7130 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7133 * If bucketsize is not a power-of-two, we may free
7134 * some pages at the end of hash table which
7135 * alloc_pages_exact() automatically does
7137 if (get_order(size) < MAX_ORDER) {
7138 table = alloc_pages_exact(size, GFP_ATOMIC);
7139 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7142 } while (!table && size > PAGE_SIZE && --log2qty);
7145 panic("Failed to allocate %s hash table\n", tablename);
7147 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7148 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7151 *_hash_shift = log2qty;
7153 *_hash_mask = (1 << log2qty) - 1;
7159 * This function checks whether pageblock includes unmovable pages or not.
7160 * If @count is not zero, it is okay to include less @count unmovable pages
7162 * PageLRU check without isolation or lru_lock could race so that
7163 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7164 * expect this function should be exact.
7166 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7167 bool skip_hwpoisoned_pages)
7169 unsigned long pfn, iter, found;
7173 * For avoiding noise data, lru_add_drain_all() should be called
7174 * If ZONE_MOVABLE, the zone never contains unmovable pages
7176 if (zone_idx(zone) == ZONE_MOVABLE)
7178 mt = get_pageblock_migratetype(page);
7179 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7182 pfn = page_to_pfn(page);
7183 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7184 unsigned long check = pfn + iter;
7186 if (!pfn_valid_within(check))
7189 page = pfn_to_page(check);
7192 * Hugepages are not in LRU lists, but they're movable.
7193 * We need not scan over tail pages bacause we don't
7194 * handle each tail page individually in migration.
7196 if (PageHuge(page)) {
7197 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7202 * We can't use page_count without pin a page
7203 * because another CPU can free compound page.
7204 * This check already skips compound tails of THP
7205 * because their page->_refcount is zero at all time.
7207 if (!page_ref_count(page)) {
7208 if (PageBuddy(page))
7209 iter += (1 << page_order(page)) - 1;
7214 * The HWPoisoned page may be not in buddy system, and
7215 * page_count() is not 0.
7217 if (skip_hwpoisoned_pages && PageHWPoison(page))
7223 * If there are RECLAIMABLE pages, we need to check
7224 * it. But now, memory offline itself doesn't call
7225 * shrink_node_slabs() and it still to be fixed.
7228 * If the page is not RAM, page_count()should be 0.
7229 * we don't need more check. This is an _used_ not-movable page.
7231 * The problematic thing here is PG_reserved pages. PG_reserved
7232 * is set to both of a memory hole page and a _used_ kernel
7241 bool is_pageblock_removable_nolock(struct page *page)
7247 * We have to be careful here because we are iterating over memory
7248 * sections which are not zone aware so we might end up outside of
7249 * the zone but still within the section.
7250 * We have to take care about the node as well. If the node is offline
7251 * its NODE_DATA will be NULL - see page_zone.
7253 if (!node_online(page_to_nid(page)))
7256 zone = page_zone(page);
7257 pfn = page_to_pfn(page);
7258 if (!zone_spans_pfn(zone, pfn))
7261 return !has_unmovable_pages(zone, page, 0, true);
7264 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7266 static unsigned long pfn_max_align_down(unsigned long pfn)
7268 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7269 pageblock_nr_pages) - 1);
7272 static unsigned long pfn_max_align_up(unsigned long pfn)
7274 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7275 pageblock_nr_pages));
7278 /* [start, end) must belong to a single zone. */
7279 static int __alloc_contig_migrate_range(struct compact_control *cc,
7280 unsigned long start, unsigned long end)
7282 /* This function is based on compact_zone() from compaction.c. */
7283 unsigned long nr_reclaimed;
7284 unsigned long pfn = start;
7285 unsigned int tries = 0;
7290 while (pfn < end || !list_empty(&cc->migratepages)) {
7291 if (fatal_signal_pending(current)) {
7296 if (list_empty(&cc->migratepages)) {
7297 cc->nr_migratepages = 0;
7298 pfn = isolate_migratepages_range(cc, pfn, end);
7304 } else if (++tries == 5) {
7305 ret = ret < 0 ? ret : -EBUSY;
7309 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7311 cc->nr_migratepages -= nr_reclaimed;
7313 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7314 NULL, 0, cc->mode, MR_CMA);
7317 putback_movable_pages(&cc->migratepages);
7324 * alloc_contig_range() -- tries to allocate given range of pages
7325 * @start: start PFN to allocate
7326 * @end: one-past-the-last PFN to allocate
7327 * @migratetype: migratetype of the underlaying pageblocks (either
7328 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7329 * in range must have the same migratetype and it must
7330 * be either of the two.
7332 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7333 * aligned, however it's the caller's responsibility to guarantee that
7334 * we are the only thread that changes migrate type of pageblocks the
7337 * The PFN range must belong to a single zone.
7339 * Returns zero on success or negative error code. On success all
7340 * pages which PFN is in [start, end) are allocated for the caller and
7341 * need to be freed with free_contig_range().
7343 int alloc_contig_range(unsigned long start, unsigned long end,
7344 unsigned migratetype)
7346 unsigned long outer_start, outer_end;
7350 struct compact_control cc = {
7351 .nr_migratepages = 0,
7353 .zone = page_zone(pfn_to_page(start)),
7354 .mode = MIGRATE_SYNC,
7355 .ignore_skip_hint = true,
7357 INIT_LIST_HEAD(&cc.migratepages);
7360 * What we do here is we mark all pageblocks in range as
7361 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7362 * have different sizes, and due to the way page allocator
7363 * work, we align the range to biggest of the two pages so
7364 * that page allocator won't try to merge buddies from
7365 * different pageblocks and change MIGRATE_ISOLATE to some
7366 * other migration type.
7368 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7369 * migrate the pages from an unaligned range (ie. pages that
7370 * we are interested in). This will put all the pages in
7371 * range back to page allocator as MIGRATE_ISOLATE.
7373 * When this is done, we take the pages in range from page
7374 * allocator removing them from the buddy system. This way
7375 * page allocator will never consider using them.
7377 * This lets us mark the pageblocks back as
7378 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7379 * aligned range but not in the unaligned, original range are
7380 * put back to page allocator so that buddy can use them.
7383 ret = start_isolate_page_range(pfn_max_align_down(start),
7384 pfn_max_align_up(end), migratetype,
7390 * In case of -EBUSY, we'd like to know which page causes problem.
7391 * So, just fall through. We will check it in test_pages_isolated().
7393 ret = __alloc_contig_migrate_range(&cc, start, end);
7394 if (ret && ret != -EBUSY)
7398 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7399 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7400 * more, all pages in [start, end) are free in page allocator.
7401 * What we are going to do is to allocate all pages from
7402 * [start, end) (that is remove them from page allocator).
7404 * The only problem is that pages at the beginning and at the
7405 * end of interesting range may be not aligned with pages that
7406 * page allocator holds, ie. they can be part of higher order
7407 * pages. Because of this, we reserve the bigger range and
7408 * once this is done free the pages we are not interested in.
7410 * We don't have to hold zone->lock here because the pages are
7411 * isolated thus they won't get removed from buddy.
7414 lru_add_drain_all();
7415 drain_all_pages(cc.zone);
7418 outer_start = start;
7419 while (!PageBuddy(pfn_to_page(outer_start))) {
7420 if (++order >= MAX_ORDER) {
7421 outer_start = start;
7424 outer_start &= ~0UL << order;
7427 if (outer_start != start) {
7428 order = page_order(pfn_to_page(outer_start));
7431 * outer_start page could be small order buddy page and
7432 * it doesn't include start page. Adjust outer_start
7433 * in this case to report failed page properly
7434 * on tracepoint in test_pages_isolated()
7436 if (outer_start + (1UL << order) <= start)
7437 outer_start = start;
7440 /* Make sure the range is really isolated. */
7441 if (test_pages_isolated(outer_start, end, false)) {
7442 pr_info("%s: [%lx, %lx) PFNs busy\n",
7443 __func__, outer_start, end);
7448 /* Grab isolated pages from freelists. */
7449 outer_end = isolate_freepages_range(&cc, outer_start, end);
7455 /* Free head and tail (if any) */
7456 if (start != outer_start)
7457 free_contig_range(outer_start, start - outer_start);
7458 if (end != outer_end)
7459 free_contig_range(end, outer_end - end);
7462 undo_isolate_page_range(pfn_max_align_down(start),
7463 pfn_max_align_up(end), migratetype);
7467 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7469 unsigned int count = 0;
7471 for (; nr_pages--; pfn++) {
7472 struct page *page = pfn_to_page(pfn);
7474 count += page_count(page) != 1;
7477 WARN(count != 0, "%d pages are still in use!\n", count);
7481 #ifdef CONFIG_MEMORY_HOTPLUG
7483 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7484 * page high values need to be recalulated.
7486 void __meminit zone_pcp_update(struct zone *zone)
7489 mutex_lock(&pcp_batch_high_lock);
7490 for_each_possible_cpu(cpu)
7491 pageset_set_high_and_batch(zone,
7492 per_cpu_ptr(zone->pageset, cpu));
7493 mutex_unlock(&pcp_batch_high_lock);
7497 void zone_pcp_reset(struct zone *zone)
7499 unsigned long flags;
7501 struct per_cpu_pageset *pset;
7503 /* avoid races with drain_pages() */
7504 local_irq_save(flags);
7505 if (zone->pageset != &boot_pageset) {
7506 for_each_online_cpu(cpu) {
7507 pset = per_cpu_ptr(zone->pageset, cpu);
7508 drain_zonestat(zone, pset);
7510 free_percpu(zone->pageset);
7511 zone->pageset = &boot_pageset;
7513 local_irq_restore(flags);
7516 #ifdef CONFIG_MEMORY_HOTREMOVE
7518 * All pages in the range must be in a single zone and isolated
7519 * before calling this.
7522 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7526 unsigned int order, i;
7528 unsigned long flags;
7529 /* find the first valid pfn */
7530 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7535 zone = page_zone(pfn_to_page(pfn));
7536 spin_lock_irqsave(&zone->lock, flags);
7538 while (pfn < end_pfn) {
7539 if (!pfn_valid(pfn)) {
7543 page = pfn_to_page(pfn);
7545 * The HWPoisoned page may be not in buddy system, and
7546 * page_count() is not 0.
7548 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7550 SetPageReserved(page);
7554 BUG_ON(page_count(page));
7555 BUG_ON(!PageBuddy(page));
7556 order = page_order(page);
7557 #ifdef CONFIG_DEBUG_VM
7558 pr_info("remove from free list %lx %d %lx\n",
7559 pfn, 1 << order, end_pfn);
7561 list_del(&page->lru);
7562 rmv_page_order(page);
7563 zone->free_area[order].nr_free--;
7564 for (i = 0; i < (1 << order); i++)
7565 SetPageReserved((page+i));
7566 pfn += (1 << order);
7568 spin_unlock_irqrestore(&zone->lock, flags);
7572 bool is_free_buddy_page(struct page *page)
7574 struct zone *zone = page_zone(page);
7575 unsigned long pfn = page_to_pfn(page);
7576 unsigned long flags;
7579 spin_lock_irqsave(&zone->lock, flags);
7580 for (order = 0; order < MAX_ORDER; order++) {
7581 struct page *page_head = page - (pfn & ((1 << order) - 1));
7583 if (PageBuddy(page_head) && page_order(page_head) >= order)
7586 spin_unlock_irqrestore(&zone->lock, flags);
7588 return order < MAX_ORDER;