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 int nid = early_pfn_to_nid(pfn);
291 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
297 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
299 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
306 * Returns false when the remaining initialisation should be deferred until
307 * later in the boot cycle when it can be parallelised.
309 static inline bool update_defer_init(pg_data_t *pgdat,
310 unsigned long pfn, unsigned long zone_end,
311 unsigned long *nr_initialised)
313 unsigned long max_initialise;
315 /* Always populate low zones for address-contrained allocations */
316 if (zone_end < pgdat_end_pfn(pgdat))
319 * Initialise at least 2G of a node but also take into account that
320 * two large system hashes that can take up 1GB for 0.25TB/node.
322 max_initialise = max(2UL << (30 - PAGE_SHIFT),
323 (pgdat->node_spanned_pages >> 8));
326 if ((*nr_initialised > max_initialise) &&
327 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
328 pgdat->first_deferred_pfn = pfn;
335 static inline void reset_deferred_meminit(pg_data_t *pgdat)
339 static inline bool early_page_uninitialised(unsigned long pfn)
344 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
349 static inline bool update_defer_init(pg_data_t *pgdat,
350 unsigned long pfn, unsigned long zone_end,
351 unsigned long *nr_initialised)
357 /* Return a pointer to the bitmap storing bits affecting a block of pages */
358 static inline unsigned long *get_pageblock_bitmap(struct page *page,
361 #ifdef CONFIG_SPARSEMEM
362 return __pfn_to_section(pfn)->pageblock_flags;
364 return page_zone(page)->pageblock_flags;
365 #endif /* CONFIG_SPARSEMEM */
368 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
370 #ifdef CONFIG_SPARSEMEM
371 pfn &= (PAGES_PER_SECTION-1);
372 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
374 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
375 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
376 #endif /* CONFIG_SPARSEMEM */
380 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
381 * @page: The page within the block of interest
382 * @pfn: The target page frame number
383 * @end_bitidx: The last bit of interest to retrieve
384 * @mask: mask of bits that the caller is interested in
386 * Return: pageblock_bits flags
388 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
390 unsigned long end_bitidx,
393 unsigned long *bitmap;
394 unsigned long bitidx, word_bitidx;
397 bitmap = get_pageblock_bitmap(page, pfn);
398 bitidx = pfn_to_bitidx(page, pfn);
399 word_bitidx = bitidx / BITS_PER_LONG;
400 bitidx &= (BITS_PER_LONG-1);
402 word = bitmap[word_bitidx];
403 bitidx += end_bitidx;
404 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
407 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
408 unsigned long end_bitidx,
411 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
414 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
416 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
420 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
421 * @page: The page within the block of interest
422 * @flags: The flags to set
423 * @pfn: The target page frame number
424 * @end_bitidx: The last bit of interest
425 * @mask: mask of bits that the caller is interested in
427 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
429 unsigned long end_bitidx,
432 unsigned long *bitmap;
433 unsigned long bitidx, word_bitidx;
434 unsigned long old_word, word;
436 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
438 bitmap = get_pageblock_bitmap(page, pfn);
439 bitidx = pfn_to_bitidx(page, pfn);
440 word_bitidx = bitidx / BITS_PER_LONG;
441 bitidx &= (BITS_PER_LONG-1);
443 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
445 bitidx += end_bitidx;
446 mask <<= (BITS_PER_LONG - bitidx - 1);
447 flags <<= (BITS_PER_LONG - bitidx - 1);
449 word = READ_ONCE(bitmap[word_bitidx]);
451 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
452 if (word == old_word)
458 void set_pageblock_migratetype(struct page *page, int migratetype)
460 if (unlikely(page_group_by_mobility_disabled &&
461 migratetype < MIGRATE_PCPTYPES))
462 migratetype = MIGRATE_UNMOVABLE;
464 set_pageblock_flags_group(page, (unsigned long)migratetype,
465 PB_migrate, PB_migrate_end);
468 #ifdef CONFIG_DEBUG_VM
469 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
473 unsigned long pfn = page_to_pfn(page);
474 unsigned long sp, start_pfn;
477 seq = zone_span_seqbegin(zone);
478 start_pfn = zone->zone_start_pfn;
479 sp = zone->spanned_pages;
480 if (!zone_spans_pfn(zone, pfn))
482 } while (zone_span_seqretry(zone, seq));
485 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
486 pfn, zone_to_nid(zone), zone->name,
487 start_pfn, start_pfn + sp);
492 static int page_is_consistent(struct zone *zone, struct page *page)
494 if (!pfn_valid_within(page_to_pfn(page)))
496 if (zone != page_zone(page))
502 * Temporary debugging check for pages not lying within a given zone.
504 static int bad_range(struct zone *zone, struct page *page)
506 if (page_outside_zone_boundaries(zone, page))
508 if (!page_is_consistent(zone, page))
514 static inline int bad_range(struct zone *zone, struct page *page)
520 static void bad_page(struct page *page, const char *reason,
521 unsigned long bad_flags)
523 static unsigned long resume;
524 static unsigned long nr_shown;
525 static unsigned long nr_unshown;
528 * Allow a burst of 60 reports, then keep quiet for that minute;
529 * or allow a steady drip of one report per second.
531 if (nr_shown == 60) {
532 if (time_before(jiffies, resume)) {
538 "BUG: Bad page state: %lu messages suppressed\n",
545 resume = jiffies + 60 * HZ;
547 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
548 current->comm, page_to_pfn(page));
549 __dump_page(page, reason);
550 bad_flags &= page->flags;
552 pr_alert("bad because of flags: %#lx(%pGp)\n",
553 bad_flags, &bad_flags);
554 dump_page_owner(page);
559 /* Leave bad fields for debug, except PageBuddy could make trouble */
560 page_mapcount_reset(page); /* remove PageBuddy */
561 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
565 * Higher-order pages are called "compound pages". They are structured thusly:
567 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
569 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
570 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
572 * The first tail page's ->compound_dtor holds the offset in array of compound
573 * page destructors. See compound_page_dtors.
575 * The first tail page's ->compound_order holds the order of allocation.
576 * This usage means that zero-order pages may not be compound.
579 void free_compound_page(struct page *page)
581 __free_pages_ok(page, compound_order(page));
584 void prep_compound_page(struct page *page, unsigned int order)
587 int nr_pages = 1 << order;
589 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
590 set_compound_order(page, order);
592 for (i = 1; i < nr_pages; i++) {
593 struct page *p = page + i;
594 set_page_count(p, 0);
595 p->mapping = TAIL_MAPPING;
596 set_compound_head(p, page);
598 atomic_set(compound_mapcount_ptr(page), -1);
601 #ifdef CONFIG_DEBUG_PAGEALLOC
602 unsigned int _debug_guardpage_minorder;
603 bool _debug_pagealloc_enabled __read_mostly
604 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
605 EXPORT_SYMBOL(_debug_pagealloc_enabled);
606 bool _debug_guardpage_enabled __read_mostly;
608 static int __init early_debug_pagealloc(char *buf)
612 return kstrtobool(buf, &_debug_pagealloc_enabled);
614 early_param("debug_pagealloc", early_debug_pagealloc);
616 static bool need_debug_guardpage(void)
618 /* If we don't use debug_pagealloc, we don't need guard page */
619 if (!debug_pagealloc_enabled())
625 static void init_debug_guardpage(void)
627 if (!debug_pagealloc_enabled())
630 _debug_guardpage_enabled = true;
633 struct page_ext_operations debug_guardpage_ops = {
634 .need = need_debug_guardpage,
635 .init = init_debug_guardpage,
638 static int __init debug_guardpage_minorder_setup(char *buf)
642 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
643 pr_err("Bad debug_guardpage_minorder value\n");
646 _debug_guardpage_minorder = res;
647 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
650 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
652 static inline void set_page_guard(struct zone *zone, struct page *page,
653 unsigned int order, int migratetype)
655 struct page_ext *page_ext;
657 if (!debug_guardpage_enabled())
660 page_ext = lookup_page_ext(page);
661 if (unlikely(!page_ext))
664 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
666 INIT_LIST_HEAD(&page->lru);
667 set_page_private(page, order);
668 /* Guard pages are not available for any usage */
669 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
672 static inline void clear_page_guard(struct zone *zone, struct page *page,
673 unsigned int order, int migratetype)
675 struct page_ext *page_ext;
677 if (!debug_guardpage_enabled())
680 page_ext = lookup_page_ext(page);
681 if (unlikely(!page_ext))
684 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
686 set_page_private(page, 0);
687 if (!is_migrate_isolate(migratetype))
688 __mod_zone_freepage_state(zone, (1 << order), migratetype);
691 struct page_ext_operations debug_guardpage_ops = { NULL, };
692 static inline void set_page_guard(struct zone *zone, struct page *page,
693 unsigned int order, int migratetype) {}
694 static inline void clear_page_guard(struct zone *zone, struct page *page,
695 unsigned int order, int migratetype) {}
698 static inline void set_page_order(struct page *page, unsigned int order)
700 set_page_private(page, order);
701 __SetPageBuddy(page);
704 static inline void rmv_page_order(struct page *page)
706 __ClearPageBuddy(page);
707 set_page_private(page, 0);
711 * This function checks whether a page is free && is the buddy
712 * we can do coalesce a page and its buddy if
713 * (a) the buddy is not in a hole &&
714 * (b) the buddy is in the buddy system &&
715 * (c) a page and its buddy have the same order &&
716 * (d) a page and its buddy are in the same zone.
718 * For recording whether a page is in the buddy system, we set ->_mapcount
719 * PAGE_BUDDY_MAPCOUNT_VALUE.
720 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
721 * serialized by zone->lock.
723 * For recording page's order, we use page_private(page).
725 static inline int page_is_buddy(struct page *page, struct page *buddy,
728 if (!pfn_valid_within(page_to_pfn(buddy)))
731 if (page_is_guard(buddy) && page_order(buddy) == order) {
732 if (page_zone_id(page) != page_zone_id(buddy))
735 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
740 if (PageBuddy(buddy) && page_order(buddy) == order) {
742 * zone check is done late to avoid uselessly
743 * calculating zone/node ids for pages that could
746 if (page_zone_id(page) != page_zone_id(buddy))
749 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
757 * Freeing function for a buddy system allocator.
759 * The concept of a buddy system is to maintain direct-mapped table
760 * (containing bit values) for memory blocks of various "orders".
761 * The bottom level table contains the map for the smallest allocatable
762 * units of memory (here, pages), and each level above it describes
763 * pairs of units from the levels below, hence, "buddies".
764 * At a high level, all that happens here is marking the table entry
765 * at the bottom level available, and propagating the changes upward
766 * as necessary, plus some accounting needed to play nicely with other
767 * parts of the VM system.
768 * At each level, we keep a list of pages, which are heads of continuous
769 * free pages of length of (1 << order) and marked with _mapcount
770 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
772 * So when we are allocating or freeing one, we can derive the state of the
773 * other. That is, if we allocate a small block, and both were
774 * free, the remainder of the region must be split into blocks.
775 * If a block is freed, and its buddy is also free, then this
776 * triggers coalescing into a block of larger size.
781 static inline void __free_one_page(struct page *page,
783 struct zone *zone, unsigned int order,
786 unsigned long page_idx;
787 unsigned long combined_idx;
788 unsigned long uninitialized_var(buddy_idx);
790 unsigned int max_order;
792 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
794 VM_BUG_ON(!zone_is_initialized(zone));
795 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
797 VM_BUG_ON(migratetype == -1);
798 if (likely(!is_migrate_isolate(migratetype)))
799 __mod_zone_freepage_state(zone, 1 << order, migratetype);
801 page_idx = pfn & ((1 << MAX_ORDER) - 1);
803 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
804 VM_BUG_ON_PAGE(bad_range(zone, page), page);
807 while (order < max_order - 1) {
808 buddy_idx = __find_buddy_index(page_idx, order);
809 buddy = page + (buddy_idx - page_idx);
810 if (!page_is_buddy(page, buddy, order))
813 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
814 * merge with it and move up one order.
816 if (page_is_guard(buddy)) {
817 clear_page_guard(zone, buddy, order, migratetype);
819 list_del(&buddy->lru);
820 zone->free_area[order].nr_free--;
821 rmv_page_order(buddy);
823 combined_idx = buddy_idx & page_idx;
824 page = page + (combined_idx - page_idx);
825 page_idx = combined_idx;
828 if (max_order < MAX_ORDER) {
829 /* If we are here, it means order is >= pageblock_order.
830 * We want to prevent merge between freepages on isolate
831 * pageblock and normal pageblock. Without this, pageblock
832 * isolation could cause incorrect freepage or CMA accounting.
834 * We don't want to hit this code for the more frequent
837 if (unlikely(has_isolate_pageblock(zone))) {
840 buddy_idx = __find_buddy_index(page_idx, order);
841 buddy = page + (buddy_idx - page_idx);
842 buddy_mt = get_pageblock_migratetype(buddy);
844 if (migratetype != buddy_mt
845 && (is_migrate_isolate(migratetype) ||
846 is_migrate_isolate(buddy_mt)))
850 goto continue_merging;
854 set_page_order(page, order);
857 * If this is not the largest possible page, check if the buddy
858 * of the next-highest order is free. If it is, it's possible
859 * that pages are being freed that will coalesce soon. In case,
860 * that is happening, add the free page to the tail of the list
861 * so it's less likely to be used soon and more likely to be merged
862 * as a higher order page
864 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
865 struct page *higher_page, *higher_buddy;
866 combined_idx = buddy_idx & page_idx;
867 higher_page = page + (combined_idx - page_idx);
868 buddy_idx = __find_buddy_index(combined_idx, order + 1);
869 higher_buddy = higher_page + (buddy_idx - combined_idx);
870 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
871 list_add_tail(&page->lru,
872 &zone->free_area[order].free_list[migratetype]);
877 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
879 zone->free_area[order].nr_free++;
883 * A bad page could be due to a number of fields. Instead of multiple branches,
884 * try and check multiple fields with one check. The caller must do a detailed
885 * check if necessary.
887 static inline bool page_expected_state(struct page *page,
888 unsigned long check_flags)
890 if (unlikely(atomic_read(&page->_mapcount) != -1))
893 if (unlikely((unsigned long)page->mapping |
894 page_ref_count(page) |
896 (unsigned long)page->mem_cgroup |
898 (page->flags & check_flags)))
904 static void free_pages_check_bad(struct page *page)
906 const char *bad_reason;
907 unsigned long bad_flags;
912 if (unlikely(atomic_read(&page->_mapcount) != -1))
913 bad_reason = "nonzero mapcount";
914 if (unlikely(page->mapping != NULL))
915 bad_reason = "non-NULL mapping";
916 if (unlikely(page_ref_count(page) != 0))
917 bad_reason = "nonzero _refcount";
918 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
919 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
920 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
923 if (unlikely(page->mem_cgroup))
924 bad_reason = "page still charged to cgroup";
926 bad_page(page, bad_reason, bad_flags);
929 static inline int free_pages_check(struct page *page)
931 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
934 /* Something has gone sideways, find it */
935 free_pages_check_bad(page);
939 static int free_tail_pages_check(struct page *head_page, struct page *page)
944 * We rely page->lru.next never has bit 0 set, unless the page
945 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
947 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
949 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
953 switch (page - head_page) {
955 /* the first tail page: ->mapping is compound_mapcount() */
956 if (unlikely(compound_mapcount(page))) {
957 bad_page(page, "nonzero compound_mapcount", 0);
963 * the second tail page: ->mapping is
964 * page_deferred_list().next -- ignore value.
968 if (page->mapping != TAIL_MAPPING) {
969 bad_page(page, "corrupted mapping in tail page", 0);
974 if (unlikely(!PageTail(page))) {
975 bad_page(page, "PageTail not set", 0);
978 if (unlikely(compound_head(page) != head_page)) {
979 bad_page(page, "compound_head not consistent", 0);
984 page->mapping = NULL;
985 clear_compound_head(page);
989 static __always_inline bool free_pages_prepare(struct page *page,
990 unsigned int order, bool check_free)
994 VM_BUG_ON_PAGE(PageTail(page), page);
996 trace_mm_page_free(page, order);
997 kmemcheck_free_shadow(page, order);
1000 * Check tail pages before head page information is cleared to
1001 * avoid checking PageCompound for order-0 pages.
1003 if (unlikely(order)) {
1004 bool compound = PageCompound(page);
1007 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1009 for (i = 1; i < (1 << order); i++) {
1011 bad += free_tail_pages_check(page, page + i);
1012 if (unlikely(free_pages_check(page + i))) {
1016 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1019 if (PageMappingFlags(page))
1020 page->mapping = NULL;
1022 bad += free_pages_check(page);
1026 page_cpupid_reset_last(page);
1027 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1028 reset_page_owner(page, order);
1030 if (!PageHighMem(page)) {
1031 debug_check_no_locks_freed(page_address(page),
1032 PAGE_SIZE << order);
1033 debug_check_no_obj_freed(page_address(page),
1034 PAGE_SIZE << order);
1036 arch_free_page(page, order);
1037 kernel_poison_pages(page, 1 << order, 0);
1038 kernel_map_pages(page, 1 << order, 0);
1039 kasan_free_pages(page, order);
1044 #ifdef CONFIG_DEBUG_VM
1045 static inline bool free_pcp_prepare(struct page *page)
1047 return free_pages_prepare(page, 0, true);
1050 static inline bool bulkfree_pcp_prepare(struct page *page)
1055 static bool free_pcp_prepare(struct page *page)
1057 return free_pages_prepare(page, 0, false);
1060 static bool bulkfree_pcp_prepare(struct page *page)
1062 return free_pages_check(page);
1064 #endif /* CONFIG_DEBUG_VM */
1067 * Frees a number of pages from the PCP lists
1068 * Assumes all pages on list are in same zone, and of same order.
1069 * count is the number of pages to free.
1071 * If the zone was previously in an "all pages pinned" state then look to
1072 * see if this freeing clears that state.
1074 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1075 * pinned" detection logic.
1077 static void free_pcppages_bulk(struct zone *zone, int count,
1078 struct per_cpu_pages *pcp)
1080 int migratetype = 0;
1082 unsigned long nr_scanned;
1083 bool isolated_pageblocks;
1085 spin_lock(&zone->lock);
1086 isolated_pageblocks = has_isolate_pageblock(zone);
1087 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1089 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1093 struct list_head *list;
1096 * Remove pages from lists in a round-robin fashion. A
1097 * batch_free count is maintained that is incremented when an
1098 * empty list is encountered. This is so more pages are freed
1099 * off fuller lists instead of spinning excessively around empty
1104 if (++migratetype == MIGRATE_PCPTYPES)
1106 list = &pcp->lists[migratetype];
1107 } while (list_empty(list));
1109 /* This is the only non-empty list. Free them all. */
1110 if (batch_free == MIGRATE_PCPTYPES)
1114 int mt; /* migratetype of the to-be-freed page */
1116 page = list_last_entry(list, struct page, lru);
1117 /* must delete as __free_one_page list manipulates */
1118 list_del(&page->lru);
1120 mt = get_pcppage_migratetype(page);
1121 /* MIGRATE_ISOLATE page should not go to pcplists */
1122 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1123 /* Pageblock could have been isolated meanwhile */
1124 if (unlikely(isolated_pageblocks))
1125 mt = get_pageblock_migratetype(page);
1127 if (bulkfree_pcp_prepare(page))
1130 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1131 trace_mm_page_pcpu_drain(page, 0, mt);
1132 } while (--count && --batch_free && !list_empty(list));
1134 spin_unlock(&zone->lock);
1137 static void free_one_page(struct zone *zone,
1138 struct page *page, unsigned long pfn,
1142 unsigned long nr_scanned;
1143 spin_lock(&zone->lock);
1144 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1146 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1148 if (unlikely(has_isolate_pageblock(zone) ||
1149 is_migrate_isolate(migratetype))) {
1150 migratetype = get_pfnblock_migratetype(page, pfn);
1152 __free_one_page(page, pfn, zone, order, migratetype);
1153 spin_unlock(&zone->lock);
1156 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1157 unsigned long zone, int nid)
1159 set_page_links(page, zone, nid, pfn);
1160 init_page_count(page);
1161 page_mapcount_reset(page);
1162 page_cpupid_reset_last(page);
1164 INIT_LIST_HEAD(&page->lru);
1165 #ifdef WANT_PAGE_VIRTUAL
1166 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1167 if (!is_highmem_idx(zone))
1168 set_page_address(page, __va(pfn << PAGE_SHIFT));
1172 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1175 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1178 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1179 static void init_reserved_page(unsigned long pfn)
1184 if (!early_page_uninitialised(pfn))
1187 nid = early_pfn_to_nid(pfn);
1188 pgdat = NODE_DATA(nid);
1190 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1191 struct zone *zone = &pgdat->node_zones[zid];
1193 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1196 __init_single_pfn(pfn, zid, nid);
1199 static inline void init_reserved_page(unsigned long pfn)
1202 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1205 * Initialised pages do not have PageReserved set. This function is
1206 * called for each range allocated by the bootmem allocator and
1207 * marks the pages PageReserved. The remaining valid pages are later
1208 * sent to the buddy page allocator.
1210 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1212 unsigned long start_pfn = PFN_DOWN(start);
1213 unsigned long end_pfn = PFN_UP(end);
1215 for (; start_pfn < end_pfn; start_pfn++) {
1216 if (pfn_valid(start_pfn)) {
1217 struct page *page = pfn_to_page(start_pfn);
1219 init_reserved_page(start_pfn);
1221 /* Avoid false-positive PageTail() */
1222 INIT_LIST_HEAD(&page->lru);
1224 SetPageReserved(page);
1229 static void __free_pages_ok(struct page *page, unsigned int order)
1231 unsigned long flags;
1233 unsigned long pfn = page_to_pfn(page);
1235 if (!free_pages_prepare(page, order, true))
1238 migratetype = get_pfnblock_migratetype(page, pfn);
1239 local_irq_save(flags);
1240 __count_vm_events(PGFREE, 1 << order);
1241 free_one_page(page_zone(page), page, pfn, order, migratetype);
1242 local_irq_restore(flags);
1245 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1247 unsigned int nr_pages = 1 << order;
1248 struct page *p = page;
1252 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1254 __ClearPageReserved(p);
1255 set_page_count(p, 0);
1257 __ClearPageReserved(p);
1258 set_page_count(p, 0);
1260 page_zone(page)->managed_pages += nr_pages;
1261 set_page_refcounted(page);
1262 __free_pages(page, order);
1265 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1266 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1268 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1270 int __meminit early_pfn_to_nid(unsigned long pfn)
1272 static DEFINE_SPINLOCK(early_pfn_lock);
1275 spin_lock(&early_pfn_lock);
1276 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1278 nid = first_online_node;
1279 spin_unlock(&early_pfn_lock);
1285 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1286 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1287 struct mminit_pfnnid_cache *state)
1291 nid = __early_pfn_to_nid(pfn, state);
1292 if (nid >= 0 && nid != node)
1297 /* Only safe to use early in boot when initialisation is single-threaded */
1298 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1300 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1305 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1309 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1310 struct mminit_pfnnid_cache *state)
1317 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1320 if (early_page_uninitialised(pfn))
1322 return __free_pages_boot_core(page, order);
1326 * Check that the whole (or subset of) a pageblock given by the interval of
1327 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1328 * with the migration of free compaction scanner. The scanners then need to
1329 * use only pfn_valid_within() check for arches that allow holes within
1332 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1334 * It's possible on some configurations to have a setup like node0 node1 node0
1335 * i.e. it's possible that all pages within a zones range of pages do not
1336 * belong to a single zone. We assume that a border between node0 and node1
1337 * can occur within a single pageblock, but not a node0 node1 node0
1338 * interleaving within a single pageblock. It is therefore sufficient to check
1339 * the first and last page of a pageblock and avoid checking each individual
1340 * page in a pageblock.
1342 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1343 unsigned long end_pfn, struct zone *zone)
1345 struct page *start_page;
1346 struct page *end_page;
1348 /* end_pfn is one past the range we are checking */
1351 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1354 start_page = pfn_to_page(start_pfn);
1356 if (page_zone(start_page) != zone)
1359 end_page = pfn_to_page(end_pfn);
1361 /* This gives a shorter code than deriving page_zone(end_page) */
1362 if (page_zone_id(start_page) != page_zone_id(end_page))
1368 void set_zone_contiguous(struct zone *zone)
1370 unsigned long block_start_pfn = zone->zone_start_pfn;
1371 unsigned long block_end_pfn;
1373 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1374 for (; block_start_pfn < zone_end_pfn(zone);
1375 block_start_pfn = block_end_pfn,
1376 block_end_pfn += pageblock_nr_pages) {
1378 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1380 if (!__pageblock_pfn_to_page(block_start_pfn,
1381 block_end_pfn, zone))
1385 /* We confirm that there is no hole */
1386 zone->contiguous = true;
1389 void clear_zone_contiguous(struct zone *zone)
1391 zone->contiguous = false;
1394 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1395 static void __init deferred_free_range(struct page *page,
1396 unsigned long pfn, int nr_pages)
1403 /* Free a large naturally-aligned chunk if possible */
1404 if (nr_pages == MAX_ORDER_NR_PAGES &&
1405 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1406 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1407 __free_pages_boot_core(page, MAX_ORDER-1);
1411 for (i = 0; i < nr_pages; i++, page++)
1412 __free_pages_boot_core(page, 0);
1415 /* Completion tracking for deferred_init_memmap() threads */
1416 static atomic_t pgdat_init_n_undone __initdata;
1417 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1419 static inline void __init pgdat_init_report_one_done(void)
1421 if (atomic_dec_and_test(&pgdat_init_n_undone))
1422 complete(&pgdat_init_all_done_comp);
1425 /* Initialise remaining memory on a node */
1426 static int __init deferred_init_memmap(void *data)
1428 pg_data_t *pgdat = data;
1429 int nid = pgdat->node_id;
1430 struct mminit_pfnnid_cache nid_init_state = { };
1431 unsigned long start = jiffies;
1432 unsigned long nr_pages = 0;
1433 unsigned long walk_start, walk_end;
1436 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1437 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1439 if (first_init_pfn == ULONG_MAX) {
1440 pgdat_init_report_one_done();
1444 /* Bind memory initialisation thread to a local node if possible */
1445 if (!cpumask_empty(cpumask))
1446 set_cpus_allowed_ptr(current, cpumask);
1448 /* Sanity check boundaries */
1449 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1450 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1451 pgdat->first_deferred_pfn = ULONG_MAX;
1453 /* Only the highest zone is deferred so find it */
1454 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1455 zone = pgdat->node_zones + zid;
1456 if (first_init_pfn < zone_end_pfn(zone))
1460 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1461 unsigned long pfn, end_pfn;
1462 struct page *page = NULL;
1463 struct page *free_base_page = NULL;
1464 unsigned long free_base_pfn = 0;
1467 end_pfn = min(walk_end, zone_end_pfn(zone));
1468 pfn = first_init_pfn;
1469 if (pfn < walk_start)
1471 if (pfn < zone->zone_start_pfn)
1472 pfn = zone->zone_start_pfn;
1474 for (; pfn < end_pfn; pfn++) {
1475 if (!pfn_valid_within(pfn))
1479 * Ensure pfn_valid is checked every
1480 * MAX_ORDER_NR_PAGES for memory holes
1482 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1483 if (!pfn_valid(pfn)) {
1489 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1494 /* Minimise pfn page lookups and scheduler checks */
1495 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1498 nr_pages += nr_to_free;
1499 deferred_free_range(free_base_page,
1500 free_base_pfn, nr_to_free);
1501 free_base_page = NULL;
1502 free_base_pfn = nr_to_free = 0;
1504 page = pfn_to_page(pfn);
1509 VM_BUG_ON(page_zone(page) != zone);
1513 __init_single_page(page, pfn, zid, nid);
1514 if (!free_base_page) {
1515 free_base_page = page;
1516 free_base_pfn = pfn;
1521 /* Where possible, batch up pages for a single free */
1524 /* Free the current block of pages to allocator */
1525 nr_pages += nr_to_free;
1526 deferred_free_range(free_base_page, free_base_pfn,
1528 free_base_page = NULL;
1529 free_base_pfn = nr_to_free = 0;
1532 first_init_pfn = max(end_pfn, first_init_pfn);
1535 /* Sanity check that the next zone really is unpopulated */
1536 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1538 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1539 jiffies_to_msecs(jiffies - start));
1541 pgdat_init_report_one_done();
1544 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1546 void __init page_alloc_init_late(void)
1550 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1553 /* There will be num_node_state(N_MEMORY) threads */
1554 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1555 for_each_node_state(nid, N_MEMORY) {
1556 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1559 /* Block until all are initialised */
1560 wait_for_completion(&pgdat_init_all_done_comp);
1562 /* Reinit limits that are based on free pages after the kernel is up */
1563 files_maxfiles_init();
1566 for_each_populated_zone(zone)
1567 set_zone_contiguous(zone);
1571 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1572 void __init init_cma_reserved_pageblock(struct page *page)
1574 unsigned i = pageblock_nr_pages;
1575 struct page *p = page;
1578 __ClearPageReserved(p);
1579 set_page_count(p, 0);
1582 set_pageblock_migratetype(page, MIGRATE_CMA);
1584 if (pageblock_order >= MAX_ORDER) {
1585 i = pageblock_nr_pages;
1588 set_page_refcounted(p);
1589 __free_pages(p, MAX_ORDER - 1);
1590 p += MAX_ORDER_NR_PAGES;
1591 } while (i -= MAX_ORDER_NR_PAGES);
1593 set_page_refcounted(page);
1594 __free_pages(page, pageblock_order);
1597 adjust_managed_page_count(page, pageblock_nr_pages);
1602 * The order of subdivision here is critical for the IO subsystem.
1603 * Please do not alter this order without good reasons and regression
1604 * testing. Specifically, as large blocks of memory are subdivided,
1605 * the order in which smaller blocks are delivered depends on the order
1606 * they're subdivided in this function. This is the primary factor
1607 * influencing the order in which pages are delivered to the IO
1608 * subsystem according to empirical testing, and this is also justified
1609 * by considering the behavior of a buddy system containing a single
1610 * large block of memory acted on by a series of small allocations.
1611 * This behavior is a critical factor in sglist merging's success.
1615 static inline void expand(struct zone *zone, struct page *page,
1616 int low, int high, struct free_area *area,
1619 unsigned long size = 1 << high;
1621 while (high > low) {
1625 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1627 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1628 debug_guardpage_enabled() &&
1629 high < debug_guardpage_minorder()) {
1631 * Mark as guard pages (or page), that will allow to
1632 * merge back to allocator when buddy will be freed.
1633 * Corresponding page table entries will not be touched,
1634 * pages will stay not present in virtual address space
1636 set_page_guard(zone, &page[size], high, migratetype);
1639 list_add(&page[size].lru, &area->free_list[migratetype]);
1641 set_page_order(&page[size], high);
1645 static void check_new_page_bad(struct page *page)
1647 const char *bad_reason = NULL;
1648 unsigned long bad_flags = 0;
1650 if (unlikely(atomic_read(&page->_mapcount) != -1))
1651 bad_reason = "nonzero mapcount";
1652 if (unlikely(page->mapping != NULL))
1653 bad_reason = "non-NULL mapping";
1654 if (unlikely(page_ref_count(page) != 0))
1655 bad_reason = "nonzero _count";
1656 if (unlikely(page->flags & __PG_HWPOISON)) {
1657 bad_reason = "HWPoisoned (hardware-corrupted)";
1658 bad_flags = __PG_HWPOISON;
1659 /* Don't complain about hwpoisoned pages */
1660 page_mapcount_reset(page); /* remove PageBuddy */
1663 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1664 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1665 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1668 if (unlikely(page->mem_cgroup))
1669 bad_reason = "page still charged to cgroup";
1671 bad_page(page, bad_reason, bad_flags);
1675 * This page is about to be returned from the page allocator
1677 static inline int check_new_page(struct page *page)
1679 if (likely(page_expected_state(page,
1680 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1683 check_new_page_bad(page);
1687 static inline bool free_pages_prezeroed(bool poisoned)
1689 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1690 page_poisoning_enabled() && poisoned;
1693 #ifdef CONFIG_DEBUG_VM
1694 static bool check_pcp_refill(struct page *page)
1699 static bool check_new_pcp(struct page *page)
1701 return check_new_page(page);
1704 static bool check_pcp_refill(struct page *page)
1706 return check_new_page(page);
1708 static bool check_new_pcp(struct page *page)
1712 #endif /* CONFIG_DEBUG_VM */
1714 static bool check_new_pages(struct page *page, unsigned int order)
1717 for (i = 0; i < (1 << order); i++) {
1718 struct page *p = page + i;
1720 if (unlikely(check_new_page(p)))
1727 inline void post_alloc_hook(struct page *page, unsigned int order,
1730 set_page_private(page, 0);
1731 set_page_refcounted(page);
1733 arch_alloc_page(page, order);
1734 kernel_map_pages(page, 1 << order, 1);
1735 kernel_poison_pages(page, 1 << order, 1);
1736 kasan_alloc_pages(page, order);
1737 set_page_owner(page, order, gfp_flags);
1740 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1741 unsigned int alloc_flags)
1744 bool poisoned = true;
1746 for (i = 0; i < (1 << order); i++) {
1747 struct page *p = page + i;
1749 poisoned &= page_is_poisoned(p);
1752 post_alloc_hook(page, order, gfp_flags);
1754 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1755 for (i = 0; i < (1 << order); i++)
1756 clear_highpage(page + i);
1758 if (order && (gfp_flags & __GFP_COMP))
1759 prep_compound_page(page, order);
1762 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1763 * allocate the page. The expectation is that the caller is taking
1764 * steps that will free more memory. The caller should avoid the page
1765 * being used for !PFMEMALLOC purposes.
1767 if (alloc_flags & ALLOC_NO_WATERMARKS)
1768 set_page_pfmemalloc(page);
1770 clear_page_pfmemalloc(page);
1774 * Go through the free lists for the given migratetype and remove
1775 * the smallest available page from the freelists
1778 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1781 unsigned int current_order;
1782 struct free_area *area;
1785 /* Find a page of the appropriate size in the preferred list */
1786 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1787 area = &(zone->free_area[current_order]);
1788 page = list_first_entry_or_null(&area->free_list[migratetype],
1792 list_del(&page->lru);
1793 rmv_page_order(page);
1795 expand(zone, page, order, current_order, area, migratetype);
1796 set_pcppage_migratetype(page, migratetype);
1805 * This array describes the order lists are fallen back to when
1806 * the free lists for the desirable migrate type are depleted
1808 static int fallbacks[MIGRATE_TYPES][4] = {
1809 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1810 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1811 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1813 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1815 #ifdef CONFIG_MEMORY_ISOLATION
1816 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1821 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1824 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1827 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1828 unsigned int order) { return NULL; }
1832 * Move the free pages in a range to the free lists of the requested type.
1833 * Note that start_page and end_pages are not aligned on a pageblock
1834 * boundary. If alignment is required, use move_freepages_block()
1836 int move_freepages(struct zone *zone,
1837 struct page *start_page, struct page *end_page,
1842 int pages_moved = 0;
1844 #ifndef CONFIG_HOLES_IN_ZONE
1846 * page_zone is not safe to call in this context when
1847 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1848 * anyway as we check zone boundaries in move_freepages_block().
1849 * Remove at a later date when no bug reports exist related to
1850 * grouping pages by mobility
1852 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1855 for (page = start_page; page <= end_page;) {
1856 /* Make sure we are not inadvertently changing nodes */
1857 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1859 if (!pfn_valid_within(page_to_pfn(page))) {
1864 if (!PageBuddy(page)) {
1869 order = page_order(page);
1870 list_move(&page->lru,
1871 &zone->free_area[order].free_list[migratetype]);
1873 pages_moved += 1 << order;
1879 int move_freepages_block(struct zone *zone, struct page *page,
1882 unsigned long start_pfn, end_pfn;
1883 struct page *start_page, *end_page;
1885 start_pfn = page_to_pfn(page);
1886 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1887 start_page = pfn_to_page(start_pfn);
1888 end_page = start_page + pageblock_nr_pages - 1;
1889 end_pfn = start_pfn + pageblock_nr_pages - 1;
1891 /* Do not cross zone boundaries */
1892 if (!zone_spans_pfn(zone, start_pfn))
1894 if (!zone_spans_pfn(zone, end_pfn))
1897 return move_freepages(zone, start_page, end_page, migratetype);
1900 static void change_pageblock_range(struct page *pageblock_page,
1901 int start_order, int migratetype)
1903 int nr_pageblocks = 1 << (start_order - pageblock_order);
1905 while (nr_pageblocks--) {
1906 set_pageblock_migratetype(pageblock_page, migratetype);
1907 pageblock_page += pageblock_nr_pages;
1912 * When we are falling back to another migratetype during allocation, try to
1913 * steal extra free pages from the same pageblocks to satisfy further
1914 * allocations, instead of polluting multiple pageblocks.
1916 * If we are stealing a relatively large buddy page, it is likely there will
1917 * be more free pages in the pageblock, so try to steal them all. For
1918 * reclaimable and unmovable allocations, we steal regardless of page size,
1919 * as fragmentation caused by those allocations polluting movable pageblocks
1920 * is worse than movable allocations stealing from unmovable and reclaimable
1923 static bool can_steal_fallback(unsigned int order, int start_mt)
1926 * Leaving this order check is intended, although there is
1927 * relaxed order check in next check. The reason is that
1928 * we can actually steal whole pageblock if this condition met,
1929 * but, below check doesn't guarantee it and that is just heuristic
1930 * so could be changed anytime.
1932 if (order >= pageblock_order)
1935 if (order >= pageblock_order / 2 ||
1936 start_mt == MIGRATE_RECLAIMABLE ||
1937 start_mt == MIGRATE_UNMOVABLE ||
1938 page_group_by_mobility_disabled)
1945 * This function implements actual steal behaviour. If order is large enough,
1946 * we can steal whole pageblock. If not, we first move freepages in this
1947 * pageblock and check whether half of pages are moved or not. If half of
1948 * pages are moved, we can change migratetype of pageblock and permanently
1949 * use it's pages as requested migratetype in the future.
1951 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1954 unsigned int current_order = page_order(page);
1957 /* Take ownership for orders >= pageblock_order */
1958 if (current_order >= pageblock_order) {
1959 change_pageblock_range(page, current_order, start_type);
1963 pages = move_freepages_block(zone, page, start_type);
1965 /* Claim the whole block if over half of it is free */
1966 if (pages >= (1 << (pageblock_order-1)) ||
1967 page_group_by_mobility_disabled)
1968 set_pageblock_migratetype(page, start_type);
1972 * Check whether there is a suitable fallback freepage with requested order.
1973 * If only_stealable is true, this function returns fallback_mt only if
1974 * we can steal other freepages all together. This would help to reduce
1975 * fragmentation due to mixed migratetype pages in one pageblock.
1977 int find_suitable_fallback(struct free_area *area, unsigned int order,
1978 int migratetype, bool only_stealable, bool *can_steal)
1983 if (area->nr_free == 0)
1988 fallback_mt = fallbacks[migratetype][i];
1989 if (fallback_mt == MIGRATE_TYPES)
1992 if (list_empty(&area->free_list[fallback_mt]))
1995 if (can_steal_fallback(order, migratetype))
1998 if (!only_stealable)
2009 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2010 * there are no empty page blocks that contain a page with a suitable order
2012 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2013 unsigned int alloc_order)
2016 unsigned long max_managed, flags;
2019 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2020 * Check is race-prone but harmless.
2022 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2023 if (zone->nr_reserved_highatomic >= max_managed)
2026 spin_lock_irqsave(&zone->lock, flags);
2028 /* Recheck the nr_reserved_highatomic limit under the lock */
2029 if (zone->nr_reserved_highatomic >= max_managed)
2033 mt = get_pageblock_migratetype(page);
2034 if (mt != MIGRATE_HIGHATOMIC &&
2035 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2036 zone->nr_reserved_highatomic += pageblock_nr_pages;
2037 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2038 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2042 spin_unlock_irqrestore(&zone->lock, flags);
2046 * Used when an allocation is about to fail under memory pressure. This
2047 * potentially hurts the reliability of high-order allocations when under
2048 * intense memory pressure but failed atomic allocations should be easier
2049 * to recover from than an OOM.
2051 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2053 struct zonelist *zonelist = ac->zonelist;
2054 unsigned long flags;
2060 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2062 /* Preserve at least one pageblock */
2063 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2066 spin_lock_irqsave(&zone->lock, flags);
2067 for (order = 0; order < MAX_ORDER; order++) {
2068 struct free_area *area = &(zone->free_area[order]);
2070 page = list_first_entry_or_null(
2071 &area->free_list[MIGRATE_HIGHATOMIC],
2077 * It should never happen but changes to locking could
2078 * inadvertently allow a per-cpu drain to add pages
2079 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2080 * and watch for underflows.
2082 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2083 zone->nr_reserved_highatomic);
2086 * Convert to ac->migratetype and avoid the normal
2087 * pageblock stealing heuristics. Minimally, the caller
2088 * is doing the work and needs the pages. More
2089 * importantly, if the block was always converted to
2090 * MIGRATE_UNMOVABLE or another type then the number
2091 * of pageblocks that cannot be completely freed
2094 set_pageblock_migratetype(page, ac->migratetype);
2095 move_freepages_block(zone, page, ac->migratetype);
2096 spin_unlock_irqrestore(&zone->lock, flags);
2099 spin_unlock_irqrestore(&zone->lock, flags);
2103 /* Remove an element from the buddy allocator from the fallback list */
2104 static inline struct page *
2105 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2107 struct free_area *area;
2108 unsigned int current_order;
2113 /* Find the largest possible block of pages in the other list */
2114 for (current_order = MAX_ORDER-1;
2115 current_order >= order && current_order <= MAX_ORDER-1;
2117 area = &(zone->free_area[current_order]);
2118 fallback_mt = find_suitable_fallback(area, current_order,
2119 start_migratetype, false, &can_steal);
2120 if (fallback_mt == -1)
2123 page = list_first_entry(&area->free_list[fallback_mt],
2126 steal_suitable_fallback(zone, page, start_migratetype);
2128 /* Remove the page from the freelists */
2130 list_del(&page->lru);
2131 rmv_page_order(page);
2133 expand(zone, page, order, current_order, area,
2136 * The pcppage_migratetype may differ from pageblock's
2137 * migratetype depending on the decisions in
2138 * find_suitable_fallback(). This is OK as long as it does not
2139 * differ for MIGRATE_CMA pageblocks. Those can be used as
2140 * fallback only via special __rmqueue_cma_fallback() function
2142 set_pcppage_migratetype(page, start_migratetype);
2144 trace_mm_page_alloc_extfrag(page, order, current_order,
2145 start_migratetype, fallback_mt);
2154 * Do the hard work of removing an element from the buddy allocator.
2155 * Call me with the zone->lock already held.
2157 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2162 page = __rmqueue_smallest(zone, order, migratetype);
2163 if (unlikely(!page)) {
2164 if (migratetype == MIGRATE_MOVABLE)
2165 page = __rmqueue_cma_fallback(zone, order);
2168 page = __rmqueue_fallback(zone, order, migratetype);
2171 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2176 * Obtain a specified number of elements from the buddy allocator, all under
2177 * a single hold of the lock, for efficiency. Add them to the supplied list.
2178 * Returns the number of new pages which were placed at *list.
2180 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2181 unsigned long count, struct list_head *list,
2182 int migratetype, bool cold)
2186 spin_lock(&zone->lock);
2187 for (i = 0; i < count; ++i) {
2188 struct page *page = __rmqueue(zone, order, migratetype);
2189 if (unlikely(page == NULL))
2192 if (unlikely(check_pcp_refill(page)))
2196 * Split buddy pages returned by expand() are received here
2197 * in physical page order. The page is added to the callers and
2198 * list and the list head then moves forward. From the callers
2199 * perspective, the linked list is ordered by page number in
2200 * some conditions. This is useful for IO devices that can
2201 * merge IO requests if the physical pages are ordered
2205 list_add(&page->lru, list);
2207 list_add_tail(&page->lru, list);
2209 if (is_migrate_cma(get_pcppage_migratetype(page)))
2210 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2213 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2214 spin_unlock(&zone->lock);
2220 * Called from the vmstat counter updater to drain pagesets of this
2221 * currently executing processor on remote nodes after they have
2224 * Note that this function must be called with the thread pinned to
2225 * a single processor.
2227 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2229 unsigned long flags;
2230 int to_drain, batch;
2232 local_irq_save(flags);
2233 batch = READ_ONCE(pcp->batch);
2234 to_drain = min(pcp->count, batch);
2236 free_pcppages_bulk(zone, to_drain, pcp);
2237 pcp->count -= to_drain;
2239 local_irq_restore(flags);
2244 * Drain pcplists of the indicated processor and zone.
2246 * The processor must either be the current processor and the
2247 * thread pinned to the current processor or a processor that
2250 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2252 unsigned long flags;
2253 struct per_cpu_pageset *pset;
2254 struct per_cpu_pages *pcp;
2256 local_irq_save(flags);
2257 pset = per_cpu_ptr(zone->pageset, cpu);
2261 free_pcppages_bulk(zone, pcp->count, pcp);
2264 local_irq_restore(flags);
2268 * Drain pcplists of all zones on the indicated processor.
2270 * The processor must either be the current processor and the
2271 * thread pinned to the current processor or a processor that
2274 static void drain_pages(unsigned int cpu)
2278 for_each_populated_zone(zone) {
2279 drain_pages_zone(cpu, zone);
2284 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2286 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2287 * the single zone's pages.
2289 void drain_local_pages(struct zone *zone)
2291 int cpu = smp_processor_id();
2294 drain_pages_zone(cpu, zone);
2300 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2302 * When zone parameter is non-NULL, spill just the single zone's pages.
2304 * Note that this code is protected against sending an IPI to an offline
2305 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2306 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2307 * nothing keeps CPUs from showing up after we populated the cpumask and
2308 * before the call to on_each_cpu_mask().
2310 void drain_all_pages(struct zone *zone)
2315 * Allocate in the BSS so we wont require allocation in
2316 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2318 static cpumask_t cpus_with_pcps;
2321 * We don't care about racing with CPU hotplug event
2322 * as offline notification will cause the notified
2323 * cpu to drain that CPU pcps and on_each_cpu_mask
2324 * disables preemption as part of its processing
2326 for_each_online_cpu(cpu) {
2327 struct per_cpu_pageset *pcp;
2329 bool has_pcps = false;
2332 pcp = per_cpu_ptr(zone->pageset, cpu);
2336 for_each_populated_zone(z) {
2337 pcp = per_cpu_ptr(z->pageset, cpu);
2338 if (pcp->pcp.count) {
2346 cpumask_set_cpu(cpu, &cpus_with_pcps);
2348 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2350 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2354 #ifdef CONFIG_HIBERNATION
2356 void mark_free_pages(struct zone *zone)
2358 unsigned long pfn, max_zone_pfn;
2359 unsigned long flags;
2360 unsigned int order, t;
2363 if (zone_is_empty(zone))
2366 spin_lock_irqsave(&zone->lock, flags);
2368 max_zone_pfn = zone_end_pfn(zone);
2369 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2370 if (pfn_valid(pfn)) {
2371 page = pfn_to_page(pfn);
2373 if (page_zone(page) != zone)
2376 if (!swsusp_page_is_forbidden(page))
2377 swsusp_unset_page_free(page);
2380 for_each_migratetype_order(order, t) {
2381 list_for_each_entry(page,
2382 &zone->free_area[order].free_list[t], lru) {
2385 pfn = page_to_pfn(page);
2386 for (i = 0; i < (1UL << order); i++)
2387 swsusp_set_page_free(pfn_to_page(pfn + i));
2390 spin_unlock_irqrestore(&zone->lock, flags);
2392 #endif /* CONFIG_PM */
2395 * Free a 0-order page
2396 * cold == true ? free a cold page : free a hot page
2398 void free_hot_cold_page(struct page *page, bool cold)
2400 struct zone *zone = page_zone(page);
2401 struct per_cpu_pages *pcp;
2402 unsigned long flags;
2403 unsigned long pfn = page_to_pfn(page);
2406 if (!free_pcp_prepare(page))
2409 migratetype = get_pfnblock_migratetype(page, pfn);
2410 set_pcppage_migratetype(page, migratetype);
2411 local_irq_save(flags);
2412 __count_vm_event(PGFREE);
2415 * We only track unmovable, reclaimable and movable on pcp lists.
2416 * Free ISOLATE pages back to the allocator because they are being
2417 * offlined but treat RESERVE as movable pages so we can get those
2418 * areas back if necessary. Otherwise, we may have to free
2419 * excessively into the page allocator
2421 if (migratetype >= MIGRATE_PCPTYPES) {
2422 if (unlikely(is_migrate_isolate(migratetype))) {
2423 free_one_page(zone, page, pfn, 0, migratetype);
2426 migratetype = MIGRATE_MOVABLE;
2429 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2431 list_add(&page->lru, &pcp->lists[migratetype]);
2433 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2435 if (pcp->count >= pcp->high) {
2436 unsigned long batch = READ_ONCE(pcp->batch);
2437 free_pcppages_bulk(zone, batch, pcp);
2438 pcp->count -= batch;
2442 local_irq_restore(flags);
2446 * Free a list of 0-order pages
2448 void free_hot_cold_page_list(struct list_head *list, bool cold)
2450 struct page *page, *next;
2452 list_for_each_entry_safe(page, next, list, lru) {
2453 trace_mm_page_free_batched(page, cold);
2454 free_hot_cold_page(page, cold);
2459 * split_page takes a non-compound higher-order page, and splits it into
2460 * n (1<<order) sub-pages: page[0..n]
2461 * Each sub-page must be freed individually.
2463 * Note: this is probably too low level an operation for use in drivers.
2464 * Please consult with lkml before using this in your driver.
2466 void split_page(struct page *page, unsigned int order)
2470 VM_BUG_ON_PAGE(PageCompound(page), page);
2471 VM_BUG_ON_PAGE(!page_count(page), page);
2473 #ifdef CONFIG_KMEMCHECK
2475 * Split shadow pages too, because free(page[0]) would
2476 * otherwise free the whole shadow.
2478 if (kmemcheck_page_is_tracked(page))
2479 split_page(virt_to_page(page[0].shadow), order);
2482 for (i = 1; i < (1 << order); i++)
2483 set_page_refcounted(page + i);
2484 split_page_owner(page, order);
2486 EXPORT_SYMBOL_GPL(split_page);
2488 int __isolate_free_page(struct page *page, unsigned int order)
2490 unsigned long watermark;
2494 BUG_ON(!PageBuddy(page));
2496 zone = page_zone(page);
2497 mt = get_pageblock_migratetype(page);
2499 if (!is_migrate_isolate(mt)) {
2500 /* Obey watermarks as if the page was being allocated */
2501 watermark = low_wmark_pages(zone) + (1 << order);
2502 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2505 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2508 /* Remove page from free list */
2509 list_del(&page->lru);
2510 zone->free_area[order].nr_free--;
2511 rmv_page_order(page);
2513 /* Set the pageblock if the isolated page is at least a pageblock */
2514 if (order >= pageblock_order - 1) {
2515 struct page *endpage = page + (1 << order) - 1;
2516 for (; page < endpage; page += pageblock_nr_pages) {
2517 int mt = get_pageblock_migratetype(page);
2518 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2519 set_pageblock_migratetype(page,
2525 return 1UL << order;
2529 * Update NUMA hit/miss statistics
2531 * Must be called with interrupts disabled.
2533 * When __GFP_OTHER_NODE is set assume the node of the preferred
2534 * zone is the local node. This is useful for daemons who allocate
2535 * memory on behalf of other processes.
2537 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2541 int local_nid = numa_node_id();
2542 enum zone_stat_item local_stat = NUMA_LOCAL;
2544 if (unlikely(flags & __GFP_OTHER_NODE)) {
2545 local_stat = NUMA_OTHER;
2546 local_nid = preferred_zone->node;
2549 if (z->node == local_nid) {
2550 __inc_zone_state(z, NUMA_HIT);
2551 __inc_zone_state(z, local_stat);
2553 __inc_zone_state(z, NUMA_MISS);
2554 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2560 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2563 struct page *buffered_rmqueue(struct zone *preferred_zone,
2564 struct zone *zone, unsigned int order,
2565 gfp_t gfp_flags, unsigned int alloc_flags,
2568 unsigned long flags;
2570 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2572 if (likely(order == 0)) {
2573 struct per_cpu_pages *pcp;
2574 struct list_head *list;
2576 local_irq_save(flags);
2578 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2579 list = &pcp->lists[migratetype];
2580 if (list_empty(list)) {
2581 pcp->count += rmqueue_bulk(zone, 0,
2584 if (unlikely(list_empty(list)))
2589 page = list_last_entry(list, struct page, lru);
2591 page = list_first_entry(list, struct page, lru);
2593 __dec_zone_state(zone, NR_ALLOC_BATCH);
2594 list_del(&page->lru);
2597 } while (check_new_pcp(page));
2600 * We most definitely don't want callers attempting to
2601 * allocate greater than order-1 page units with __GFP_NOFAIL.
2603 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2604 spin_lock_irqsave(&zone->lock, flags);
2608 if (alloc_flags & ALLOC_HARDER) {
2609 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2611 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2614 page = __rmqueue(zone, order, migratetype);
2615 } while (page && check_new_pages(page, order));
2616 spin_unlock(&zone->lock);
2619 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2620 __mod_zone_freepage_state(zone, -(1 << order),
2621 get_pcppage_migratetype(page));
2624 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2625 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2626 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2628 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2629 zone_statistics(preferred_zone, zone, gfp_flags);
2630 local_irq_restore(flags);
2632 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2636 local_irq_restore(flags);
2640 #ifdef CONFIG_FAIL_PAGE_ALLOC
2643 struct fault_attr attr;
2645 bool ignore_gfp_highmem;
2646 bool ignore_gfp_reclaim;
2648 } fail_page_alloc = {
2649 .attr = FAULT_ATTR_INITIALIZER,
2650 .ignore_gfp_reclaim = true,
2651 .ignore_gfp_highmem = true,
2655 static int __init setup_fail_page_alloc(char *str)
2657 return setup_fault_attr(&fail_page_alloc.attr, str);
2659 __setup("fail_page_alloc=", setup_fail_page_alloc);
2661 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2663 if (order < fail_page_alloc.min_order)
2665 if (gfp_mask & __GFP_NOFAIL)
2667 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2669 if (fail_page_alloc.ignore_gfp_reclaim &&
2670 (gfp_mask & __GFP_DIRECT_RECLAIM))
2673 return should_fail(&fail_page_alloc.attr, 1 << order);
2676 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2678 static int __init fail_page_alloc_debugfs(void)
2680 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2683 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2684 &fail_page_alloc.attr);
2686 return PTR_ERR(dir);
2688 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2689 &fail_page_alloc.ignore_gfp_reclaim))
2691 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2692 &fail_page_alloc.ignore_gfp_highmem))
2694 if (!debugfs_create_u32("min-order", mode, dir,
2695 &fail_page_alloc.min_order))
2700 debugfs_remove_recursive(dir);
2705 late_initcall(fail_page_alloc_debugfs);
2707 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2709 #else /* CONFIG_FAIL_PAGE_ALLOC */
2711 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2716 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2719 * Return true if free base pages are above 'mark'. For high-order checks it
2720 * will return true of the order-0 watermark is reached and there is at least
2721 * one free page of a suitable size. Checking now avoids taking the zone lock
2722 * to check in the allocation paths if no pages are free.
2724 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2725 int classzone_idx, unsigned int alloc_flags,
2730 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2732 /* free_pages may go negative - that's OK */
2733 free_pages -= (1 << order) - 1;
2735 if (alloc_flags & ALLOC_HIGH)
2739 * If the caller does not have rights to ALLOC_HARDER then subtract
2740 * the high-atomic reserves. This will over-estimate the size of the
2741 * atomic reserve but it avoids a search.
2743 if (likely(!alloc_harder))
2744 free_pages -= z->nr_reserved_highatomic;
2749 /* If allocation can't use CMA areas don't use free CMA pages */
2750 if (!(alloc_flags & ALLOC_CMA))
2751 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2755 * Check watermarks for an order-0 allocation request. If these
2756 * are not met, then a high-order request also cannot go ahead
2757 * even if a suitable page happened to be free.
2759 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2762 /* If this is an order-0 request then the watermark is fine */
2766 /* For a high-order request, check at least one suitable page is free */
2767 for (o = order; o < MAX_ORDER; o++) {
2768 struct free_area *area = &z->free_area[o];
2777 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2778 if (!list_empty(&area->free_list[mt]))
2783 if ((alloc_flags & ALLOC_CMA) &&
2784 !list_empty(&area->free_list[MIGRATE_CMA])) {
2792 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2793 int classzone_idx, unsigned int alloc_flags)
2795 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2796 zone_page_state(z, NR_FREE_PAGES));
2799 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2800 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2802 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2806 /* If allocation can't use CMA areas don't use free CMA pages */
2807 if (!(alloc_flags & ALLOC_CMA))
2808 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2812 * Fast check for order-0 only. If this fails then the reserves
2813 * need to be calculated. There is a corner case where the check
2814 * passes but only the high-order atomic reserve are free. If
2815 * the caller is !atomic then it'll uselessly search the free
2816 * list. That corner case is then slower but it is harmless.
2818 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2821 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2825 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2826 unsigned long mark, int classzone_idx)
2828 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2830 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2831 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2833 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2838 static bool zone_local(struct zone *local_zone, struct zone *zone)
2840 return local_zone->node == zone->node;
2843 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2845 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2848 #else /* CONFIG_NUMA */
2849 static bool zone_local(struct zone *local_zone, struct zone *zone)
2854 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2858 #endif /* CONFIG_NUMA */
2860 static void reset_alloc_batches(struct zone *preferred_zone)
2862 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2865 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2866 high_wmark_pages(zone) - low_wmark_pages(zone) -
2867 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2868 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2869 } while (zone++ != preferred_zone);
2873 * get_page_from_freelist goes through the zonelist trying to allocate
2876 static struct page *
2877 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2878 const struct alloc_context *ac)
2880 struct zoneref *z = ac->preferred_zoneref;
2882 bool fair_skipped = false;
2883 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2887 * Scan zonelist, looking for a zone with enough free.
2888 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2890 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2895 if (cpusets_enabled() &&
2896 (alloc_flags & ALLOC_CPUSET) &&
2897 !__cpuset_zone_allowed(zone, gfp_mask))
2900 * Distribute pages in proportion to the individual
2901 * zone size to ensure fair page aging. The zone a
2902 * page was allocated in should have no effect on the
2903 * time the page has in memory before being reclaimed.
2906 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2907 fair_skipped = true;
2910 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2917 * When allocating a page cache page for writing, we
2918 * want to get it from a zone that is within its dirty
2919 * limit, such that no single zone holds more than its
2920 * proportional share of globally allowed dirty pages.
2921 * The dirty limits take into account the zone's
2922 * lowmem reserves and high watermark so that kswapd
2923 * should be able to balance it without having to
2924 * write pages from its LRU list.
2926 * This may look like it could increase pressure on
2927 * lower zones by failing allocations in higher zones
2928 * before they are full. But the pages that do spill
2929 * over are limited as the lower zones are protected
2930 * by this very same mechanism. It should not become
2931 * a practical burden to them.
2933 * XXX: For now, allow allocations to potentially
2934 * exceed the per-zone dirty limit in the slowpath
2935 * (spread_dirty_pages unset) before going into reclaim,
2936 * which is important when on a NUMA setup the allowed
2937 * zones are together not big enough to reach the
2938 * global limit. The proper fix for these situations
2939 * will require awareness of zones in the
2940 * dirty-throttling and the flusher threads.
2942 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2945 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2946 if (!zone_watermark_fast(zone, order, mark,
2947 ac_classzone_idx(ac), alloc_flags)) {
2950 /* Checked here to keep the fast path fast */
2951 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2952 if (alloc_flags & ALLOC_NO_WATERMARKS)
2955 if (zone_reclaim_mode == 0 ||
2956 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2959 ret = zone_reclaim(zone, gfp_mask, order);
2961 case ZONE_RECLAIM_NOSCAN:
2964 case ZONE_RECLAIM_FULL:
2965 /* scanned but unreclaimable */
2968 /* did we reclaim enough */
2969 if (zone_watermark_ok(zone, order, mark,
2970 ac_classzone_idx(ac), alloc_flags))
2978 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2979 gfp_mask, alloc_flags, ac->migratetype);
2981 prep_new_page(page, order, gfp_mask, alloc_flags);
2984 * If this is a high-order atomic allocation then check
2985 * if the pageblock should be reserved for the future
2987 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2988 reserve_highatomic_pageblock(page, zone, order);
2995 * The first pass makes sure allocations are spread fairly within the
2996 * local node. However, the local node might have free pages left
2997 * after the fairness batches are exhausted, and remote zones haven't
2998 * even been considered yet. Try once more without fairness, and
2999 * include remote zones now, before entering the slowpath and waking
3000 * kswapd: prefer spilling to a remote zone over swapping locally.
3005 fair_skipped = false;
3006 reset_alloc_batches(ac->preferred_zoneref->zone);
3007 z = ac->preferred_zoneref;
3015 * Large machines with many possible nodes should not always dump per-node
3016 * meminfo in irq context.
3018 static inline bool should_suppress_show_mem(void)
3023 ret = in_interrupt();
3028 static DEFINE_RATELIMIT_STATE(nopage_rs,
3029 DEFAULT_RATELIMIT_INTERVAL,
3030 DEFAULT_RATELIMIT_BURST);
3032 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3034 unsigned int filter = SHOW_MEM_FILTER_NODES;
3036 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3037 debug_guardpage_minorder() > 0)
3041 * This documents exceptions given to allocations in certain
3042 * contexts that are allowed to allocate outside current's set
3045 if (!(gfp_mask & __GFP_NOMEMALLOC))
3046 if (test_thread_flag(TIF_MEMDIE) ||
3047 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3048 filter &= ~SHOW_MEM_FILTER_NODES;
3049 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3050 filter &= ~SHOW_MEM_FILTER_NODES;
3053 struct va_format vaf;
3056 va_start(args, fmt);
3061 pr_warn("%pV", &vaf);
3066 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3067 current->comm, order, gfp_mask, &gfp_mask);
3069 if (!should_suppress_show_mem())
3073 static inline struct page *
3074 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3075 const struct alloc_context *ac, unsigned long *did_some_progress)
3077 struct oom_control oc = {
3078 .zonelist = ac->zonelist,
3079 .nodemask = ac->nodemask,
3081 .gfp_mask = gfp_mask,
3086 *did_some_progress = 0;
3089 * Acquire the oom lock. If that fails, somebody else is
3090 * making progress for us.
3092 if (!mutex_trylock(&oom_lock)) {
3093 *did_some_progress = 1;
3094 schedule_timeout_uninterruptible(1);
3099 * Go through the zonelist yet one more time, keep very high watermark
3100 * here, this is only to catch a parallel oom killing, we must fail if
3101 * we're still under heavy pressure.
3103 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3104 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3108 if (!(gfp_mask & __GFP_NOFAIL)) {
3109 /* Coredumps can quickly deplete all memory reserves */
3110 if (current->flags & PF_DUMPCORE)
3112 /* The OOM killer will not help higher order allocs */
3113 if (order > PAGE_ALLOC_COSTLY_ORDER)
3115 /* The OOM killer does not needlessly kill tasks for lowmem */
3116 if (ac->high_zoneidx < ZONE_NORMAL)
3118 if (pm_suspended_storage())
3121 * XXX: GFP_NOFS allocations should rather fail than rely on
3122 * other request to make a forward progress.
3123 * We are in an unfortunate situation where out_of_memory cannot
3124 * do much for this context but let's try it to at least get
3125 * access to memory reserved if the current task is killed (see
3126 * out_of_memory). Once filesystems are ready to handle allocation
3127 * failures more gracefully we should just bail out here.
3130 /* The OOM killer may not free memory on a specific node */
3131 if (gfp_mask & __GFP_THISNODE)
3134 /* Exhausted what can be done so it's blamo time */
3135 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3136 *did_some_progress = 1;
3138 if (gfp_mask & __GFP_NOFAIL) {
3139 page = get_page_from_freelist(gfp_mask, order,
3140 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3142 * fallback to ignore cpuset restriction if our nodes
3146 page = get_page_from_freelist(gfp_mask, order,
3147 ALLOC_NO_WATERMARKS, ac);
3151 mutex_unlock(&oom_lock);
3157 * Maximum number of compaction retries wit a progress before OOM
3158 * killer is consider as the only way to move forward.
3160 #define MAX_COMPACT_RETRIES 16
3162 #ifdef CONFIG_COMPACTION
3163 /* Try memory compaction for high-order allocations before reclaim */
3164 static struct page *
3165 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3166 unsigned int alloc_flags, const struct alloc_context *ac,
3167 enum migrate_mode mode, enum compact_result *compact_result)
3170 int contended_compaction;
3175 current->flags |= PF_MEMALLOC;
3176 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3177 mode, &contended_compaction);
3178 current->flags &= ~PF_MEMALLOC;
3180 if (*compact_result <= COMPACT_INACTIVE)
3184 * At least in one zone compaction wasn't deferred or skipped, so let's
3185 * count a compaction stall
3187 count_vm_event(COMPACTSTALL);
3189 page = get_page_from_freelist(gfp_mask, order,
3190 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3193 struct zone *zone = page_zone(page);
3195 zone->compact_blockskip_flush = false;
3196 compaction_defer_reset(zone, order, true);
3197 count_vm_event(COMPACTSUCCESS);
3202 * It's bad if compaction run occurs and fails. The most likely reason
3203 * is that pages exist, but not enough to satisfy watermarks.
3205 count_vm_event(COMPACTFAIL);
3208 * In all zones where compaction was attempted (and not
3209 * deferred or skipped), lock contention has been detected.
3210 * For THP allocation we do not want to disrupt the others
3211 * so we fallback to base pages instead.
3213 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3214 *compact_result = COMPACT_CONTENDED;
3217 * If compaction was aborted due to need_resched(), we do not
3218 * want to further increase allocation latency, unless it is
3219 * khugepaged trying to collapse.
3221 if (contended_compaction == COMPACT_CONTENDED_SCHED
3222 && !(current->flags & PF_KTHREAD))
3223 *compact_result = COMPACT_CONTENDED;
3231 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3232 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3233 int compaction_retries)
3235 int max_retries = MAX_COMPACT_RETRIES;
3241 * compaction considers all the zone as desperately out of memory
3242 * so it doesn't really make much sense to retry except when the
3243 * failure could be caused by weak migration mode.
3245 if (compaction_failed(compact_result)) {
3246 if (*migrate_mode == MIGRATE_ASYNC) {
3247 *migrate_mode = MIGRATE_SYNC_LIGHT;
3254 * make sure the compaction wasn't deferred or didn't bail out early
3255 * due to locks contention before we declare that we should give up.
3256 * But do not retry if the given zonelist is not suitable for
3259 if (compaction_withdrawn(compact_result))
3260 return compaction_zonelist_suitable(ac, order, alloc_flags);
3263 * !costly requests are much more important than __GFP_REPEAT
3264 * costly ones because they are de facto nofail and invoke OOM
3265 * killer to move on while costly can fail and users are ready
3266 * to cope with that. 1/4 retries is rather arbitrary but we
3267 * would need much more detailed feedback from compaction to
3268 * make a better decision.
3270 if (order > PAGE_ALLOC_COSTLY_ORDER)
3272 if (compaction_retries <= max_retries)
3278 static inline struct page *
3279 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3280 unsigned int alloc_flags, const struct alloc_context *ac,
3281 enum migrate_mode mode, enum compact_result *compact_result)
3283 *compact_result = COMPACT_SKIPPED;
3288 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3289 enum compact_result compact_result,
3290 enum migrate_mode *migrate_mode,
3291 int compaction_retries)
3296 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3300 * There are setups with compaction disabled which would prefer to loop
3301 * inside the allocator rather than hit the oom killer prematurely.
3302 * Let's give them a good hope and keep retrying while the order-0
3303 * watermarks are OK.
3305 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3307 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3308 ac_classzone_idx(ac), alloc_flags))
3313 #endif /* CONFIG_COMPACTION */
3315 /* Perform direct synchronous page reclaim */
3317 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3318 const struct alloc_context *ac)
3320 struct reclaim_state reclaim_state;
3325 /* We now go into synchronous reclaim */
3326 cpuset_memory_pressure_bump();
3327 current->flags |= PF_MEMALLOC;
3328 lockdep_set_current_reclaim_state(gfp_mask);
3329 reclaim_state.reclaimed_slab = 0;
3330 current->reclaim_state = &reclaim_state;
3332 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3335 current->reclaim_state = NULL;
3336 lockdep_clear_current_reclaim_state();
3337 current->flags &= ~PF_MEMALLOC;
3344 /* The really slow allocator path where we enter direct reclaim */
3345 static inline struct page *
3346 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3347 unsigned int alloc_flags, const struct alloc_context *ac,
3348 unsigned long *did_some_progress)
3350 struct page *page = NULL;
3351 bool drained = false;
3353 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3354 if (unlikely(!(*did_some_progress)))
3358 page = get_page_from_freelist(gfp_mask, order,
3359 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3362 * If an allocation failed after direct reclaim, it could be because
3363 * pages are pinned on the per-cpu lists or in high alloc reserves.
3364 * Shrink them them and try again
3366 if (!page && !drained) {
3367 unreserve_highatomic_pageblock(ac);
3368 drain_all_pages(NULL);
3376 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3381 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3382 ac->high_zoneidx, ac->nodemask)
3383 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3386 static inline unsigned int
3387 gfp_to_alloc_flags(gfp_t gfp_mask)
3389 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3391 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3392 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3395 * The caller may dip into page reserves a bit more if the caller
3396 * cannot run direct reclaim, or if the caller has realtime scheduling
3397 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3398 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3400 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3402 if (gfp_mask & __GFP_ATOMIC) {
3404 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3405 * if it can't schedule.
3407 if (!(gfp_mask & __GFP_NOMEMALLOC))
3408 alloc_flags |= ALLOC_HARDER;
3410 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3411 * comment for __cpuset_node_allowed().
3413 alloc_flags &= ~ALLOC_CPUSET;
3414 } else if (unlikely(rt_task(current)) && !in_interrupt())
3415 alloc_flags |= ALLOC_HARDER;
3417 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3418 if (gfp_mask & __GFP_MEMALLOC)
3419 alloc_flags |= ALLOC_NO_WATERMARKS;
3420 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3421 alloc_flags |= ALLOC_NO_WATERMARKS;
3422 else if (!in_interrupt() &&
3423 ((current->flags & PF_MEMALLOC) ||
3424 unlikely(test_thread_flag(TIF_MEMDIE))))
3425 alloc_flags |= ALLOC_NO_WATERMARKS;
3428 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3429 alloc_flags |= ALLOC_CMA;
3434 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3436 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3439 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3441 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3445 * Maximum number of reclaim retries without any progress before OOM killer
3446 * is consider as the only way to move forward.
3448 #define MAX_RECLAIM_RETRIES 16
3451 * Checks whether it makes sense to retry the reclaim to make a forward progress
3452 * for the given allocation request.
3453 * The reclaim feedback represented by did_some_progress (any progress during
3454 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3455 * any progress in a row) is considered as well as the reclaimable pages on the
3456 * applicable zone list (with a backoff mechanism which is a function of
3457 * no_progress_loops).
3459 * Returns true if a retry is viable or false to enter the oom path.
3462 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3463 struct alloc_context *ac, int alloc_flags,
3464 bool did_some_progress, int no_progress_loops)
3470 * Make sure we converge to OOM if we cannot make any progress
3471 * several times in the row.
3473 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3477 * Keep reclaiming pages while there is a chance this will lead somewhere.
3478 * If none of the target zones can satisfy our allocation request even
3479 * if all reclaimable pages are considered then we are screwed and have
3482 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3484 unsigned long available;
3485 unsigned long reclaimable;
3487 available = reclaimable = zone_reclaimable_pages(zone);
3488 available -= DIV_ROUND_UP(no_progress_loops * available,
3489 MAX_RECLAIM_RETRIES);
3490 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3493 * Would the allocation succeed if we reclaimed the whole
3496 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3497 ac_classzone_idx(ac), alloc_flags, available)) {
3499 * If we didn't make any progress and have a lot of
3500 * dirty + writeback pages then we should wait for
3501 * an IO to complete to slow down the reclaim and
3502 * prevent from pre mature OOM
3504 if (!did_some_progress) {
3505 unsigned long writeback;
3506 unsigned long dirty;
3508 writeback = zone_page_state_snapshot(zone,
3510 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3512 if (2*(writeback + dirty) > reclaimable) {
3513 congestion_wait(BLK_RW_ASYNC, HZ/10);
3519 * Memory allocation/reclaim might be called from a WQ
3520 * context and the current implementation of the WQ
3521 * concurrency control doesn't recognize that
3522 * a particular WQ is congested if the worker thread is
3523 * looping without ever sleeping. Therefore we have to
3524 * do a short sleep here rather than calling
3527 if (current->flags & PF_WQ_WORKER)
3528 schedule_timeout_uninterruptible(1);
3539 static inline struct page *
3540 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3541 struct alloc_context *ac)
3543 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3544 struct page *page = NULL;
3545 unsigned int alloc_flags;
3546 unsigned long did_some_progress;
3547 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3548 enum compact_result compact_result;
3549 int compaction_retries = 0;
3550 int no_progress_loops = 0;
3553 * In the slowpath, we sanity check order to avoid ever trying to
3554 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3555 * be using allocators in order of preference for an area that is
3558 if (order >= MAX_ORDER) {
3559 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3564 * We also sanity check to catch abuse of atomic reserves being used by
3565 * callers that are not in atomic context.
3567 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3568 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3569 gfp_mask &= ~__GFP_ATOMIC;
3572 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3573 wake_all_kswapds(order, ac);
3576 * OK, we're below the kswapd watermark and have kicked background
3577 * reclaim. Now things get more complex, so set up alloc_flags according
3578 * to how we want to proceed.
3580 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3583 * Reset the zonelist iterators if memory policies can be ignored.
3584 * These allocations are high priority and system rather than user
3587 if ((alloc_flags & ALLOC_NO_WATERMARKS) || !(alloc_flags & ALLOC_CPUSET)) {
3588 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3589 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3590 ac->high_zoneidx, ac->nodemask);
3593 /* This is the last chance, in general, before the goto nopage. */
3594 page = get_page_from_freelist(gfp_mask, order,
3595 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3599 /* Allocate without watermarks if the context allows */
3600 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3601 page = get_page_from_freelist(gfp_mask, order,
3602 ALLOC_NO_WATERMARKS, ac);
3607 /* Caller is not willing to reclaim, we can't balance anything */
3608 if (!can_direct_reclaim) {
3610 * All existing users of the __GFP_NOFAIL are blockable, so warn
3611 * of any new users that actually allow this type of allocation
3614 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3618 /* Avoid recursion of direct reclaim */
3619 if (current->flags & PF_MEMALLOC) {
3621 * __GFP_NOFAIL request from this context is rather bizarre
3622 * because we cannot reclaim anything and only can loop waiting
3623 * for somebody to do a work for us.
3625 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3632 /* Avoid allocations with no watermarks from looping endlessly */
3633 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3637 * Try direct compaction. The first pass is asynchronous. Subsequent
3638 * attempts after direct reclaim are synchronous
3640 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3646 /* Checks for THP-specific high-order allocations */
3647 if (is_thp_gfp_mask(gfp_mask)) {
3649 * If compaction is deferred for high-order allocations, it is
3650 * because sync compaction recently failed. If this is the case
3651 * and the caller requested a THP allocation, we do not want
3652 * to heavily disrupt the system, so we fail the allocation
3653 * instead of entering direct reclaim.
3655 if (compact_result == COMPACT_DEFERRED)
3659 * Compaction is contended so rather back off than cause
3662 if(compact_result == COMPACT_CONTENDED)
3666 if (order && compaction_made_progress(compact_result))
3667 compaction_retries++;
3669 /* Try direct reclaim and then allocating */
3670 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3671 &did_some_progress);
3675 /* Do not loop if specifically requested */
3676 if (gfp_mask & __GFP_NORETRY)
3680 * Do not retry costly high order allocations unless they are
3683 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3687 * Costly allocations might have made a progress but this doesn't mean
3688 * their order will become available due to high fragmentation so
3689 * always increment the no progress counter for them
3691 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3692 no_progress_loops = 0;
3694 no_progress_loops++;
3696 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3697 did_some_progress > 0, no_progress_loops))
3701 * It doesn't make any sense to retry for the compaction if the order-0
3702 * reclaim is not able to make any progress because the current
3703 * implementation of the compaction depends on the sufficient amount
3704 * of free memory (see __compaction_suitable)
3706 if (did_some_progress > 0 &&
3707 should_compact_retry(ac, order, alloc_flags,
3708 compact_result, &migration_mode,
3709 compaction_retries))
3712 /* Reclaim has failed us, start killing things */
3713 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3717 /* Retry as long as the OOM killer is making progress */
3718 if (did_some_progress) {
3719 no_progress_loops = 0;
3725 * High-order allocations do not necessarily loop after direct reclaim
3726 * and reclaim/compaction depends on compaction being called after
3727 * reclaim so call directly if necessary.
3728 * It can become very expensive to allocate transparent hugepages at
3729 * fault, so use asynchronous memory compaction for THP unless it is
3730 * khugepaged trying to collapse. All other requests should tolerate
3731 * at least light sync migration.
3733 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3734 migration_mode = MIGRATE_ASYNC;
3736 migration_mode = MIGRATE_SYNC_LIGHT;
3737 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3743 warn_alloc_failed(gfp_mask, order, NULL);
3749 * This is the 'heart' of the zoned buddy allocator.
3752 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3753 struct zonelist *zonelist, nodemask_t *nodemask)
3756 unsigned int cpuset_mems_cookie;
3757 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3758 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3759 struct alloc_context ac = {
3760 .high_zoneidx = gfp_zone(gfp_mask),
3761 .zonelist = zonelist,
3762 .nodemask = nodemask,
3763 .migratetype = gfpflags_to_migratetype(gfp_mask),
3766 if (cpusets_enabled()) {
3767 alloc_mask |= __GFP_HARDWALL;
3768 alloc_flags |= ALLOC_CPUSET;
3770 ac.nodemask = &cpuset_current_mems_allowed;
3773 gfp_mask &= gfp_allowed_mask;
3775 lockdep_trace_alloc(gfp_mask);
3777 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3779 if (should_fail_alloc_page(gfp_mask, order))
3783 * Check the zones suitable for the gfp_mask contain at least one
3784 * valid zone. It's possible to have an empty zonelist as a result
3785 * of __GFP_THISNODE and a memoryless node
3787 if (unlikely(!zonelist->_zonerefs->zone))
3790 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3791 alloc_flags |= ALLOC_CMA;
3794 cpuset_mems_cookie = read_mems_allowed_begin();
3796 /* Dirty zone balancing only done in the fast path */
3797 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3800 * The preferred zone is used for statistics but crucially it is
3801 * also used as the starting point for the zonelist iterator. It
3802 * may get reset for allocations that ignore memory policies.
3804 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3805 ac.high_zoneidx, ac.nodemask);
3806 if (!ac.preferred_zoneref) {
3811 /* First allocation attempt */
3812 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3817 * Runtime PM, block IO and its error handling path can deadlock
3818 * because I/O on the device might not complete.
3820 alloc_mask = memalloc_noio_flags(gfp_mask);
3821 ac.spread_dirty_pages = false;
3824 * Restore the original nodemask if it was potentially replaced with
3825 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3827 if (cpusets_enabled())
3828 ac.nodemask = nodemask;
3829 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3833 * When updating a task's mems_allowed, it is possible to race with
3834 * parallel threads in such a way that an allocation can fail while
3835 * the mask is being updated. If a page allocation is about to fail,
3836 * check if the cpuset changed during allocation and if so, retry.
3838 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3839 alloc_mask = gfp_mask;
3844 if (kmemcheck_enabled && page)
3845 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3847 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3851 EXPORT_SYMBOL(__alloc_pages_nodemask);
3854 * Common helper functions.
3856 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3861 * __get_free_pages() returns a 32-bit address, which cannot represent
3864 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3866 page = alloc_pages(gfp_mask, order);
3869 return (unsigned long) page_address(page);
3871 EXPORT_SYMBOL(__get_free_pages);
3873 unsigned long get_zeroed_page(gfp_t gfp_mask)
3875 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3877 EXPORT_SYMBOL(get_zeroed_page);
3879 void __free_pages(struct page *page, unsigned int order)
3881 if (put_page_testzero(page)) {
3883 free_hot_cold_page(page, false);
3885 __free_pages_ok(page, order);
3889 EXPORT_SYMBOL(__free_pages);
3891 void free_pages(unsigned long addr, unsigned int order)
3894 VM_BUG_ON(!virt_addr_valid((void *)addr));
3895 __free_pages(virt_to_page((void *)addr), order);
3899 EXPORT_SYMBOL(free_pages);
3903 * An arbitrary-length arbitrary-offset area of memory which resides
3904 * within a 0 or higher order page. Multiple fragments within that page
3905 * are individually refcounted, in the page's reference counter.
3907 * The page_frag functions below provide a simple allocation framework for
3908 * page fragments. This is used by the network stack and network device
3909 * drivers to provide a backing region of memory for use as either an
3910 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3912 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3915 struct page *page = NULL;
3916 gfp_t gfp = gfp_mask;
3918 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3919 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3921 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3922 PAGE_FRAG_CACHE_MAX_ORDER);
3923 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3925 if (unlikely(!page))
3926 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3928 nc->va = page ? page_address(page) : NULL;
3933 void *__alloc_page_frag(struct page_frag_cache *nc,
3934 unsigned int fragsz, gfp_t gfp_mask)
3936 unsigned int size = PAGE_SIZE;
3940 if (unlikely(!nc->va)) {
3942 page = __page_frag_refill(nc, gfp_mask);
3946 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3947 /* if size can vary use size else just use PAGE_SIZE */
3950 /* Even if we own the page, we do not use atomic_set().
3951 * This would break get_page_unless_zero() users.
3953 page_ref_add(page, size - 1);
3955 /* reset page count bias and offset to start of new frag */
3956 nc->pfmemalloc = page_is_pfmemalloc(page);
3957 nc->pagecnt_bias = size;
3961 offset = nc->offset - fragsz;
3962 if (unlikely(offset < 0)) {
3963 page = virt_to_page(nc->va);
3965 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3968 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3969 /* if size can vary use size else just use PAGE_SIZE */
3972 /* OK, page count is 0, we can safely set it */
3973 set_page_count(page, size);
3975 /* reset page count bias and offset to start of new frag */
3976 nc->pagecnt_bias = size;
3977 offset = size - fragsz;
3981 nc->offset = offset;
3983 return nc->va + offset;
3985 EXPORT_SYMBOL(__alloc_page_frag);
3988 * Frees a page fragment allocated out of either a compound or order 0 page.
3990 void __free_page_frag(void *addr)
3992 struct page *page = virt_to_head_page(addr);
3994 if (unlikely(put_page_testzero(page)))
3995 __free_pages_ok(page, compound_order(page));
3997 EXPORT_SYMBOL(__free_page_frag);
4000 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
4001 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
4002 * equivalent to alloc_pages.
4004 * It should be used when the caller would like to use kmalloc, but since the
4005 * allocation is large, it has to fall back to the page allocator.
4007 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
4011 page = alloc_pages(gfp_mask, order);
4012 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4013 __free_pages(page, order);
4019 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
4023 page = alloc_pages_node(nid, gfp_mask, order);
4024 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4025 __free_pages(page, order);
4032 * __free_kmem_pages and free_kmem_pages will free pages allocated with
4035 void __free_kmem_pages(struct page *page, unsigned int order)
4037 memcg_kmem_uncharge(page, order);
4038 __free_pages(page, order);
4041 void free_kmem_pages(unsigned long addr, unsigned int order)
4044 VM_BUG_ON(!virt_addr_valid((void *)addr));
4045 __free_kmem_pages(virt_to_page((void *)addr), order);
4049 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4053 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4054 unsigned long used = addr + PAGE_ALIGN(size);
4056 split_page(virt_to_page((void *)addr), order);
4057 while (used < alloc_end) {
4062 return (void *)addr;
4066 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4067 * @size: the number of bytes to allocate
4068 * @gfp_mask: GFP flags for the allocation
4070 * This function is similar to alloc_pages(), except that it allocates the
4071 * minimum number of pages to satisfy the request. alloc_pages() can only
4072 * allocate memory in power-of-two pages.
4074 * This function is also limited by MAX_ORDER.
4076 * Memory allocated by this function must be released by free_pages_exact().
4078 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4080 unsigned int order = get_order(size);
4083 addr = __get_free_pages(gfp_mask, order);
4084 return make_alloc_exact(addr, order, size);
4086 EXPORT_SYMBOL(alloc_pages_exact);
4089 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4091 * @nid: the preferred node ID where memory should be allocated
4092 * @size: the number of bytes to allocate
4093 * @gfp_mask: GFP flags for the allocation
4095 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4098 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4100 unsigned int order = get_order(size);
4101 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4104 return make_alloc_exact((unsigned long)page_address(p), order, size);
4108 * free_pages_exact - release memory allocated via alloc_pages_exact()
4109 * @virt: the value returned by alloc_pages_exact.
4110 * @size: size of allocation, same value as passed to alloc_pages_exact().
4112 * Release the memory allocated by a previous call to alloc_pages_exact.
4114 void free_pages_exact(void *virt, size_t size)
4116 unsigned long addr = (unsigned long)virt;
4117 unsigned long end = addr + PAGE_ALIGN(size);
4119 while (addr < end) {
4124 EXPORT_SYMBOL(free_pages_exact);
4127 * nr_free_zone_pages - count number of pages beyond high watermark
4128 * @offset: The zone index of the highest zone
4130 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4131 * high watermark within all zones at or below a given zone index. For each
4132 * zone, the number of pages is calculated as:
4133 * managed_pages - high_pages
4135 static unsigned long nr_free_zone_pages(int offset)
4140 /* Just pick one node, since fallback list is circular */
4141 unsigned long sum = 0;
4143 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4145 for_each_zone_zonelist(zone, z, zonelist, offset) {
4146 unsigned long size = zone->managed_pages;
4147 unsigned long high = high_wmark_pages(zone);
4156 * nr_free_buffer_pages - count number of pages beyond high watermark
4158 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4159 * watermark within ZONE_DMA and ZONE_NORMAL.
4161 unsigned long nr_free_buffer_pages(void)
4163 return nr_free_zone_pages(gfp_zone(GFP_USER));
4165 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4168 * nr_free_pagecache_pages - count number of pages beyond high watermark
4170 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4171 * high watermark within all zones.
4173 unsigned long nr_free_pagecache_pages(void)
4175 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4178 static inline void show_node(struct zone *zone)
4180 if (IS_ENABLED(CONFIG_NUMA))
4181 printk("Node %d ", zone_to_nid(zone));
4184 long si_mem_available(void)
4187 unsigned long pagecache;
4188 unsigned long wmark_low = 0;
4189 unsigned long pages[NR_LRU_LISTS];
4193 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4194 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4197 wmark_low += zone->watermark[WMARK_LOW];
4200 * Estimate the amount of memory available for userspace allocations,
4201 * without causing swapping.
4203 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4206 * Not all the page cache can be freed, otherwise the system will
4207 * start swapping. Assume at least half of the page cache, or the
4208 * low watermark worth of cache, needs to stay.
4210 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4211 pagecache -= min(pagecache / 2, wmark_low);
4212 available += pagecache;
4215 * Part of the reclaimable slab consists of items that are in use,
4216 * and cannot be freed. Cap this estimate at the low watermark.
4218 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4219 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4225 EXPORT_SYMBOL_GPL(si_mem_available);
4227 void si_meminfo(struct sysinfo *val)
4229 val->totalram = totalram_pages;
4230 val->sharedram = global_page_state(NR_SHMEM);
4231 val->freeram = global_page_state(NR_FREE_PAGES);
4232 val->bufferram = nr_blockdev_pages();
4233 val->totalhigh = totalhigh_pages;
4234 val->freehigh = nr_free_highpages();
4235 val->mem_unit = PAGE_SIZE;
4238 EXPORT_SYMBOL(si_meminfo);
4241 void si_meminfo_node(struct sysinfo *val, int nid)
4243 int zone_type; /* needs to be signed */
4244 unsigned long managed_pages = 0;
4245 unsigned long managed_highpages = 0;
4246 unsigned long free_highpages = 0;
4247 pg_data_t *pgdat = NODE_DATA(nid);
4249 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4250 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4251 val->totalram = managed_pages;
4252 val->sharedram = node_page_state(nid, NR_SHMEM);
4253 val->freeram = node_page_state(nid, NR_FREE_PAGES);
4254 #ifdef CONFIG_HIGHMEM
4255 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4256 struct zone *zone = &pgdat->node_zones[zone_type];
4258 if (is_highmem(zone)) {
4259 managed_highpages += zone->managed_pages;
4260 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4263 val->totalhigh = managed_highpages;
4264 val->freehigh = free_highpages;
4266 val->totalhigh = managed_highpages;
4267 val->freehigh = free_highpages;
4269 val->mem_unit = PAGE_SIZE;
4274 * Determine whether the node should be displayed or not, depending on whether
4275 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4277 bool skip_free_areas_node(unsigned int flags, int nid)
4280 unsigned int cpuset_mems_cookie;
4282 if (!(flags & SHOW_MEM_FILTER_NODES))
4286 cpuset_mems_cookie = read_mems_allowed_begin();
4287 ret = !node_isset(nid, cpuset_current_mems_allowed);
4288 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4293 #define K(x) ((x) << (PAGE_SHIFT-10))
4295 static void show_migration_types(unsigned char type)
4297 static const char types[MIGRATE_TYPES] = {
4298 [MIGRATE_UNMOVABLE] = 'U',
4299 [MIGRATE_MOVABLE] = 'M',
4300 [MIGRATE_RECLAIMABLE] = 'E',
4301 [MIGRATE_HIGHATOMIC] = 'H',
4303 [MIGRATE_CMA] = 'C',
4305 #ifdef CONFIG_MEMORY_ISOLATION
4306 [MIGRATE_ISOLATE] = 'I',
4309 char tmp[MIGRATE_TYPES + 1];
4313 for (i = 0; i < MIGRATE_TYPES; i++) {
4314 if (type & (1 << i))
4319 printk("(%s) ", tmp);
4323 * Show free area list (used inside shift_scroll-lock stuff)
4324 * We also calculate the percentage fragmentation. We do this by counting the
4325 * memory on each free list with the exception of the first item on the list.
4328 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4331 void show_free_areas(unsigned int filter)
4333 unsigned long free_pcp = 0;
4337 for_each_populated_zone(zone) {
4338 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4341 for_each_online_cpu(cpu)
4342 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4345 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4346 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4347 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4348 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4349 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4350 " free:%lu free_pcp:%lu free_cma:%lu\n",
4351 global_page_state(NR_ACTIVE_ANON),
4352 global_page_state(NR_INACTIVE_ANON),
4353 global_page_state(NR_ISOLATED_ANON),
4354 global_page_state(NR_ACTIVE_FILE),
4355 global_page_state(NR_INACTIVE_FILE),
4356 global_page_state(NR_ISOLATED_FILE),
4357 global_page_state(NR_UNEVICTABLE),
4358 global_page_state(NR_FILE_DIRTY),
4359 global_page_state(NR_WRITEBACK),
4360 global_page_state(NR_UNSTABLE_NFS),
4361 global_page_state(NR_SLAB_RECLAIMABLE),
4362 global_page_state(NR_SLAB_UNRECLAIMABLE),
4363 global_page_state(NR_FILE_MAPPED),
4364 global_page_state(NR_SHMEM),
4365 global_page_state(NR_PAGETABLE),
4366 global_page_state(NR_BOUNCE),
4367 global_page_state(NR_FREE_PAGES),
4369 global_page_state(NR_FREE_CMA_PAGES));
4371 for_each_populated_zone(zone) {
4374 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4378 for_each_online_cpu(cpu)
4379 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4387 " active_anon:%lukB"
4388 " inactive_anon:%lukB"
4389 " active_file:%lukB"
4390 " inactive_file:%lukB"
4391 " unevictable:%lukB"
4392 " isolated(anon):%lukB"
4393 " isolated(file):%lukB"
4401 " slab_reclaimable:%lukB"
4402 " slab_unreclaimable:%lukB"
4403 " kernel_stack:%lukB"
4410 " writeback_tmp:%lukB"
4411 " pages_scanned:%lu"
4412 " all_unreclaimable? %s"
4415 K(zone_page_state(zone, NR_FREE_PAGES)),
4416 K(min_wmark_pages(zone)),
4417 K(low_wmark_pages(zone)),
4418 K(high_wmark_pages(zone)),
4419 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4420 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4421 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4422 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4423 K(zone_page_state(zone, NR_UNEVICTABLE)),
4424 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4425 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4426 K(zone->present_pages),
4427 K(zone->managed_pages),
4428 K(zone_page_state(zone, NR_MLOCK)),
4429 K(zone_page_state(zone, NR_FILE_DIRTY)),
4430 K(zone_page_state(zone, NR_WRITEBACK)),
4431 K(zone_page_state(zone, NR_FILE_MAPPED)),
4432 K(zone_page_state(zone, NR_SHMEM)),
4433 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4434 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4435 zone_page_state(zone, NR_KERNEL_STACK) *
4437 K(zone_page_state(zone, NR_PAGETABLE)),
4438 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4439 K(zone_page_state(zone, NR_BOUNCE)),
4441 K(this_cpu_read(zone->pageset->pcp.count)),
4442 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4443 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4444 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4445 (!zone_reclaimable(zone) ? "yes" : "no")
4447 printk("lowmem_reserve[]:");
4448 for (i = 0; i < MAX_NR_ZONES; i++)
4449 printk(" %ld", zone->lowmem_reserve[i]);
4453 for_each_populated_zone(zone) {
4455 unsigned long nr[MAX_ORDER], flags, total = 0;
4456 unsigned char types[MAX_ORDER];
4458 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4461 printk("%s: ", zone->name);
4463 spin_lock_irqsave(&zone->lock, flags);
4464 for (order = 0; order < MAX_ORDER; order++) {
4465 struct free_area *area = &zone->free_area[order];
4468 nr[order] = area->nr_free;
4469 total += nr[order] << order;
4472 for (type = 0; type < MIGRATE_TYPES; type++) {
4473 if (!list_empty(&area->free_list[type]))
4474 types[order] |= 1 << type;
4477 spin_unlock_irqrestore(&zone->lock, flags);
4478 for (order = 0; order < MAX_ORDER; order++) {
4479 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4481 show_migration_types(types[order]);
4483 printk("= %lukB\n", K(total));
4486 hugetlb_show_meminfo();
4488 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4490 show_swap_cache_info();
4493 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4495 zoneref->zone = zone;
4496 zoneref->zone_idx = zone_idx(zone);
4500 * Builds allocation fallback zone lists.
4502 * Add all populated zones of a node to the zonelist.
4504 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4508 enum zone_type zone_type = MAX_NR_ZONES;
4512 zone = pgdat->node_zones + zone_type;
4513 if (populated_zone(zone)) {
4514 zoneref_set_zone(zone,
4515 &zonelist->_zonerefs[nr_zones++]);
4516 check_highest_zone(zone_type);
4518 } while (zone_type);
4526 * 0 = automatic detection of better ordering.
4527 * 1 = order by ([node] distance, -zonetype)
4528 * 2 = order by (-zonetype, [node] distance)
4530 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4531 * the same zonelist. So only NUMA can configure this param.
4533 #define ZONELIST_ORDER_DEFAULT 0
4534 #define ZONELIST_ORDER_NODE 1
4535 #define ZONELIST_ORDER_ZONE 2
4537 /* zonelist order in the kernel.
4538 * set_zonelist_order() will set this to NODE or ZONE.
4540 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4541 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4545 /* The value user specified ....changed by config */
4546 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4547 /* string for sysctl */
4548 #define NUMA_ZONELIST_ORDER_LEN 16
4549 char numa_zonelist_order[16] = "default";
4552 * interface for configure zonelist ordering.
4553 * command line option "numa_zonelist_order"
4554 * = "[dD]efault - default, automatic configuration.
4555 * = "[nN]ode - order by node locality, then by zone within node
4556 * = "[zZ]one - order by zone, then by locality within zone
4559 static int __parse_numa_zonelist_order(char *s)
4561 if (*s == 'd' || *s == 'D') {
4562 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4563 } else if (*s == 'n' || *s == 'N') {
4564 user_zonelist_order = ZONELIST_ORDER_NODE;
4565 } else if (*s == 'z' || *s == 'Z') {
4566 user_zonelist_order = ZONELIST_ORDER_ZONE;
4568 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4574 static __init int setup_numa_zonelist_order(char *s)
4581 ret = __parse_numa_zonelist_order(s);
4583 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4587 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4590 * sysctl handler for numa_zonelist_order
4592 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4593 void __user *buffer, size_t *length,
4596 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4598 static DEFINE_MUTEX(zl_order_mutex);
4600 mutex_lock(&zl_order_mutex);
4602 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4606 strcpy(saved_string, (char *)table->data);
4608 ret = proc_dostring(table, write, buffer, length, ppos);
4612 int oldval = user_zonelist_order;
4614 ret = __parse_numa_zonelist_order((char *)table->data);
4617 * bogus value. restore saved string
4619 strncpy((char *)table->data, saved_string,
4620 NUMA_ZONELIST_ORDER_LEN);
4621 user_zonelist_order = oldval;
4622 } else if (oldval != user_zonelist_order) {
4623 mutex_lock(&zonelists_mutex);
4624 build_all_zonelists(NULL, NULL);
4625 mutex_unlock(&zonelists_mutex);
4629 mutex_unlock(&zl_order_mutex);
4634 #define MAX_NODE_LOAD (nr_online_nodes)
4635 static int node_load[MAX_NUMNODES];
4638 * find_next_best_node - find the next node that should appear in a given node's fallback list
4639 * @node: node whose fallback list we're appending
4640 * @used_node_mask: nodemask_t of already used nodes
4642 * We use a number of factors to determine which is the next node that should
4643 * appear on a given node's fallback list. The node should not have appeared
4644 * already in @node's fallback list, and it should be the next closest node
4645 * according to the distance array (which contains arbitrary distance values
4646 * from each node to each node in the system), and should also prefer nodes
4647 * with no CPUs, since presumably they'll have very little allocation pressure
4648 * on them otherwise.
4649 * It returns -1 if no node is found.
4651 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4654 int min_val = INT_MAX;
4655 int best_node = NUMA_NO_NODE;
4656 const struct cpumask *tmp = cpumask_of_node(0);
4658 /* Use the local node if we haven't already */
4659 if (!node_isset(node, *used_node_mask)) {
4660 node_set(node, *used_node_mask);
4664 for_each_node_state(n, N_MEMORY) {
4666 /* Don't want a node to appear more than once */
4667 if (node_isset(n, *used_node_mask))
4670 /* Use the distance array to find the distance */
4671 val = node_distance(node, n);
4673 /* Penalize nodes under us ("prefer the next node") */
4676 /* Give preference to headless and unused nodes */
4677 tmp = cpumask_of_node(n);
4678 if (!cpumask_empty(tmp))
4679 val += PENALTY_FOR_NODE_WITH_CPUS;
4681 /* Slight preference for less loaded node */
4682 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4683 val += node_load[n];
4685 if (val < min_val) {
4692 node_set(best_node, *used_node_mask);
4699 * Build zonelists ordered by node and zones within node.
4700 * This results in maximum locality--normal zone overflows into local
4701 * DMA zone, if any--but risks exhausting DMA zone.
4703 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4706 struct zonelist *zonelist;
4708 zonelist = &pgdat->node_zonelists[0];
4709 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4711 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4712 zonelist->_zonerefs[j].zone = NULL;
4713 zonelist->_zonerefs[j].zone_idx = 0;
4717 * Build gfp_thisnode zonelists
4719 static void build_thisnode_zonelists(pg_data_t *pgdat)
4722 struct zonelist *zonelist;
4724 zonelist = &pgdat->node_zonelists[1];
4725 j = build_zonelists_node(pgdat, zonelist, 0);
4726 zonelist->_zonerefs[j].zone = NULL;
4727 zonelist->_zonerefs[j].zone_idx = 0;
4731 * Build zonelists ordered by zone and nodes within zones.
4732 * This results in conserving DMA zone[s] until all Normal memory is
4733 * exhausted, but results in overflowing to remote node while memory
4734 * may still exist in local DMA zone.
4736 static int node_order[MAX_NUMNODES];
4738 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4741 int zone_type; /* needs to be signed */
4743 struct zonelist *zonelist;
4745 zonelist = &pgdat->node_zonelists[0];
4747 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4748 for (j = 0; j < nr_nodes; j++) {
4749 node = node_order[j];
4750 z = &NODE_DATA(node)->node_zones[zone_type];
4751 if (populated_zone(z)) {
4753 &zonelist->_zonerefs[pos++]);
4754 check_highest_zone(zone_type);
4758 zonelist->_zonerefs[pos].zone = NULL;
4759 zonelist->_zonerefs[pos].zone_idx = 0;
4762 #if defined(CONFIG_64BIT)
4764 * Devices that require DMA32/DMA are relatively rare and do not justify a
4765 * penalty to every machine in case the specialised case applies. Default
4766 * to Node-ordering on 64-bit NUMA machines
4768 static int default_zonelist_order(void)
4770 return ZONELIST_ORDER_NODE;
4774 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4775 * by the kernel. If processes running on node 0 deplete the low memory zone
4776 * then reclaim will occur more frequency increasing stalls and potentially
4777 * be easier to OOM if a large percentage of the zone is under writeback or
4778 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4779 * Hence, default to zone ordering on 32-bit.
4781 static int default_zonelist_order(void)
4783 return ZONELIST_ORDER_ZONE;
4785 #endif /* CONFIG_64BIT */
4787 static void set_zonelist_order(void)
4789 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4790 current_zonelist_order = default_zonelist_order();
4792 current_zonelist_order = user_zonelist_order;
4795 static void build_zonelists(pg_data_t *pgdat)
4798 nodemask_t used_mask;
4799 int local_node, prev_node;
4800 struct zonelist *zonelist;
4801 unsigned int order = current_zonelist_order;
4803 /* initialize zonelists */
4804 for (i = 0; i < MAX_ZONELISTS; i++) {
4805 zonelist = pgdat->node_zonelists + i;
4806 zonelist->_zonerefs[0].zone = NULL;
4807 zonelist->_zonerefs[0].zone_idx = 0;
4810 /* NUMA-aware ordering of nodes */
4811 local_node = pgdat->node_id;
4812 load = nr_online_nodes;
4813 prev_node = local_node;
4814 nodes_clear(used_mask);
4816 memset(node_order, 0, sizeof(node_order));
4819 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4821 * We don't want to pressure a particular node.
4822 * So adding penalty to the first node in same
4823 * distance group to make it round-robin.
4825 if (node_distance(local_node, node) !=
4826 node_distance(local_node, prev_node))
4827 node_load[node] = load;
4831 if (order == ZONELIST_ORDER_NODE)
4832 build_zonelists_in_node_order(pgdat, node);
4834 node_order[i++] = node; /* remember order */
4837 if (order == ZONELIST_ORDER_ZONE) {
4838 /* calculate node order -- i.e., DMA last! */
4839 build_zonelists_in_zone_order(pgdat, i);
4842 build_thisnode_zonelists(pgdat);
4845 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4847 * Return node id of node used for "local" allocations.
4848 * I.e., first node id of first zone in arg node's generic zonelist.
4849 * Used for initializing percpu 'numa_mem', which is used primarily
4850 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4852 int local_memory_node(int node)
4856 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4857 gfp_zone(GFP_KERNEL),
4859 return z->zone->node;
4863 #else /* CONFIG_NUMA */
4865 static void set_zonelist_order(void)
4867 current_zonelist_order = ZONELIST_ORDER_ZONE;
4870 static void build_zonelists(pg_data_t *pgdat)
4872 int node, local_node;
4874 struct zonelist *zonelist;
4876 local_node = pgdat->node_id;
4878 zonelist = &pgdat->node_zonelists[0];
4879 j = build_zonelists_node(pgdat, zonelist, 0);
4882 * Now we build the zonelist so that it contains the zones
4883 * of all the other nodes.
4884 * We don't want to pressure a particular node, so when
4885 * building the zones for node N, we make sure that the
4886 * zones coming right after the local ones are those from
4887 * node N+1 (modulo N)
4889 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4890 if (!node_online(node))
4892 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4894 for (node = 0; node < local_node; node++) {
4895 if (!node_online(node))
4897 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4900 zonelist->_zonerefs[j].zone = NULL;
4901 zonelist->_zonerefs[j].zone_idx = 0;
4904 #endif /* CONFIG_NUMA */
4907 * Boot pageset table. One per cpu which is going to be used for all
4908 * zones and all nodes. The parameters will be set in such a way
4909 * that an item put on a list will immediately be handed over to
4910 * the buddy list. This is safe since pageset manipulation is done
4911 * with interrupts disabled.
4913 * The boot_pagesets must be kept even after bootup is complete for
4914 * unused processors and/or zones. They do play a role for bootstrapping
4915 * hotplugged processors.
4917 * zoneinfo_show() and maybe other functions do
4918 * not check if the processor is online before following the pageset pointer.
4919 * Other parts of the kernel may not check if the zone is available.
4921 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4922 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4923 static void setup_zone_pageset(struct zone *zone);
4926 * Global mutex to protect against size modification of zonelists
4927 * as well as to serialize pageset setup for the new populated zone.
4929 DEFINE_MUTEX(zonelists_mutex);
4931 /* return values int ....just for stop_machine() */
4932 static int __build_all_zonelists(void *data)
4936 pg_data_t *self = data;
4939 memset(node_load, 0, sizeof(node_load));
4942 if (self && !node_online(self->node_id)) {
4943 build_zonelists(self);
4946 for_each_online_node(nid) {
4947 pg_data_t *pgdat = NODE_DATA(nid);
4949 build_zonelists(pgdat);
4953 * Initialize the boot_pagesets that are going to be used
4954 * for bootstrapping processors. The real pagesets for
4955 * each zone will be allocated later when the per cpu
4956 * allocator is available.
4958 * boot_pagesets are used also for bootstrapping offline
4959 * cpus if the system is already booted because the pagesets
4960 * are needed to initialize allocators on a specific cpu too.
4961 * F.e. the percpu allocator needs the page allocator which
4962 * needs the percpu allocator in order to allocate its pagesets
4963 * (a chicken-egg dilemma).
4965 for_each_possible_cpu(cpu) {
4966 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4968 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4970 * We now know the "local memory node" for each node--
4971 * i.e., the node of the first zone in the generic zonelist.
4972 * Set up numa_mem percpu variable for on-line cpus. During
4973 * boot, only the boot cpu should be on-line; we'll init the
4974 * secondary cpus' numa_mem as they come on-line. During
4975 * node/memory hotplug, we'll fixup all on-line cpus.
4977 if (cpu_online(cpu))
4978 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4985 static noinline void __init
4986 build_all_zonelists_init(void)
4988 __build_all_zonelists(NULL);
4989 mminit_verify_zonelist();
4990 cpuset_init_current_mems_allowed();
4994 * Called with zonelists_mutex held always
4995 * unless system_state == SYSTEM_BOOTING.
4997 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4998 * [we're only called with non-NULL zone through __meminit paths] and
4999 * (2) call of __init annotated helper build_all_zonelists_init
5000 * [protected by SYSTEM_BOOTING].
5002 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5004 set_zonelist_order();
5006 if (system_state == SYSTEM_BOOTING) {
5007 build_all_zonelists_init();
5009 #ifdef CONFIG_MEMORY_HOTPLUG
5011 setup_zone_pageset(zone);
5013 /* we have to stop all cpus to guarantee there is no user
5015 stop_machine(__build_all_zonelists, pgdat, NULL);
5016 /* cpuset refresh routine should be here */
5018 vm_total_pages = nr_free_pagecache_pages();
5020 * Disable grouping by mobility if the number of pages in the
5021 * system is too low to allow the mechanism to work. It would be
5022 * more accurate, but expensive to check per-zone. This check is
5023 * made on memory-hotadd so a system can start with mobility
5024 * disabled and enable it later
5026 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5027 page_group_by_mobility_disabled = 1;
5029 page_group_by_mobility_disabled = 0;
5031 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5033 zonelist_order_name[current_zonelist_order],
5034 page_group_by_mobility_disabled ? "off" : "on",
5037 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5042 * Helper functions to size the waitqueue hash table.
5043 * Essentially these want to choose hash table sizes sufficiently
5044 * large so that collisions trying to wait on pages are rare.
5045 * But in fact, the number of active page waitqueues on typical
5046 * systems is ridiculously low, less than 200. So this is even
5047 * conservative, even though it seems large.
5049 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5050 * waitqueues, i.e. the size of the waitq table given the number of pages.
5052 #define PAGES_PER_WAITQUEUE 256
5054 #ifndef CONFIG_MEMORY_HOTPLUG
5055 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5057 unsigned long size = 1;
5059 pages /= PAGES_PER_WAITQUEUE;
5061 while (size < pages)
5065 * Once we have dozens or even hundreds of threads sleeping
5066 * on IO we've got bigger problems than wait queue collision.
5067 * Limit the size of the wait table to a reasonable size.
5069 size = min(size, 4096UL);
5071 return max(size, 4UL);
5075 * A zone's size might be changed by hot-add, so it is not possible to determine
5076 * a suitable size for its wait_table. So we use the maximum size now.
5078 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5080 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5081 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5082 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5084 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5085 * or more by the traditional way. (See above). It equals:
5087 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5088 * ia64(16K page size) : = ( 8G + 4M)byte.
5089 * powerpc (64K page size) : = (32G +16M)byte.
5091 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5098 * This is an integer logarithm so that shifts can be used later
5099 * to extract the more random high bits from the multiplicative
5100 * hash function before the remainder is taken.
5102 static inline unsigned long wait_table_bits(unsigned long size)
5108 * Initially all pages are reserved - free ones are freed
5109 * up by free_all_bootmem() once the early boot process is
5110 * done. Non-atomic initialization, single-pass.
5112 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5113 unsigned long start_pfn, enum memmap_context context)
5115 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5116 unsigned long end_pfn = start_pfn + size;
5117 pg_data_t *pgdat = NODE_DATA(nid);
5119 unsigned long nr_initialised = 0;
5120 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5121 struct memblock_region *r = NULL, *tmp;
5124 if (highest_memmap_pfn < end_pfn - 1)
5125 highest_memmap_pfn = end_pfn - 1;
5128 * Honor reservation requested by the driver for this ZONE_DEVICE
5131 if (altmap && start_pfn == altmap->base_pfn)
5132 start_pfn += altmap->reserve;
5134 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5136 * There can be holes in boot-time mem_map[]s handed to this
5137 * function. They do not exist on hotplugged memory.
5139 if (context != MEMMAP_EARLY)
5142 if (!early_pfn_valid(pfn))
5144 if (!early_pfn_in_nid(pfn, nid))
5146 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5149 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5151 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5152 * from zone_movable_pfn[nid] to end of each node should be
5153 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5155 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5156 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5160 * Check given memblock attribute by firmware which can affect
5161 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5162 * mirrored, it's an overlapped memmap init. skip it.
5164 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5165 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5166 for_each_memblock(memory, tmp)
5167 if (pfn < memblock_region_memory_end_pfn(tmp))
5171 if (pfn >= memblock_region_memory_base_pfn(r) &&
5172 memblock_is_mirror(r)) {
5173 /* already initialized as NORMAL */
5174 pfn = memblock_region_memory_end_pfn(r);
5182 * Mark the block movable so that blocks are reserved for
5183 * movable at startup. This will force kernel allocations
5184 * to reserve their blocks rather than leaking throughout
5185 * the address space during boot when many long-lived
5186 * kernel allocations are made.
5188 * bitmap is created for zone's valid pfn range. but memmap
5189 * can be created for invalid pages (for alignment)
5190 * check here not to call set_pageblock_migratetype() against
5193 if (!(pfn & (pageblock_nr_pages - 1))) {
5194 struct page *page = pfn_to_page(pfn);
5196 __init_single_page(page, pfn, zone, nid);
5197 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5199 __init_single_pfn(pfn, zone, nid);
5204 static void __meminit zone_init_free_lists(struct zone *zone)
5206 unsigned int order, t;
5207 for_each_migratetype_order(order, t) {
5208 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5209 zone->free_area[order].nr_free = 0;
5213 #ifndef __HAVE_ARCH_MEMMAP_INIT
5214 #define memmap_init(size, nid, zone, start_pfn) \
5215 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5218 static int zone_batchsize(struct zone *zone)
5224 * The per-cpu-pages pools are set to around 1000th of the
5225 * size of the zone. But no more than 1/2 of a meg.
5227 * OK, so we don't know how big the cache is. So guess.
5229 batch = zone->managed_pages / 1024;
5230 if (batch * PAGE_SIZE > 512 * 1024)
5231 batch = (512 * 1024) / PAGE_SIZE;
5232 batch /= 4; /* We effectively *= 4 below */
5237 * Clamp the batch to a 2^n - 1 value. Having a power
5238 * of 2 value was found to be more likely to have
5239 * suboptimal cache aliasing properties in some cases.
5241 * For example if 2 tasks are alternately allocating
5242 * batches of pages, one task can end up with a lot
5243 * of pages of one half of the possible page colors
5244 * and the other with pages of the other colors.
5246 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5251 /* The deferral and batching of frees should be suppressed under NOMMU
5254 * The problem is that NOMMU needs to be able to allocate large chunks
5255 * of contiguous memory as there's no hardware page translation to
5256 * assemble apparent contiguous memory from discontiguous pages.
5258 * Queueing large contiguous runs of pages for batching, however,
5259 * causes the pages to actually be freed in smaller chunks. As there
5260 * can be a significant delay between the individual batches being
5261 * recycled, this leads to the once large chunks of space being
5262 * fragmented and becoming unavailable for high-order allocations.
5269 * pcp->high and pcp->batch values are related and dependent on one another:
5270 * ->batch must never be higher then ->high.
5271 * The following function updates them in a safe manner without read side
5274 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5275 * those fields changing asynchronously (acording the the above rule).
5277 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5278 * outside of boot time (or some other assurance that no concurrent updaters
5281 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5282 unsigned long batch)
5284 /* start with a fail safe value for batch */
5288 /* Update high, then batch, in order */
5295 /* a companion to pageset_set_high() */
5296 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5298 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5301 static void pageset_init(struct per_cpu_pageset *p)
5303 struct per_cpu_pages *pcp;
5306 memset(p, 0, sizeof(*p));
5310 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5311 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5314 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5317 pageset_set_batch(p, batch);
5321 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5322 * to the value high for the pageset p.
5324 static void pageset_set_high(struct per_cpu_pageset *p,
5327 unsigned long batch = max(1UL, high / 4);
5328 if ((high / 4) > (PAGE_SHIFT * 8))
5329 batch = PAGE_SHIFT * 8;
5331 pageset_update(&p->pcp, high, batch);
5334 static void pageset_set_high_and_batch(struct zone *zone,
5335 struct per_cpu_pageset *pcp)
5337 if (percpu_pagelist_fraction)
5338 pageset_set_high(pcp,
5339 (zone->managed_pages /
5340 percpu_pagelist_fraction));
5342 pageset_set_batch(pcp, zone_batchsize(zone));
5345 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5347 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5350 pageset_set_high_and_batch(zone, pcp);
5353 static void __meminit setup_zone_pageset(struct zone *zone)
5356 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5357 for_each_possible_cpu(cpu)
5358 zone_pageset_init(zone, cpu);
5362 * Allocate per cpu pagesets and initialize them.
5363 * Before this call only boot pagesets were available.
5365 void __init setup_per_cpu_pageset(void)
5369 for_each_populated_zone(zone)
5370 setup_zone_pageset(zone);
5373 static noinline __init_refok
5374 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5380 * The per-page waitqueue mechanism uses hashed waitqueues
5383 zone->wait_table_hash_nr_entries =
5384 wait_table_hash_nr_entries(zone_size_pages);
5385 zone->wait_table_bits =
5386 wait_table_bits(zone->wait_table_hash_nr_entries);
5387 alloc_size = zone->wait_table_hash_nr_entries
5388 * sizeof(wait_queue_head_t);
5390 if (!slab_is_available()) {
5391 zone->wait_table = (wait_queue_head_t *)
5392 memblock_virt_alloc_node_nopanic(
5393 alloc_size, zone->zone_pgdat->node_id);
5396 * This case means that a zone whose size was 0 gets new memory
5397 * via memory hot-add.
5398 * But it may be the case that a new node was hot-added. In
5399 * this case vmalloc() will not be able to use this new node's
5400 * memory - this wait_table must be initialized to use this new
5401 * node itself as well.
5402 * To use this new node's memory, further consideration will be
5405 zone->wait_table = vmalloc(alloc_size);
5407 if (!zone->wait_table)
5410 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5411 init_waitqueue_head(zone->wait_table + i);
5416 static __meminit void zone_pcp_init(struct zone *zone)
5419 * per cpu subsystem is not up at this point. The following code
5420 * relies on the ability of the linker to provide the
5421 * offset of a (static) per cpu variable into the per cpu area.
5423 zone->pageset = &boot_pageset;
5425 if (populated_zone(zone))
5426 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5427 zone->name, zone->present_pages,
5428 zone_batchsize(zone));
5431 int __meminit init_currently_empty_zone(struct zone *zone,
5432 unsigned long zone_start_pfn,
5435 struct pglist_data *pgdat = zone->zone_pgdat;
5437 ret = zone_wait_table_init(zone, size);
5440 pgdat->nr_zones = zone_idx(zone) + 1;
5442 zone->zone_start_pfn = zone_start_pfn;
5444 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5445 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5447 (unsigned long)zone_idx(zone),
5448 zone_start_pfn, (zone_start_pfn + size));
5450 zone_init_free_lists(zone);
5455 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5456 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5459 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5461 int __meminit __early_pfn_to_nid(unsigned long pfn,
5462 struct mminit_pfnnid_cache *state)
5464 unsigned long start_pfn, end_pfn;
5467 if (state->last_start <= pfn && pfn < state->last_end)
5468 return state->last_nid;
5470 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5472 state->last_start = start_pfn;
5473 state->last_end = end_pfn;
5474 state->last_nid = nid;
5479 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5482 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5483 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5484 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5486 * If an architecture guarantees that all ranges registered contain no holes
5487 * and may be freed, this this function may be used instead of calling
5488 * memblock_free_early_nid() manually.
5490 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5492 unsigned long start_pfn, end_pfn;
5495 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5496 start_pfn = min(start_pfn, max_low_pfn);
5497 end_pfn = min(end_pfn, max_low_pfn);
5499 if (start_pfn < end_pfn)
5500 memblock_free_early_nid(PFN_PHYS(start_pfn),
5501 (end_pfn - start_pfn) << PAGE_SHIFT,
5507 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5508 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5510 * If an architecture guarantees that all ranges registered contain no holes and may
5511 * be freed, this function may be used instead of calling memory_present() manually.
5513 void __init sparse_memory_present_with_active_regions(int nid)
5515 unsigned long start_pfn, end_pfn;
5518 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5519 memory_present(this_nid, start_pfn, end_pfn);
5523 * get_pfn_range_for_nid - Return the start and end page frames for a node
5524 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5525 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5526 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5528 * It returns the start and end page frame of a node based on information
5529 * provided by memblock_set_node(). If called for a node
5530 * with no available memory, a warning is printed and the start and end
5533 void __meminit get_pfn_range_for_nid(unsigned int nid,
5534 unsigned long *start_pfn, unsigned long *end_pfn)
5536 unsigned long this_start_pfn, this_end_pfn;
5542 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5543 *start_pfn = min(*start_pfn, this_start_pfn);
5544 *end_pfn = max(*end_pfn, this_end_pfn);
5547 if (*start_pfn == -1UL)
5552 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5553 * assumption is made that zones within a node are ordered in monotonic
5554 * increasing memory addresses so that the "highest" populated zone is used
5556 static void __init find_usable_zone_for_movable(void)
5559 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5560 if (zone_index == ZONE_MOVABLE)
5563 if (arch_zone_highest_possible_pfn[zone_index] >
5564 arch_zone_lowest_possible_pfn[zone_index])
5568 VM_BUG_ON(zone_index == -1);
5569 movable_zone = zone_index;
5573 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5574 * because it is sized independent of architecture. Unlike the other zones,
5575 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5576 * in each node depending on the size of each node and how evenly kernelcore
5577 * is distributed. This helper function adjusts the zone ranges
5578 * provided by the architecture for a given node by using the end of the
5579 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5580 * zones within a node are in order of monotonic increases memory addresses
5582 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5583 unsigned long zone_type,
5584 unsigned long node_start_pfn,
5585 unsigned long node_end_pfn,
5586 unsigned long *zone_start_pfn,
5587 unsigned long *zone_end_pfn)
5589 /* Only adjust if ZONE_MOVABLE is on this node */
5590 if (zone_movable_pfn[nid]) {
5591 /* Size ZONE_MOVABLE */
5592 if (zone_type == ZONE_MOVABLE) {
5593 *zone_start_pfn = zone_movable_pfn[nid];
5594 *zone_end_pfn = min(node_end_pfn,
5595 arch_zone_highest_possible_pfn[movable_zone]);
5597 /* Check if this whole range is within ZONE_MOVABLE */
5598 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5599 *zone_start_pfn = *zone_end_pfn;
5604 * Return the number of pages a zone spans in a node, including holes
5605 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5607 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5608 unsigned long zone_type,
5609 unsigned long node_start_pfn,
5610 unsigned long node_end_pfn,
5611 unsigned long *zone_start_pfn,
5612 unsigned long *zone_end_pfn,
5613 unsigned long *ignored)
5615 /* When hotadd a new node from cpu_up(), the node should be empty */
5616 if (!node_start_pfn && !node_end_pfn)
5619 /* Get the start and end of the zone */
5620 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5621 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5622 adjust_zone_range_for_zone_movable(nid, zone_type,
5623 node_start_pfn, node_end_pfn,
5624 zone_start_pfn, zone_end_pfn);
5626 /* Check that this node has pages within the zone's required range */
5627 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5630 /* Move the zone boundaries inside the node if necessary */
5631 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5632 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5634 /* Return the spanned pages */
5635 return *zone_end_pfn - *zone_start_pfn;
5639 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5640 * then all holes in the requested range will be accounted for.
5642 unsigned long __meminit __absent_pages_in_range(int nid,
5643 unsigned long range_start_pfn,
5644 unsigned long range_end_pfn)
5646 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5647 unsigned long start_pfn, end_pfn;
5650 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5651 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5652 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5653 nr_absent -= end_pfn - start_pfn;
5659 * absent_pages_in_range - Return number of page frames in holes within a range
5660 * @start_pfn: The start PFN to start searching for holes
5661 * @end_pfn: The end PFN to stop searching for holes
5663 * It returns the number of pages frames in memory holes within a range.
5665 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5666 unsigned long end_pfn)
5668 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5671 /* Return the number of page frames in holes in a zone on a node */
5672 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5673 unsigned long zone_type,
5674 unsigned long node_start_pfn,
5675 unsigned long node_end_pfn,
5676 unsigned long *ignored)
5678 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5679 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5680 unsigned long zone_start_pfn, zone_end_pfn;
5681 unsigned long nr_absent;
5683 /* When hotadd a new node from cpu_up(), the node should be empty */
5684 if (!node_start_pfn && !node_end_pfn)
5687 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5688 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5690 adjust_zone_range_for_zone_movable(nid, zone_type,
5691 node_start_pfn, node_end_pfn,
5692 &zone_start_pfn, &zone_end_pfn);
5693 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5696 * ZONE_MOVABLE handling.
5697 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5700 if (zone_movable_pfn[nid]) {
5701 if (mirrored_kernelcore) {
5702 unsigned long start_pfn, end_pfn;
5703 struct memblock_region *r;
5705 for_each_memblock(memory, r) {
5706 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5707 zone_start_pfn, zone_end_pfn);
5708 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5709 zone_start_pfn, zone_end_pfn);
5711 if (zone_type == ZONE_MOVABLE &&
5712 memblock_is_mirror(r))
5713 nr_absent += end_pfn - start_pfn;
5715 if (zone_type == ZONE_NORMAL &&
5716 !memblock_is_mirror(r))
5717 nr_absent += end_pfn - start_pfn;
5720 if (zone_type == ZONE_NORMAL)
5721 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5728 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5729 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5730 unsigned long zone_type,
5731 unsigned long node_start_pfn,
5732 unsigned long node_end_pfn,
5733 unsigned long *zone_start_pfn,
5734 unsigned long *zone_end_pfn,
5735 unsigned long *zones_size)
5739 *zone_start_pfn = node_start_pfn;
5740 for (zone = 0; zone < zone_type; zone++)
5741 *zone_start_pfn += zones_size[zone];
5743 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5745 return zones_size[zone_type];
5748 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5749 unsigned long zone_type,
5750 unsigned long node_start_pfn,
5751 unsigned long node_end_pfn,
5752 unsigned long *zholes_size)
5757 return zholes_size[zone_type];
5760 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5762 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5763 unsigned long node_start_pfn,
5764 unsigned long node_end_pfn,
5765 unsigned long *zones_size,
5766 unsigned long *zholes_size)
5768 unsigned long realtotalpages = 0, totalpages = 0;
5771 for (i = 0; i < MAX_NR_ZONES; i++) {
5772 struct zone *zone = pgdat->node_zones + i;
5773 unsigned long zone_start_pfn, zone_end_pfn;
5774 unsigned long size, real_size;
5776 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5782 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5783 node_start_pfn, node_end_pfn,
5786 zone->zone_start_pfn = zone_start_pfn;
5788 zone->zone_start_pfn = 0;
5789 zone->spanned_pages = size;
5790 zone->present_pages = real_size;
5793 realtotalpages += real_size;
5796 pgdat->node_spanned_pages = totalpages;
5797 pgdat->node_present_pages = realtotalpages;
5798 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5802 #ifndef CONFIG_SPARSEMEM
5804 * Calculate the size of the zone->blockflags rounded to an unsigned long
5805 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5806 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5807 * round what is now in bits to nearest long in bits, then return it in
5810 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5812 unsigned long usemapsize;
5814 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5815 usemapsize = roundup(zonesize, pageblock_nr_pages);
5816 usemapsize = usemapsize >> pageblock_order;
5817 usemapsize *= NR_PAGEBLOCK_BITS;
5818 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5820 return usemapsize / 8;
5823 static void __init setup_usemap(struct pglist_data *pgdat,
5825 unsigned long zone_start_pfn,
5826 unsigned long zonesize)
5828 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5829 zone->pageblock_flags = NULL;
5831 zone->pageblock_flags =
5832 memblock_virt_alloc_node_nopanic(usemapsize,
5836 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5837 unsigned long zone_start_pfn, unsigned long zonesize) {}
5838 #endif /* CONFIG_SPARSEMEM */
5840 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5842 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5843 void __paginginit set_pageblock_order(void)
5847 /* Check that pageblock_nr_pages has not already been setup */
5848 if (pageblock_order)
5851 if (HPAGE_SHIFT > PAGE_SHIFT)
5852 order = HUGETLB_PAGE_ORDER;
5854 order = MAX_ORDER - 1;
5857 * Assume the largest contiguous order of interest is a huge page.
5858 * This value may be variable depending on boot parameters on IA64 and
5861 pageblock_order = order;
5863 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5866 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5867 * is unused as pageblock_order is set at compile-time. See
5868 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5871 void __paginginit set_pageblock_order(void)
5875 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5877 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5878 unsigned long present_pages)
5880 unsigned long pages = spanned_pages;
5883 * Provide a more accurate estimation if there are holes within
5884 * the zone and SPARSEMEM is in use. If there are holes within the
5885 * zone, each populated memory region may cost us one or two extra
5886 * memmap pages due to alignment because memmap pages for each
5887 * populated regions may not naturally algined on page boundary.
5888 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5890 if (spanned_pages > present_pages + (present_pages >> 4) &&
5891 IS_ENABLED(CONFIG_SPARSEMEM))
5892 pages = present_pages;
5894 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5898 * Set up the zone data structures:
5899 * - mark all pages reserved
5900 * - mark all memory queues empty
5901 * - clear the memory bitmaps
5903 * NOTE: pgdat should get zeroed by caller.
5905 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5908 int nid = pgdat->node_id;
5911 pgdat_resize_init(pgdat);
5912 #ifdef CONFIG_NUMA_BALANCING
5913 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5914 pgdat->numabalancing_migrate_nr_pages = 0;
5915 pgdat->numabalancing_migrate_next_window = jiffies;
5917 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5918 spin_lock_init(&pgdat->split_queue_lock);
5919 INIT_LIST_HEAD(&pgdat->split_queue);
5920 pgdat->split_queue_len = 0;
5922 init_waitqueue_head(&pgdat->kswapd_wait);
5923 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5924 #ifdef CONFIG_COMPACTION
5925 init_waitqueue_head(&pgdat->kcompactd_wait);
5927 pgdat_page_ext_init(pgdat);
5929 for (j = 0; j < MAX_NR_ZONES; j++) {
5930 struct zone *zone = pgdat->node_zones + j;
5931 unsigned long size, realsize, freesize, memmap_pages;
5932 unsigned long zone_start_pfn = zone->zone_start_pfn;
5934 size = zone->spanned_pages;
5935 realsize = freesize = zone->present_pages;
5938 * Adjust freesize so that it accounts for how much memory
5939 * is used by this zone for memmap. This affects the watermark
5940 * and per-cpu initialisations
5942 memmap_pages = calc_memmap_size(size, realsize);
5943 if (!is_highmem_idx(j)) {
5944 if (freesize >= memmap_pages) {
5945 freesize -= memmap_pages;
5948 " %s zone: %lu pages used for memmap\n",
5949 zone_names[j], memmap_pages);
5951 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5952 zone_names[j], memmap_pages, freesize);
5955 /* Account for reserved pages */
5956 if (j == 0 && freesize > dma_reserve) {
5957 freesize -= dma_reserve;
5958 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5959 zone_names[0], dma_reserve);
5962 if (!is_highmem_idx(j))
5963 nr_kernel_pages += freesize;
5964 /* Charge for highmem memmap if there are enough kernel pages */
5965 else if (nr_kernel_pages > memmap_pages * 2)
5966 nr_kernel_pages -= memmap_pages;
5967 nr_all_pages += freesize;
5970 * Set an approximate value for lowmem here, it will be adjusted
5971 * when the bootmem allocator frees pages into the buddy system.
5972 * And all highmem pages will be managed by the buddy system.
5974 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5977 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5979 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5981 zone->name = zone_names[j];
5982 spin_lock_init(&zone->lock);
5983 spin_lock_init(&zone->lru_lock);
5984 zone_seqlock_init(zone);
5985 zone->zone_pgdat = pgdat;
5986 zone_pcp_init(zone);
5988 /* For bootup, initialized properly in watermark setup */
5989 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5991 lruvec_init(&zone->lruvec);
5995 set_pageblock_order();
5996 setup_usemap(pgdat, zone, zone_start_pfn, size);
5997 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5999 memmap_init(size, nid, j, zone_start_pfn);
6003 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
6005 unsigned long __maybe_unused start = 0;
6006 unsigned long __maybe_unused offset = 0;
6008 /* Skip empty nodes */
6009 if (!pgdat->node_spanned_pages)
6012 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6013 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6014 offset = pgdat->node_start_pfn - start;
6015 /* ia64 gets its own node_mem_map, before this, without bootmem */
6016 if (!pgdat->node_mem_map) {
6017 unsigned long size, end;
6021 * The zone's endpoints aren't required to be MAX_ORDER
6022 * aligned but the node_mem_map endpoints must be in order
6023 * for the buddy allocator to function correctly.
6025 end = pgdat_end_pfn(pgdat);
6026 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6027 size = (end - start) * sizeof(struct page);
6028 map = alloc_remap(pgdat->node_id, size);
6030 map = memblock_virt_alloc_node_nopanic(size,
6032 pgdat->node_mem_map = map + offset;
6034 #ifndef CONFIG_NEED_MULTIPLE_NODES
6036 * With no DISCONTIG, the global mem_map is just set as node 0's
6038 if (pgdat == NODE_DATA(0)) {
6039 mem_map = NODE_DATA(0)->node_mem_map;
6040 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6041 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6043 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6046 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6049 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6050 unsigned long node_start_pfn, unsigned long *zholes_size)
6052 pg_data_t *pgdat = NODE_DATA(nid);
6053 unsigned long start_pfn = 0;
6054 unsigned long end_pfn = 0;
6056 /* pg_data_t should be reset to zero when it's allocated */
6057 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6059 reset_deferred_meminit(pgdat);
6060 pgdat->node_id = nid;
6061 pgdat->node_start_pfn = node_start_pfn;
6062 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6063 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6064 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6065 (u64)start_pfn << PAGE_SHIFT,
6066 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6068 start_pfn = node_start_pfn;
6070 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6071 zones_size, zholes_size);
6073 alloc_node_mem_map(pgdat);
6074 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6075 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6076 nid, (unsigned long)pgdat,
6077 (unsigned long)pgdat->node_mem_map);
6080 free_area_init_core(pgdat);
6083 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6085 #if MAX_NUMNODES > 1
6087 * Figure out the number of possible node ids.
6089 void __init setup_nr_node_ids(void)
6091 unsigned int highest;
6093 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6094 nr_node_ids = highest + 1;
6099 * node_map_pfn_alignment - determine the maximum internode alignment
6101 * This function should be called after node map is populated and sorted.
6102 * It calculates the maximum power of two alignment which can distinguish
6105 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6106 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6107 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6108 * shifted, 1GiB is enough and this function will indicate so.
6110 * This is used to test whether pfn -> nid mapping of the chosen memory
6111 * model has fine enough granularity to avoid incorrect mapping for the
6112 * populated node map.
6114 * Returns the determined alignment in pfn's. 0 if there is no alignment
6115 * requirement (single node).
6117 unsigned long __init node_map_pfn_alignment(void)
6119 unsigned long accl_mask = 0, last_end = 0;
6120 unsigned long start, end, mask;
6124 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6125 if (!start || last_nid < 0 || last_nid == nid) {
6132 * Start with a mask granular enough to pin-point to the
6133 * start pfn and tick off bits one-by-one until it becomes
6134 * too coarse to separate the current node from the last.
6136 mask = ~((1 << __ffs(start)) - 1);
6137 while (mask && last_end <= (start & (mask << 1)))
6140 /* accumulate all internode masks */
6144 /* convert mask to number of pages */
6145 return ~accl_mask + 1;
6148 /* Find the lowest pfn for a node */
6149 static unsigned long __init find_min_pfn_for_node(int nid)
6151 unsigned long min_pfn = ULONG_MAX;
6152 unsigned long start_pfn;
6155 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6156 min_pfn = min(min_pfn, start_pfn);
6158 if (min_pfn == ULONG_MAX) {
6159 pr_warn("Could not find start_pfn for node %d\n", nid);
6167 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6169 * It returns the minimum PFN based on information provided via
6170 * memblock_set_node().
6172 unsigned long __init find_min_pfn_with_active_regions(void)
6174 return find_min_pfn_for_node(MAX_NUMNODES);
6178 * early_calculate_totalpages()
6179 * Sum pages in active regions for movable zone.
6180 * Populate N_MEMORY for calculating usable_nodes.
6182 static unsigned long __init early_calculate_totalpages(void)
6184 unsigned long totalpages = 0;
6185 unsigned long start_pfn, end_pfn;
6188 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6189 unsigned long pages = end_pfn - start_pfn;
6191 totalpages += pages;
6193 node_set_state(nid, N_MEMORY);
6199 * Find the PFN the Movable zone begins in each node. Kernel memory
6200 * is spread evenly between nodes as long as the nodes have enough
6201 * memory. When they don't, some nodes will have more kernelcore than
6204 static void __init find_zone_movable_pfns_for_nodes(void)
6207 unsigned long usable_startpfn;
6208 unsigned long kernelcore_node, kernelcore_remaining;
6209 /* save the state before borrow the nodemask */
6210 nodemask_t saved_node_state = node_states[N_MEMORY];
6211 unsigned long totalpages = early_calculate_totalpages();
6212 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6213 struct memblock_region *r;
6215 /* Need to find movable_zone earlier when movable_node is specified. */
6216 find_usable_zone_for_movable();
6219 * If movable_node is specified, ignore kernelcore and movablecore
6222 if (movable_node_is_enabled()) {
6223 for_each_memblock(memory, r) {
6224 if (!memblock_is_hotpluggable(r))
6229 usable_startpfn = PFN_DOWN(r->base);
6230 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6231 min(usable_startpfn, zone_movable_pfn[nid]) :
6239 * If kernelcore=mirror is specified, ignore movablecore option
6241 if (mirrored_kernelcore) {
6242 bool mem_below_4gb_not_mirrored = false;
6244 for_each_memblock(memory, r) {
6245 if (memblock_is_mirror(r))
6250 usable_startpfn = memblock_region_memory_base_pfn(r);
6252 if (usable_startpfn < 0x100000) {
6253 mem_below_4gb_not_mirrored = true;
6257 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6258 min(usable_startpfn, zone_movable_pfn[nid]) :
6262 if (mem_below_4gb_not_mirrored)
6263 pr_warn("This configuration results in unmirrored kernel memory.");
6269 * If movablecore=nn[KMG] was specified, calculate what size of
6270 * kernelcore that corresponds so that memory usable for
6271 * any allocation type is evenly spread. If both kernelcore
6272 * and movablecore are specified, then the value of kernelcore
6273 * will be used for required_kernelcore if it's greater than
6274 * what movablecore would have allowed.
6276 if (required_movablecore) {
6277 unsigned long corepages;
6280 * Round-up so that ZONE_MOVABLE is at least as large as what
6281 * was requested by the user
6283 required_movablecore =
6284 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6285 required_movablecore = min(totalpages, required_movablecore);
6286 corepages = totalpages - required_movablecore;
6288 required_kernelcore = max(required_kernelcore, corepages);
6292 * If kernelcore was not specified or kernelcore size is larger
6293 * than totalpages, there is no ZONE_MOVABLE.
6295 if (!required_kernelcore || required_kernelcore >= totalpages)
6298 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6299 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6302 /* Spread kernelcore memory as evenly as possible throughout nodes */
6303 kernelcore_node = required_kernelcore / usable_nodes;
6304 for_each_node_state(nid, N_MEMORY) {
6305 unsigned long start_pfn, end_pfn;
6308 * Recalculate kernelcore_node if the division per node
6309 * now exceeds what is necessary to satisfy the requested
6310 * amount of memory for the kernel
6312 if (required_kernelcore < kernelcore_node)
6313 kernelcore_node = required_kernelcore / usable_nodes;
6316 * As the map is walked, we track how much memory is usable
6317 * by the kernel using kernelcore_remaining. When it is
6318 * 0, the rest of the node is usable by ZONE_MOVABLE
6320 kernelcore_remaining = kernelcore_node;
6322 /* Go through each range of PFNs within this node */
6323 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6324 unsigned long size_pages;
6326 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6327 if (start_pfn >= end_pfn)
6330 /* Account for what is only usable for kernelcore */
6331 if (start_pfn < usable_startpfn) {
6332 unsigned long kernel_pages;
6333 kernel_pages = min(end_pfn, usable_startpfn)
6336 kernelcore_remaining -= min(kernel_pages,
6337 kernelcore_remaining);
6338 required_kernelcore -= min(kernel_pages,
6339 required_kernelcore);
6341 /* Continue if range is now fully accounted */
6342 if (end_pfn <= usable_startpfn) {
6345 * Push zone_movable_pfn to the end so
6346 * that if we have to rebalance
6347 * kernelcore across nodes, we will
6348 * not double account here
6350 zone_movable_pfn[nid] = end_pfn;
6353 start_pfn = usable_startpfn;
6357 * The usable PFN range for ZONE_MOVABLE is from
6358 * start_pfn->end_pfn. Calculate size_pages as the
6359 * number of pages used as kernelcore
6361 size_pages = end_pfn - start_pfn;
6362 if (size_pages > kernelcore_remaining)
6363 size_pages = kernelcore_remaining;
6364 zone_movable_pfn[nid] = start_pfn + size_pages;
6367 * Some kernelcore has been met, update counts and
6368 * break if the kernelcore for this node has been
6371 required_kernelcore -= min(required_kernelcore,
6373 kernelcore_remaining -= size_pages;
6374 if (!kernelcore_remaining)
6380 * If there is still required_kernelcore, we do another pass with one
6381 * less node in the count. This will push zone_movable_pfn[nid] further
6382 * along on the nodes that still have memory until kernelcore is
6386 if (usable_nodes && required_kernelcore > usable_nodes)
6390 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6391 for (nid = 0; nid < MAX_NUMNODES; nid++)
6392 zone_movable_pfn[nid] =
6393 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6396 /* restore the node_state */
6397 node_states[N_MEMORY] = saved_node_state;
6400 /* Any regular or high memory on that node ? */
6401 static void check_for_memory(pg_data_t *pgdat, int nid)
6403 enum zone_type zone_type;
6405 if (N_MEMORY == N_NORMAL_MEMORY)
6408 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6409 struct zone *zone = &pgdat->node_zones[zone_type];
6410 if (populated_zone(zone)) {
6411 node_set_state(nid, N_HIGH_MEMORY);
6412 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6413 zone_type <= ZONE_NORMAL)
6414 node_set_state(nid, N_NORMAL_MEMORY);
6421 * free_area_init_nodes - Initialise all pg_data_t and zone data
6422 * @max_zone_pfn: an array of max PFNs for each zone
6424 * This will call free_area_init_node() for each active node in the system.
6425 * Using the page ranges provided by memblock_set_node(), the size of each
6426 * zone in each node and their holes is calculated. If the maximum PFN
6427 * between two adjacent zones match, it is assumed that the zone is empty.
6428 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6429 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6430 * starts where the previous one ended. For example, ZONE_DMA32 starts
6431 * at arch_max_dma_pfn.
6433 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6435 unsigned long start_pfn, end_pfn;
6438 /* Record where the zone boundaries are */
6439 memset(arch_zone_lowest_possible_pfn, 0,
6440 sizeof(arch_zone_lowest_possible_pfn));
6441 memset(arch_zone_highest_possible_pfn, 0,
6442 sizeof(arch_zone_highest_possible_pfn));
6444 start_pfn = find_min_pfn_with_active_regions();
6446 for (i = 0; i < MAX_NR_ZONES; i++) {
6447 if (i == ZONE_MOVABLE)
6450 end_pfn = max(max_zone_pfn[i], start_pfn);
6451 arch_zone_lowest_possible_pfn[i] = start_pfn;
6452 arch_zone_highest_possible_pfn[i] = end_pfn;
6454 start_pfn = end_pfn;
6456 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6457 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6459 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6460 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6461 find_zone_movable_pfns_for_nodes();
6463 /* Print out the zone ranges */
6464 pr_info("Zone ranges:\n");
6465 for (i = 0; i < MAX_NR_ZONES; i++) {
6466 if (i == ZONE_MOVABLE)
6468 pr_info(" %-8s ", zone_names[i]);
6469 if (arch_zone_lowest_possible_pfn[i] ==
6470 arch_zone_highest_possible_pfn[i])
6473 pr_cont("[mem %#018Lx-%#018Lx]\n",
6474 (u64)arch_zone_lowest_possible_pfn[i]
6476 ((u64)arch_zone_highest_possible_pfn[i]
6477 << PAGE_SHIFT) - 1);
6480 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6481 pr_info("Movable zone start for each node\n");
6482 for (i = 0; i < MAX_NUMNODES; i++) {
6483 if (zone_movable_pfn[i])
6484 pr_info(" Node %d: %#018Lx\n", i,
6485 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6488 /* Print out the early node map */
6489 pr_info("Early memory node ranges\n");
6490 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6491 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6492 (u64)start_pfn << PAGE_SHIFT,
6493 ((u64)end_pfn << PAGE_SHIFT) - 1);
6495 /* Initialise every node */
6496 mminit_verify_pageflags_layout();
6497 setup_nr_node_ids();
6498 for_each_online_node(nid) {
6499 pg_data_t *pgdat = NODE_DATA(nid);
6500 free_area_init_node(nid, NULL,
6501 find_min_pfn_for_node(nid), NULL);
6503 /* Any memory on that node */
6504 if (pgdat->node_present_pages)
6505 node_set_state(nid, N_MEMORY);
6506 check_for_memory(pgdat, nid);
6510 static int __init cmdline_parse_core(char *p, unsigned long *core)
6512 unsigned long long coremem;
6516 coremem = memparse(p, &p);
6517 *core = coremem >> PAGE_SHIFT;
6519 /* Paranoid check that UL is enough for the coremem value */
6520 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6526 * kernelcore=size sets the amount of memory for use for allocations that
6527 * cannot be reclaimed or migrated.
6529 static int __init cmdline_parse_kernelcore(char *p)
6531 /* parse kernelcore=mirror */
6532 if (parse_option_str(p, "mirror")) {
6533 mirrored_kernelcore = true;
6537 return cmdline_parse_core(p, &required_kernelcore);
6541 * movablecore=size sets the amount of memory for use for allocations that
6542 * can be reclaimed or migrated.
6544 static int __init cmdline_parse_movablecore(char *p)
6546 return cmdline_parse_core(p, &required_movablecore);
6549 early_param("kernelcore", cmdline_parse_kernelcore);
6550 early_param("movablecore", cmdline_parse_movablecore);
6552 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6554 void adjust_managed_page_count(struct page *page, long count)
6556 spin_lock(&managed_page_count_lock);
6557 page_zone(page)->managed_pages += count;
6558 totalram_pages += count;
6559 #ifdef CONFIG_HIGHMEM
6560 if (PageHighMem(page))
6561 totalhigh_pages += count;
6563 spin_unlock(&managed_page_count_lock);
6565 EXPORT_SYMBOL(adjust_managed_page_count);
6567 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6570 unsigned long pages = 0;
6572 start = (void *)PAGE_ALIGN((unsigned long)start);
6573 end = (void *)((unsigned long)end & PAGE_MASK);
6574 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6575 if ((unsigned int)poison <= 0xFF)
6576 memset(pos, poison, PAGE_SIZE);
6577 free_reserved_page(virt_to_page(pos));
6581 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6582 s, pages << (PAGE_SHIFT - 10), start, end);
6586 EXPORT_SYMBOL(free_reserved_area);
6588 #ifdef CONFIG_HIGHMEM
6589 void free_highmem_page(struct page *page)
6591 __free_reserved_page(page);
6593 page_zone(page)->managed_pages++;
6599 void __init mem_init_print_info(const char *str)
6601 unsigned long physpages, codesize, datasize, rosize, bss_size;
6602 unsigned long init_code_size, init_data_size;
6604 physpages = get_num_physpages();
6605 codesize = _etext - _stext;
6606 datasize = _edata - _sdata;
6607 rosize = __end_rodata - __start_rodata;
6608 bss_size = __bss_stop - __bss_start;
6609 init_data_size = __init_end - __init_begin;
6610 init_code_size = _einittext - _sinittext;
6613 * Detect special cases and adjust section sizes accordingly:
6614 * 1) .init.* may be embedded into .data sections
6615 * 2) .init.text.* may be out of [__init_begin, __init_end],
6616 * please refer to arch/tile/kernel/vmlinux.lds.S.
6617 * 3) .rodata.* may be embedded into .text or .data sections.
6619 #define adj_init_size(start, end, size, pos, adj) \
6621 if (start <= pos && pos < end && size > adj) \
6625 adj_init_size(__init_begin, __init_end, init_data_size,
6626 _sinittext, init_code_size);
6627 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6628 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6629 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6630 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6632 #undef adj_init_size
6634 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6635 #ifdef CONFIG_HIGHMEM
6639 nr_free_pages() << (PAGE_SHIFT - 10),
6640 physpages << (PAGE_SHIFT - 10),
6641 codesize >> 10, datasize >> 10, rosize >> 10,
6642 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6643 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6644 totalcma_pages << (PAGE_SHIFT - 10),
6645 #ifdef CONFIG_HIGHMEM
6646 totalhigh_pages << (PAGE_SHIFT - 10),
6648 str ? ", " : "", str ? str : "");
6652 * set_dma_reserve - set the specified number of pages reserved in the first zone
6653 * @new_dma_reserve: The number of pages to mark reserved
6655 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6656 * In the DMA zone, a significant percentage may be consumed by kernel image
6657 * and other unfreeable allocations which can skew the watermarks badly. This
6658 * function may optionally be used to account for unfreeable pages in the
6659 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6660 * smaller per-cpu batchsize.
6662 void __init set_dma_reserve(unsigned long new_dma_reserve)
6664 dma_reserve = new_dma_reserve;
6667 void __init free_area_init(unsigned long *zones_size)
6669 free_area_init_node(0, zones_size,
6670 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6673 static int page_alloc_cpu_notify(struct notifier_block *self,
6674 unsigned long action, void *hcpu)
6676 int cpu = (unsigned long)hcpu;
6678 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6679 lru_add_drain_cpu(cpu);
6683 * Spill the event counters of the dead processor
6684 * into the current processors event counters.
6685 * This artificially elevates the count of the current
6688 vm_events_fold_cpu(cpu);
6691 * Zero the differential counters of the dead processor
6692 * so that the vm statistics are consistent.
6694 * This is only okay since the processor is dead and cannot
6695 * race with what we are doing.
6697 cpu_vm_stats_fold(cpu);
6702 void __init page_alloc_init(void)
6704 hotcpu_notifier(page_alloc_cpu_notify, 0);
6708 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6709 * or min_free_kbytes changes.
6711 static void calculate_totalreserve_pages(void)
6713 struct pglist_data *pgdat;
6714 unsigned long reserve_pages = 0;
6715 enum zone_type i, j;
6717 for_each_online_pgdat(pgdat) {
6718 for (i = 0; i < MAX_NR_ZONES; i++) {
6719 struct zone *zone = pgdat->node_zones + i;
6722 /* Find valid and maximum lowmem_reserve in the zone */
6723 for (j = i; j < MAX_NR_ZONES; j++) {
6724 if (zone->lowmem_reserve[j] > max)
6725 max = zone->lowmem_reserve[j];
6728 /* we treat the high watermark as reserved pages. */
6729 max += high_wmark_pages(zone);
6731 if (max > zone->managed_pages)
6732 max = zone->managed_pages;
6734 zone->totalreserve_pages = max;
6736 reserve_pages += max;
6739 totalreserve_pages = reserve_pages;
6743 * setup_per_zone_lowmem_reserve - called whenever
6744 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6745 * has a correct pages reserved value, so an adequate number of
6746 * pages are left in the zone after a successful __alloc_pages().
6748 static void setup_per_zone_lowmem_reserve(void)
6750 struct pglist_data *pgdat;
6751 enum zone_type j, idx;
6753 for_each_online_pgdat(pgdat) {
6754 for (j = 0; j < MAX_NR_ZONES; j++) {
6755 struct zone *zone = pgdat->node_zones + j;
6756 unsigned long managed_pages = zone->managed_pages;
6758 zone->lowmem_reserve[j] = 0;
6762 struct zone *lower_zone;
6766 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6767 sysctl_lowmem_reserve_ratio[idx] = 1;
6769 lower_zone = pgdat->node_zones + idx;
6770 lower_zone->lowmem_reserve[j] = managed_pages /
6771 sysctl_lowmem_reserve_ratio[idx];
6772 managed_pages += lower_zone->managed_pages;
6777 /* update totalreserve_pages */
6778 calculate_totalreserve_pages();
6781 static void __setup_per_zone_wmarks(void)
6783 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6784 unsigned long lowmem_pages = 0;
6786 unsigned long flags;
6788 /* Calculate total number of !ZONE_HIGHMEM pages */
6789 for_each_zone(zone) {
6790 if (!is_highmem(zone))
6791 lowmem_pages += zone->managed_pages;
6794 for_each_zone(zone) {
6797 spin_lock_irqsave(&zone->lock, flags);
6798 tmp = (u64)pages_min * zone->managed_pages;
6799 do_div(tmp, lowmem_pages);
6800 if (is_highmem(zone)) {
6802 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6803 * need highmem pages, so cap pages_min to a small
6806 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6807 * deltas control asynch page reclaim, and so should
6808 * not be capped for highmem.
6810 unsigned long min_pages;
6812 min_pages = zone->managed_pages / 1024;
6813 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6814 zone->watermark[WMARK_MIN] = min_pages;
6817 * If it's a lowmem zone, reserve a number of pages
6818 * proportionate to the zone's size.
6820 zone->watermark[WMARK_MIN] = tmp;
6824 * Set the kswapd watermarks distance according to the
6825 * scale factor in proportion to available memory, but
6826 * ensure a minimum size on small systems.
6828 tmp = max_t(u64, tmp >> 2,
6829 mult_frac(zone->managed_pages,
6830 watermark_scale_factor, 10000));
6832 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6833 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6835 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6836 high_wmark_pages(zone) - low_wmark_pages(zone) -
6837 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6839 spin_unlock_irqrestore(&zone->lock, flags);
6842 /* update totalreserve_pages */
6843 calculate_totalreserve_pages();
6847 * setup_per_zone_wmarks - called when min_free_kbytes changes
6848 * or when memory is hot-{added|removed}
6850 * Ensures that the watermark[min,low,high] values for each zone are set
6851 * correctly with respect to min_free_kbytes.
6853 void setup_per_zone_wmarks(void)
6855 mutex_lock(&zonelists_mutex);
6856 __setup_per_zone_wmarks();
6857 mutex_unlock(&zonelists_mutex);
6861 * Initialise min_free_kbytes.
6863 * For small machines we want it small (128k min). For large machines
6864 * we want it large (64MB max). But it is not linear, because network
6865 * bandwidth does not increase linearly with machine size. We use
6867 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6868 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6884 int __meminit init_per_zone_wmark_min(void)
6886 unsigned long lowmem_kbytes;
6887 int new_min_free_kbytes;
6889 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6890 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6892 if (new_min_free_kbytes > user_min_free_kbytes) {
6893 min_free_kbytes = new_min_free_kbytes;
6894 if (min_free_kbytes < 128)
6895 min_free_kbytes = 128;
6896 if (min_free_kbytes > 65536)
6897 min_free_kbytes = 65536;
6899 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6900 new_min_free_kbytes, user_min_free_kbytes);
6902 setup_per_zone_wmarks();
6903 refresh_zone_stat_thresholds();
6904 setup_per_zone_lowmem_reserve();
6907 core_initcall(init_per_zone_wmark_min)
6910 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6911 * that we can call two helper functions whenever min_free_kbytes
6914 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6915 void __user *buffer, size_t *length, loff_t *ppos)
6919 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6924 user_min_free_kbytes = min_free_kbytes;
6925 setup_per_zone_wmarks();
6930 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6931 void __user *buffer, size_t *length, loff_t *ppos)
6935 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6940 setup_per_zone_wmarks();
6946 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6947 void __user *buffer, size_t *length, loff_t *ppos)
6952 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6957 zone->min_unmapped_pages = (zone->managed_pages *
6958 sysctl_min_unmapped_ratio) / 100;
6962 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6963 void __user *buffer, size_t *length, loff_t *ppos)
6968 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6973 zone->min_slab_pages = (zone->managed_pages *
6974 sysctl_min_slab_ratio) / 100;
6980 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6981 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6982 * whenever sysctl_lowmem_reserve_ratio changes.
6984 * The reserve ratio obviously has absolutely no relation with the
6985 * minimum watermarks. The lowmem reserve ratio can only make sense
6986 * if in function of the boot time zone sizes.
6988 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6989 void __user *buffer, size_t *length, loff_t *ppos)
6991 proc_dointvec_minmax(table, write, buffer, length, ppos);
6992 setup_per_zone_lowmem_reserve();
6997 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6998 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6999 * pagelist can have before it gets flushed back to buddy allocator.
7001 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7002 void __user *buffer, size_t *length, loff_t *ppos)
7005 int old_percpu_pagelist_fraction;
7008 mutex_lock(&pcp_batch_high_lock);
7009 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7011 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7012 if (!write || ret < 0)
7015 /* Sanity checking to avoid pcp imbalance */
7016 if (percpu_pagelist_fraction &&
7017 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7018 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7024 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7027 for_each_populated_zone(zone) {
7030 for_each_possible_cpu(cpu)
7031 pageset_set_high_and_batch(zone,
7032 per_cpu_ptr(zone->pageset, cpu));
7035 mutex_unlock(&pcp_batch_high_lock);
7040 int hashdist = HASHDIST_DEFAULT;
7042 static int __init set_hashdist(char *str)
7046 hashdist = simple_strtoul(str, &str, 0);
7049 __setup("hashdist=", set_hashdist);
7053 * allocate a large system hash table from bootmem
7054 * - it is assumed that the hash table must contain an exact power-of-2
7055 * quantity of entries
7056 * - limit is the number of hash buckets, not the total allocation size
7058 void *__init alloc_large_system_hash(const char *tablename,
7059 unsigned long bucketsize,
7060 unsigned long numentries,
7063 unsigned int *_hash_shift,
7064 unsigned int *_hash_mask,
7065 unsigned long low_limit,
7066 unsigned long high_limit)
7068 unsigned long long max = high_limit;
7069 unsigned long log2qty, size;
7072 /* allow the kernel cmdline to have a say */
7074 /* round applicable memory size up to nearest megabyte */
7075 numentries = nr_kernel_pages;
7077 /* It isn't necessary when PAGE_SIZE >= 1MB */
7078 if (PAGE_SHIFT < 20)
7079 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7081 /* limit to 1 bucket per 2^scale bytes of low memory */
7082 if (scale > PAGE_SHIFT)
7083 numentries >>= (scale - PAGE_SHIFT);
7085 numentries <<= (PAGE_SHIFT - scale);
7087 /* Make sure we've got at least a 0-order allocation.. */
7088 if (unlikely(flags & HASH_SMALL)) {
7089 /* Makes no sense without HASH_EARLY */
7090 WARN_ON(!(flags & HASH_EARLY));
7091 if (!(numentries >> *_hash_shift)) {
7092 numentries = 1UL << *_hash_shift;
7093 BUG_ON(!numentries);
7095 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7096 numentries = PAGE_SIZE / bucketsize;
7098 numentries = roundup_pow_of_two(numentries);
7100 /* limit allocation size to 1/16 total memory by default */
7102 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7103 do_div(max, bucketsize);
7105 max = min(max, 0x80000000ULL);
7107 if (numentries < low_limit)
7108 numentries = low_limit;
7109 if (numentries > max)
7112 log2qty = ilog2(numentries);
7115 size = bucketsize << log2qty;
7116 if (flags & HASH_EARLY)
7117 table = memblock_virt_alloc_nopanic(size, 0);
7119 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7122 * If bucketsize is not a power-of-two, we may free
7123 * some pages at the end of hash table which
7124 * alloc_pages_exact() automatically does
7126 if (get_order(size) < MAX_ORDER) {
7127 table = alloc_pages_exact(size, GFP_ATOMIC);
7128 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7131 } while (!table && size > PAGE_SIZE && --log2qty);
7134 panic("Failed to allocate %s hash table\n", tablename);
7136 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7137 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7140 *_hash_shift = log2qty;
7142 *_hash_mask = (1 << log2qty) - 1;
7148 * This function checks whether pageblock includes unmovable pages or not.
7149 * If @count is not zero, it is okay to include less @count unmovable pages
7151 * PageLRU check without isolation or lru_lock could race so that
7152 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7153 * expect this function should be exact.
7155 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7156 bool skip_hwpoisoned_pages)
7158 unsigned long pfn, iter, found;
7162 * For avoiding noise data, lru_add_drain_all() should be called
7163 * If ZONE_MOVABLE, the zone never contains unmovable pages
7165 if (zone_idx(zone) == ZONE_MOVABLE)
7167 mt = get_pageblock_migratetype(page);
7168 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7171 pfn = page_to_pfn(page);
7172 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7173 unsigned long check = pfn + iter;
7175 if (!pfn_valid_within(check))
7178 page = pfn_to_page(check);
7181 * Hugepages are not in LRU lists, but they're movable.
7182 * We need not scan over tail pages bacause we don't
7183 * handle each tail page individually in migration.
7185 if (PageHuge(page)) {
7186 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7191 * We can't use page_count without pin a page
7192 * because another CPU can free compound page.
7193 * This check already skips compound tails of THP
7194 * because their page->_refcount is zero at all time.
7196 if (!page_ref_count(page)) {
7197 if (PageBuddy(page))
7198 iter += (1 << page_order(page)) - 1;
7203 * The HWPoisoned page may be not in buddy system, and
7204 * page_count() is not 0.
7206 if (skip_hwpoisoned_pages && PageHWPoison(page))
7212 * If there are RECLAIMABLE pages, we need to check
7213 * it. But now, memory offline itself doesn't call
7214 * shrink_node_slabs() and it still to be fixed.
7217 * If the page is not RAM, page_count()should be 0.
7218 * we don't need more check. This is an _used_ not-movable page.
7220 * The problematic thing here is PG_reserved pages. PG_reserved
7221 * is set to both of a memory hole page and a _used_ kernel
7230 bool is_pageblock_removable_nolock(struct page *page)
7236 * We have to be careful here because we are iterating over memory
7237 * sections which are not zone aware so we might end up outside of
7238 * the zone but still within the section.
7239 * We have to take care about the node as well. If the node is offline
7240 * its NODE_DATA will be NULL - see page_zone.
7242 if (!node_online(page_to_nid(page)))
7245 zone = page_zone(page);
7246 pfn = page_to_pfn(page);
7247 if (!zone_spans_pfn(zone, pfn))
7250 return !has_unmovable_pages(zone, page, 0, true);
7253 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7255 static unsigned long pfn_max_align_down(unsigned long pfn)
7257 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7258 pageblock_nr_pages) - 1);
7261 static unsigned long pfn_max_align_up(unsigned long pfn)
7263 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7264 pageblock_nr_pages));
7267 /* [start, end) must belong to a single zone. */
7268 static int __alloc_contig_migrate_range(struct compact_control *cc,
7269 unsigned long start, unsigned long end)
7271 /* This function is based on compact_zone() from compaction.c. */
7272 unsigned long nr_reclaimed;
7273 unsigned long pfn = start;
7274 unsigned int tries = 0;
7279 while (pfn < end || !list_empty(&cc->migratepages)) {
7280 if (fatal_signal_pending(current)) {
7285 if (list_empty(&cc->migratepages)) {
7286 cc->nr_migratepages = 0;
7287 pfn = isolate_migratepages_range(cc, pfn, end);
7293 } else if (++tries == 5) {
7294 ret = ret < 0 ? ret : -EBUSY;
7298 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7300 cc->nr_migratepages -= nr_reclaimed;
7302 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7303 NULL, 0, cc->mode, MR_CMA);
7306 putback_movable_pages(&cc->migratepages);
7313 * alloc_contig_range() -- tries to allocate given range of pages
7314 * @start: start PFN to allocate
7315 * @end: one-past-the-last PFN to allocate
7316 * @migratetype: migratetype of the underlaying pageblocks (either
7317 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7318 * in range must have the same migratetype and it must
7319 * be either of the two.
7321 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7322 * aligned, however it's the caller's responsibility to guarantee that
7323 * we are the only thread that changes migrate type of pageblocks the
7326 * The PFN range must belong to a single zone.
7328 * Returns zero on success or negative error code. On success all
7329 * pages which PFN is in [start, end) are allocated for the caller and
7330 * need to be freed with free_contig_range().
7332 int alloc_contig_range(unsigned long start, unsigned long end,
7333 unsigned migratetype)
7335 unsigned long outer_start, outer_end;
7339 struct compact_control cc = {
7340 .nr_migratepages = 0,
7342 .zone = page_zone(pfn_to_page(start)),
7343 .mode = MIGRATE_SYNC,
7344 .ignore_skip_hint = true,
7346 INIT_LIST_HEAD(&cc.migratepages);
7349 * What we do here is we mark all pageblocks in range as
7350 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7351 * have different sizes, and due to the way page allocator
7352 * work, we align the range to biggest of the two pages so
7353 * that page allocator won't try to merge buddies from
7354 * different pageblocks and change MIGRATE_ISOLATE to some
7355 * other migration type.
7357 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7358 * migrate the pages from an unaligned range (ie. pages that
7359 * we are interested in). This will put all the pages in
7360 * range back to page allocator as MIGRATE_ISOLATE.
7362 * When this is done, we take the pages in range from page
7363 * allocator removing them from the buddy system. This way
7364 * page allocator will never consider using them.
7366 * This lets us mark the pageblocks back as
7367 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7368 * aligned range but not in the unaligned, original range are
7369 * put back to page allocator so that buddy can use them.
7372 ret = start_isolate_page_range(pfn_max_align_down(start),
7373 pfn_max_align_up(end), migratetype,
7379 * In case of -EBUSY, we'd like to know which page causes problem.
7380 * So, just fall through. We will check it in test_pages_isolated().
7382 ret = __alloc_contig_migrate_range(&cc, start, end);
7383 if (ret && ret != -EBUSY)
7387 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7388 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7389 * more, all pages in [start, end) are free in page allocator.
7390 * What we are going to do is to allocate all pages from
7391 * [start, end) (that is remove them from page allocator).
7393 * The only problem is that pages at the beginning and at the
7394 * end of interesting range may be not aligned with pages that
7395 * page allocator holds, ie. they can be part of higher order
7396 * pages. Because of this, we reserve the bigger range and
7397 * once this is done free the pages we are not interested in.
7399 * We don't have to hold zone->lock here because the pages are
7400 * isolated thus they won't get removed from buddy.
7403 lru_add_drain_all();
7404 drain_all_pages(cc.zone);
7407 outer_start = start;
7408 while (!PageBuddy(pfn_to_page(outer_start))) {
7409 if (++order >= MAX_ORDER) {
7410 outer_start = start;
7413 outer_start &= ~0UL << order;
7416 if (outer_start != start) {
7417 order = page_order(pfn_to_page(outer_start));
7420 * outer_start page could be small order buddy page and
7421 * it doesn't include start page. Adjust outer_start
7422 * in this case to report failed page properly
7423 * on tracepoint in test_pages_isolated()
7425 if (outer_start + (1UL << order) <= start)
7426 outer_start = start;
7429 /* Make sure the range is really isolated. */
7430 if (test_pages_isolated(outer_start, end, false)) {
7431 pr_info("%s: [%lx, %lx) PFNs busy\n",
7432 __func__, outer_start, end);
7437 /* Grab isolated pages from freelists. */
7438 outer_end = isolate_freepages_range(&cc, outer_start, end);
7444 /* Free head and tail (if any) */
7445 if (start != outer_start)
7446 free_contig_range(outer_start, start - outer_start);
7447 if (end != outer_end)
7448 free_contig_range(end, outer_end - end);
7451 undo_isolate_page_range(pfn_max_align_down(start),
7452 pfn_max_align_up(end), migratetype);
7456 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7458 unsigned int count = 0;
7460 for (; nr_pages--; pfn++) {
7461 struct page *page = pfn_to_page(pfn);
7463 count += page_count(page) != 1;
7466 WARN(count != 0, "%d pages are still in use!\n", count);
7470 #ifdef CONFIG_MEMORY_HOTPLUG
7472 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7473 * page high values need to be recalulated.
7475 void __meminit zone_pcp_update(struct zone *zone)
7478 mutex_lock(&pcp_batch_high_lock);
7479 for_each_possible_cpu(cpu)
7480 pageset_set_high_and_batch(zone,
7481 per_cpu_ptr(zone->pageset, cpu));
7482 mutex_unlock(&pcp_batch_high_lock);
7486 void zone_pcp_reset(struct zone *zone)
7488 unsigned long flags;
7490 struct per_cpu_pageset *pset;
7492 /* avoid races with drain_pages() */
7493 local_irq_save(flags);
7494 if (zone->pageset != &boot_pageset) {
7495 for_each_online_cpu(cpu) {
7496 pset = per_cpu_ptr(zone->pageset, cpu);
7497 drain_zonestat(zone, pset);
7499 free_percpu(zone->pageset);
7500 zone->pageset = &boot_pageset;
7502 local_irq_restore(flags);
7505 #ifdef CONFIG_MEMORY_HOTREMOVE
7507 * All pages in the range must be in a single zone and isolated
7508 * before calling this.
7511 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7515 unsigned int order, i;
7517 unsigned long flags;
7518 /* find the first valid pfn */
7519 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7524 zone = page_zone(pfn_to_page(pfn));
7525 spin_lock_irqsave(&zone->lock, flags);
7527 while (pfn < end_pfn) {
7528 if (!pfn_valid(pfn)) {
7532 page = pfn_to_page(pfn);
7534 * The HWPoisoned page may be not in buddy system, and
7535 * page_count() is not 0.
7537 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7539 SetPageReserved(page);
7543 BUG_ON(page_count(page));
7544 BUG_ON(!PageBuddy(page));
7545 order = page_order(page);
7546 #ifdef CONFIG_DEBUG_VM
7547 pr_info("remove from free list %lx %d %lx\n",
7548 pfn, 1 << order, end_pfn);
7550 list_del(&page->lru);
7551 rmv_page_order(page);
7552 zone->free_area[order].nr_free--;
7553 for (i = 0; i < (1 << order); i++)
7554 SetPageReserved((page+i));
7555 pfn += (1 << order);
7557 spin_unlock_irqrestore(&zone->lock, flags);
7561 bool is_free_buddy_page(struct page *page)
7563 struct zone *zone = page_zone(page);
7564 unsigned long pfn = page_to_pfn(page);
7565 unsigned long flags;
7568 spin_lock_irqsave(&zone->lock, flags);
7569 for (order = 0; order < MAX_ORDER; order++) {
7570 struct page *page_head = page - (pfn & ((1 << order) - 1));
7572 if (PageBuddy(page_head) && page_order(page_head) >= order)
7575 spin_unlock_irqrestore(&zone->lock, flags);
7577 return order < MAX_ORDER;