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>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node);
79 EXPORT_PER_CPU_SYMBOL(numa_node);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91 int _node_numa_mem_[MAX_NUMNODES];
95 * Array of node states.
97 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
98 [N_POSSIBLE] = NODE_MASK_ALL,
99 [N_ONLINE] = { { [0] = 1UL } },
101 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_HIGHMEM
103 [N_HIGH_MEMORY] = { { [0] = 1UL } },
105 #ifdef CONFIG_MOVABLE_NODE
106 [N_MEMORY] = { { [0] = 1UL } },
108 [N_CPU] = { { [0] = 1UL } },
111 EXPORT_SYMBOL(node_states);
113 /* Protect totalram_pages and zone->managed_pages */
114 static DEFINE_SPINLOCK(managed_page_count_lock);
116 unsigned long totalram_pages __read_mostly;
117 unsigned long totalreserve_pages __read_mostly;
118 unsigned long totalcma_pages __read_mostly;
120 int percpu_pagelist_fraction;
121 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 * A cached value of the page's pageblock's migratetype, used when the page is
125 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
126 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
127 * Also the migratetype set in the page does not necessarily match the pcplist
128 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
129 * other index - this ensures that it will be put on the correct CMA freelist.
131 static inline int get_pcppage_migratetype(struct page *page)
136 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
138 page->index = migratetype;
141 #ifdef CONFIG_PM_SLEEP
143 * The following functions are used by the suspend/hibernate code to temporarily
144 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
145 * while devices are suspended. To avoid races with the suspend/hibernate code,
146 * they should always be called with pm_mutex held (gfp_allowed_mask also should
147 * only be modified with pm_mutex held, unless the suspend/hibernate code is
148 * guaranteed not to run in parallel with that modification).
151 static gfp_t saved_gfp_mask;
153 void pm_restore_gfp_mask(void)
155 WARN_ON(!mutex_is_locked(&pm_mutex));
156 if (saved_gfp_mask) {
157 gfp_allowed_mask = saved_gfp_mask;
162 void pm_restrict_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex));
165 WARN_ON(saved_gfp_mask);
166 saved_gfp_mask = gfp_allowed_mask;
167 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
170 bool pm_suspended_storage(void)
172 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
176 #endif /* CONFIG_PM_SLEEP */
178 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
179 unsigned int pageblock_order __read_mostly;
182 static void __free_pages_ok(struct page *page, unsigned int order);
185 * results with 256, 32 in the lowmem_reserve sysctl:
186 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
187 * 1G machine -> (16M dma, 784M normal, 224M high)
188 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
189 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
190 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
192 * TBD: should special case ZONE_DMA32 machines here - in those we normally
193 * don't need any ZONE_NORMAL reservation
195 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
202 #ifdef CONFIG_HIGHMEM
208 EXPORT_SYMBOL(totalram_pages);
210 static char * const zone_names[MAX_NR_ZONES] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
222 #ifdef CONFIG_ZONE_DEVICE
227 char * const migratetype_names[MIGRATE_TYPES] = {
235 #ifdef CONFIG_MEMORY_ISOLATION
240 compound_page_dtor * const compound_page_dtors[] = {
243 #ifdef CONFIG_HUGETLB_PAGE
246 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
251 int min_free_kbytes = 1024;
252 int user_min_free_kbytes = -1;
253 int watermark_scale_factor = 10;
255 static unsigned long __meminitdata nr_kernel_pages;
256 static unsigned long __meminitdata nr_all_pages;
257 static unsigned long __meminitdata dma_reserve;
259 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
260 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
261 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
262 static unsigned long __initdata required_kernelcore;
263 static unsigned long __initdata required_movablecore;
264 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
265 static bool mirrored_kernelcore;
267 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
269 EXPORT_SYMBOL(movable_zone);
270 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
273 int nr_node_ids __read_mostly = MAX_NUMNODES;
274 int nr_online_nodes __read_mostly = 1;
275 EXPORT_SYMBOL(nr_node_ids);
276 EXPORT_SYMBOL(nr_online_nodes);
279 int page_group_by_mobility_disabled __read_mostly;
281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
282 static inline void reset_deferred_meminit(pg_data_t *pgdat)
284 pgdat->first_deferred_pfn = ULONG_MAX;
287 /* Returns true if the struct page for the pfn is uninitialised */
288 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
290 int nid = early_pfn_to_nid(pfn);
292 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
298 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
300 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
307 * Returns false when the remaining initialisation should be deferred until
308 * later in the boot cycle when it can be parallelised.
310 static inline bool update_defer_init(pg_data_t *pgdat,
311 unsigned long pfn, unsigned long zone_end,
312 unsigned long *nr_initialised)
314 unsigned long max_initialise;
316 /* Always populate low zones for address-contrained allocations */
317 if (zone_end < pgdat_end_pfn(pgdat))
320 * Initialise at least 2G of a node but also take into account that
321 * two large system hashes that can take up 1GB for 0.25TB/node.
323 max_initialise = max(2UL << (30 - PAGE_SHIFT),
324 (pgdat->node_spanned_pages >> 8));
327 if ((*nr_initialised > max_initialise) &&
328 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
329 pgdat->first_deferred_pfn = pfn;
336 static inline void reset_deferred_meminit(pg_data_t *pgdat)
340 static inline bool early_page_uninitialised(unsigned long pfn)
345 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
350 static inline bool update_defer_init(pg_data_t *pgdat,
351 unsigned long pfn, unsigned long zone_end,
352 unsigned long *nr_initialised)
358 /* Return a pointer to the bitmap storing bits affecting a block of pages */
359 static inline unsigned long *get_pageblock_bitmap(struct page *page,
362 #ifdef CONFIG_SPARSEMEM
363 return __pfn_to_section(pfn)->pageblock_flags;
365 return page_zone(page)->pageblock_flags;
366 #endif /* CONFIG_SPARSEMEM */
369 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
371 #ifdef CONFIG_SPARSEMEM
372 pfn &= (PAGES_PER_SECTION-1);
373 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
375 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
376 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
377 #endif /* CONFIG_SPARSEMEM */
381 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
382 * @page: The page within the block of interest
383 * @pfn: The target page frame number
384 * @end_bitidx: The last bit of interest to retrieve
385 * @mask: mask of bits that the caller is interested in
387 * Return: pageblock_bits flags
389 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
391 unsigned long end_bitidx,
394 unsigned long *bitmap;
395 unsigned long bitidx, word_bitidx;
398 bitmap = get_pageblock_bitmap(page, pfn);
399 bitidx = pfn_to_bitidx(page, pfn);
400 word_bitidx = bitidx / BITS_PER_LONG;
401 bitidx &= (BITS_PER_LONG-1);
403 word = bitmap[word_bitidx];
404 bitidx += end_bitidx;
405 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
408 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
409 unsigned long end_bitidx,
412 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
415 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
417 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
421 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
422 * @page: The page within the block of interest
423 * @flags: The flags to set
424 * @pfn: The target page frame number
425 * @end_bitidx: The last bit of interest
426 * @mask: mask of bits that the caller is interested in
428 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
430 unsigned long end_bitidx,
433 unsigned long *bitmap;
434 unsigned long bitidx, word_bitidx;
435 unsigned long old_word, word;
437 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
439 bitmap = get_pageblock_bitmap(page, pfn);
440 bitidx = pfn_to_bitidx(page, pfn);
441 word_bitidx = bitidx / BITS_PER_LONG;
442 bitidx &= (BITS_PER_LONG-1);
444 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
446 bitidx += end_bitidx;
447 mask <<= (BITS_PER_LONG - bitidx - 1);
448 flags <<= (BITS_PER_LONG - bitidx - 1);
450 word = READ_ONCE(bitmap[word_bitidx]);
452 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
453 if (word == old_word)
459 void set_pageblock_migratetype(struct page *page, int migratetype)
461 if (unlikely(page_group_by_mobility_disabled &&
462 migratetype < MIGRATE_PCPTYPES))
463 migratetype = MIGRATE_UNMOVABLE;
465 set_pageblock_flags_group(page, (unsigned long)migratetype,
466 PB_migrate, PB_migrate_end);
469 #ifdef CONFIG_DEBUG_VM
470 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
474 unsigned long pfn = page_to_pfn(page);
475 unsigned long sp, start_pfn;
478 seq = zone_span_seqbegin(zone);
479 start_pfn = zone->zone_start_pfn;
480 sp = zone->spanned_pages;
481 if (!zone_spans_pfn(zone, pfn))
483 } while (zone_span_seqretry(zone, seq));
486 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
487 pfn, zone_to_nid(zone), zone->name,
488 start_pfn, start_pfn + sp);
493 static int page_is_consistent(struct zone *zone, struct page *page)
495 if (!pfn_valid_within(page_to_pfn(page)))
497 if (zone != page_zone(page))
503 * Temporary debugging check for pages not lying within a given zone.
505 static int bad_range(struct zone *zone, struct page *page)
507 if (page_outside_zone_boundaries(zone, page))
509 if (!page_is_consistent(zone, page))
515 static inline int bad_range(struct zone *zone, struct page *page)
521 static void bad_page(struct page *page, const char *reason,
522 unsigned long bad_flags)
524 static unsigned long resume;
525 static unsigned long nr_shown;
526 static unsigned long nr_unshown;
529 * Allow a burst of 60 reports, then keep quiet for that minute;
530 * or allow a steady drip of one report per second.
532 if (nr_shown == 60) {
533 if (time_before(jiffies, resume)) {
539 "BUG: Bad page state: %lu messages suppressed\n",
546 resume = jiffies + 60 * HZ;
548 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
549 current->comm, page_to_pfn(page));
550 __dump_page(page, reason);
551 bad_flags &= page->flags;
553 pr_alert("bad because of flags: %#lx(%pGp)\n",
554 bad_flags, &bad_flags);
555 dump_page_owner(page);
560 /* Leave bad fields for debug, except PageBuddy could make trouble */
561 page_mapcount_reset(page); /* remove PageBuddy */
562 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
566 * Higher-order pages are called "compound pages". They are structured thusly:
568 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
570 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
571 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
573 * The first tail page's ->compound_dtor holds the offset in array of compound
574 * page destructors. See compound_page_dtors.
576 * The first tail page's ->compound_order holds the order of allocation.
577 * This usage means that zero-order pages may not be compound.
580 void free_compound_page(struct page *page)
582 __free_pages_ok(page, compound_order(page));
585 void prep_compound_page(struct page *page, unsigned int order)
588 int nr_pages = 1 << order;
590 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
591 set_compound_order(page, order);
593 for (i = 1; i < nr_pages; i++) {
594 struct page *p = page + i;
595 set_page_count(p, 0);
596 p->mapping = TAIL_MAPPING;
597 set_compound_head(p, page);
599 atomic_set(compound_mapcount_ptr(page), -1);
602 #ifdef CONFIG_DEBUG_PAGEALLOC
603 unsigned int _debug_guardpage_minorder;
604 bool _debug_pagealloc_enabled __read_mostly
605 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
606 EXPORT_SYMBOL(_debug_pagealloc_enabled);
607 bool _debug_guardpage_enabled __read_mostly;
609 static int __init early_debug_pagealloc(char *buf)
613 return kstrtobool(buf, &_debug_pagealloc_enabled);
615 early_param("debug_pagealloc", early_debug_pagealloc);
617 static bool need_debug_guardpage(void)
619 /* If we don't use debug_pagealloc, we don't need guard page */
620 if (!debug_pagealloc_enabled())
626 static void init_debug_guardpage(void)
628 if (!debug_pagealloc_enabled())
631 _debug_guardpage_enabled = true;
634 struct page_ext_operations debug_guardpage_ops = {
635 .need = need_debug_guardpage,
636 .init = init_debug_guardpage,
639 static int __init debug_guardpage_minorder_setup(char *buf)
643 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
644 pr_err("Bad debug_guardpage_minorder value\n");
647 _debug_guardpage_minorder = res;
648 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
651 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
653 static inline void set_page_guard(struct zone *zone, struct page *page,
654 unsigned int order, int migratetype)
656 struct page_ext *page_ext;
658 if (!debug_guardpage_enabled())
661 page_ext = lookup_page_ext(page);
662 if (unlikely(!page_ext))
665 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
667 INIT_LIST_HEAD(&page->lru);
668 set_page_private(page, order);
669 /* Guard pages are not available for any usage */
670 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
673 static inline void clear_page_guard(struct zone *zone, struct page *page,
674 unsigned int order, int migratetype)
676 struct page_ext *page_ext;
678 if (!debug_guardpage_enabled())
681 page_ext = lookup_page_ext(page);
682 if (unlikely(!page_ext))
685 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
687 set_page_private(page, 0);
688 if (!is_migrate_isolate(migratetype))
689 __mod_zone_freepage_state(zone, (1 << order), migratetype);
692 struct page_ext_operations debug_guardpage_ops = { NULL, };
693 static inline void set_page_guard(struct zone *zone, struct page *page,
694 unsigned int order, int migratetype) {}
695 static inline void clear_page_guard(struct zone *zone, struct page *page,
696 unsigned int order, int migratetype) {}
699 static inline void set_page_order(struct page *page, unsigned int order)
701 set_page_private(page, order);
702 __SetPageBuddy(page);
705 static inline void rmv_page_order(struct page *page)
707 __ClearPageBuddy(page);
708 set_page_private(page, 0);
712 * This function checks whether a page is free && is the buddy
713 * we can do coalesce a page and its buddy if
714 * (a) the buddy is not in a hole &&
715 * (b) the buddy is in the buddy system &&
716 * (c) a page and its buddy have the same order &&
717 * (d) a page and its buddy are in the same zone.
719 * For recording whether a page is in the buddy system, we set ->_mapcount
720 * PAGE_BUDDY_MAPCOUNT_VALUE.
721 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
722 * serialized by zone->lock.
724 * For recording page's order, we use page_private(page).
726 static inline int page_is_buddy(struct page *page, struct page *buddy,
729 if (!pfn_valid_within(page_to_pfn(buddy)))
732 if (page_is_guard(buddy) && page_order(buddy) == order) {
733 if (page_zone_id(page) != page_zone_id(buddy))
736 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
741 if (PageBuddy(buddy) && page_order(buddy) == order) {
743 * zone check is done late to avoid uselessly
744 * calculating zone/node ids for pages that could
747 if (page_zone_id(page) != page_zone_id(buddy))
750 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
758 * Freeing function for a buddy system allocator.
760 * The concept of a buddy system is to maintain direct-mapped table
761 * (containing bit values) for memory blocks of various "orders".
762 * The bottom level table contains the map for the smallest allocatable
763 * units of memory (here, pages), and each level above it describes
764 * pairs of units from the levels below, hence, "buddies".
765 * At a high level, all that happens here is marking the table entry
766 * at the bottom level available, and propagating the changes upward
767 * as necessary, plus some accounting needed to play nicely with other
768 * parts of the VM system.
769 * At each level, we keep a list of pages, which are heads of continuous
770 * free pages of length of (1 << order) and marked with _mapcount
771 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
773 * So when we are allocating or freeing one, we can derive the state of the
774 * other. That is, if we allocate a small block, and both were
775 * free, the remainder of the region must be split into blocks.
776 * If a block is freed, and its buddy is also free, then this
777 * triggers coalescing into a block of larger size.
782 static inline void __free_one_page(struct page *page,
784 struct zone *zone, unsigned int order,
787 unsigned long page_idx;
788 unsigned long combined_idx;
789 unsigned long uninitialized_var(buddy_idx);
791 unsigned int max_order;
793 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
795 VM_BUG_ON(!zone_is_initialized(zone));
796 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
798 VM_BUG_ON(migratetype == -1);
799 if (likely(!is_migrate_isolate(migratetype)))
800 __mod_zone_freepage_state(zone, 1 << order, migratetype);
802 page_idx = pfn & ((1 << MAX_ORDER) - 1);
804 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
805 VM_BUG_ON_PAGE(bad_range(zone, page), page);
808 while (order < max_order - 1) {
809 buddy_idx = __find_buddy_index(page_idx, order);
810 buddy = page + (buddy_idx - page_idx);
811 if (!page_is_buddy(page, buddy, order))
814 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
815 * merge with it and move up one order.
817 if (page_is_guard(buddy)) {
818 clear_page_guard(zone, buddy, order, migratetype);
820 list_del(&buddy->lru);
821 zone->free_area[order].nr_free--;
822 rmv_page_order(buddy);
824 combined_idx = buddy_idx & page_idx;
825 page = page + (combined_idx - page_idx);
826 page_idx = combined_idx;
829 if (max_order < MAX_ORDER) {
830 /* If we are here, it means order is >= pageblock_order.
831 * We want to prevent merge between freepages on isolate
832 * pageblock and normal pageblock. Without this, pageblock
833 * isolation could cause incorrect freepage or CMA accounting.
835 * We don't want to hit this code for the more frequent
838 if (unlikely(has_isolate_pageblock(zone))) {
841 buddy_idx = __find_buddy_index(page_idx, order);
842 buddy = page + (buddy_idx - page_idx);
843 buddy_mt = get_pageblock_migratetype(buddy);
845 if (migratetype != buddy_mt
846 && (is_migrate_isolate(migratetype) ||
847 is_migrate_isolate(buddy_mt)))
851 goto continue_merging;
855 set_page_order(page, order);
858 * If this is not the largest possible page, check if the buddy
859 * of the next-highest order is free. If it is, it's possible
860 * that pages are being freed that will coalesce soon. In case,
861 * that is happening, add the free page to the tail of the list
862 * so it's less likely to be used soon and more likely to be merged
863 * as a higher order page
865 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
866 struct page *higher_page, *higher_buddy;
867 combined_idx = buddy_idx & page_idx;
868 higher_page = page + (combined_idx - page_idx);
869 buddy_idx = __find_buddy_index(combined_idx, order + 1);
870 higher_buddy = higher_page + (buddy_idx - combined_idx);
871 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
872 list_add_tail(&page->lru,
873 &zone->free_area[order].free_list[migratetype]);
878 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
880 zone->free_area[order].nr_free++;
884 * A bad page could be due to a number of fields. Instead of multiple branches,
885 * try and check multiple fields with one check. The caller must do a detailed
886 * check if necessary.
888 static inline bool page_expected_state(struct page *page,
889 unsigned long check_flags)
891 if (unlikely(atomic_read(&page->_mapcount) != -1))
894 if (unlikely((unsigned long)page->mapping |
895 page_ref_count(page) |
897 (unsigned long)page->mem_cgroup |
899 (page->flags & check_flags)))
905 static void free_pages_check_bad(struct page *page)
907 const char *bad_reason;
908 unsigned long bad_flags;
913 if (unlikely(atomic_read(&page->_mapcount) != -1))
914 bad_reason = "nonzero mapcount";
915 if (unlikely(page->mapping != NULL))
916 bad_reason = "non-NULL mapping";
917 if (unlikely(page_ref_count(page) != 0))
918 bad_reason = "nonzero _refcount";
919 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
920 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
921 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
924 if (unlikely(page->mem_cgroup))
925 bad_reason = "page still charged to cgroup";
927 bad_page(page, bad_reason, bad_flags);
930 static inline int free_pages_check(struct page *page)
932 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
935 /* Something has gone sideways, find it */
936 free_pages_check_bad(page);
940 static int free_tail_pages_check(struct page *head_page, struct page *page)
945 * We rely page->lru.next never has bit 0 set, unless the page
946 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
948 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
950 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
954 switch (page - head_page) {
956 /* the first tail page: ->mapping is compound_mapcount() */
957 if (unlikely(compound_mapcount(page))) {
958 bad_page(page, "nonzero compound_mapcount", 0);
964 * the second tail page: ->mapping is
965 * page_deferred_list().next -- ignore value.
969 if (page->mapping != TAIL_MAPPING) {
970 bad_page(page, "corrupted mapping in tail page", 0);
975 if (unlikely(!PageTail(page))) {
976 bad_page(page, "PageTail not set", 0);
979 if (unlikely(compound_head(page) != head_page)) {
980 bad_page(page, "compound_head not consistent", 0);
985 page->mapping = NULL;
986 clear_compound_head(page);
990 static __always_inline bool free_pages_prepare(struct page *page,
991 unsigned int order, bool check_free)
995 VM_BUG_ON_PAGE(PageTail(page), page);
997 trace_mm_page_free(page, order);
998 kmemcheck_free_shadow(page, order);
1001 * Check tail pages before head page information is cleared to
1002 * avoid checking PageCompound for order-0 pages.
1004 if (unlikely(order)) {
1005 bool compound = PageCompound(page);
1008 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1011 ClearPageDoubleMap(page);
1012 for (i = 1; i < (1 << order); i++) {
1014 bad += free_tail_pages_check(page, page + i);
1015 if (unlikely(free_pages_check(page + i))) {
1019 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1022 if (PageMappingFlags(page))
1023 page->mapping = NULL;
1024 if (memcg_kmem_enabled() && PageKmemcg(page)) {
1025 memcg_kmem_uncharge(page, order);
1026 __ClearPageKmemcg(page);
1029 bad += free_pages_check(page);
1033 page_cpupid_reset_last(page);
1034 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1035 reset_page_owner(page, order);
1037 if (!PageHighMem(page)) {
1038 debug_check_no_locks_freed(page_address(page),
1039 PAGE_SIZE << order);
1040 debug_check_no_obj_freed(page_address(page),
1041 PAGE_SIZE << order);
1043 arch_free_page(page, order);
1044 kernel_poison_pages(page, 1 << order, 0);
1045 kernel_map_pages(page, 1 << order, 0);
1046 kasan_free_pages(page, order);
1051 #ifdef CONFIG_DEBUG_VM
1052 static inline bool free_pcp_prepare(struct page *page)
1054 return free_pages_prepare(page, 0, true);
1057 static inline bool bulkfree_pcp_prepare(struct page *page)
1062 static bool free_pcp_prepare(struct page *page)
1064 return free_pages_prepare(page, 0, false);
1067 static bool bulkfree_pcp_prepare(struct page *page)
1069 return free_pages_check(page);
1071 #endif /* CONFIG_DEBUG_VM */
1074 * Frees a number of pages from the PCP lists
1075 * Assumes all pages on list are in same zone, and of same order.
1076 * count is the number of pages to free.
1078 * If the zone was previously in an "all pages pinned" state then look to
1079 * see if this freeing clears that state.
1081 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1082 * pinned" detection logic.
1084 static void free_pcppages_bulk(struct zone *zone, int count,
1085 struct per_cpu_pages *pcp)
1087 int migratetype = 0;
1089 unsigned long nr_scanned;
1090 bool isolated_pageblocks;
1092 spin_lock(&zone->lock);
1093 isolated_pageblocks = has_isolate_pageblock(zone);
1094 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1096 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1100 struct list_head *list;
1103 * Remove pages from lists in a round-robin fashion. A
1104 * batch_free count is maintained that is incremented when an
1105 * empty list is encountered. This is so more pages are freed
1106 * off fuller lists instead of spinning excessively around empty
1111 if (++migratetype == MIGRATE_PCPTYPES)
1113 list = &pcp->lists[migratetype];
1114 } while (list_empty(list));
1116 /* This is the only non-empty list. Free them all. */
1117 if (batch_free == MIGRATE_PCPTYPES)
1121 int mt; /* migratetype of the to-be-freed page */
1123 page = list_last_entry(list, struct page, lru);
1124 /* must delete as __free_one_page list manipulates */
1125 list_del(&page->lru);
1127 mt = get_pcppage_migratetype(page);
1128 /* MIGRATE_ISOLATE page should not go to pcplists */
1129 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1130 /* Pageblock could have been isolated meanwhile */
1131 if (unlikely(isolated_pageblocks))
1132 mt = get_pageblock_migratetype(page);
1134 if (bulkfree_pcp_prepare(page))
1137 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1138 trace_mm_page_pcpu_drain(page, 0, mt);
1139 } while (--count && --batch_free && !list_empty(list));
1141 spin_unlock(&zone->lock);
1144 static void free_one_page(struct zone *zone,
1145 struct page *page, unsigned long pfn,
1149 unsigned long nr_scanned;
1150 spin_lock(&zone->lock);
1151 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1153 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1155 if (unlikely(has_isolate_pageblock(zone) ||
1156 is_migrate_isolate(migratetype))) {
1157 migratetype = get_pfnblock_migratetype(page, pfn);
1159 __free_one_page(page, pfn, zone, order, migratetype);
1160 spin_unlock(&zone->lock);
1163 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1164 unsigned long zone, int nid)
1166 set_page_links(page, zone, nid, pfn);
1167 init_page_count(page);
1168 page_mapcount_reset(page);
1169 page_cpupid_reset_last(page);
1171 INIT_LIST_HEAD(&page->lru);
1172 #ifdef WANT_PAGE_VIRTUAL
1173 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1174 if (!is_highmem_idx(zone))
1175 set_page_address(page, __va(pfn << PAGE_SHIFT));
1179 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1182 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1185 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1186 static void init_reserved_page(unsigned long pfn)
1191 if (!early_page_uninitialised(pfn))
1194 nid = early_pfn_to_nid(pfn);
1195 pgdat = NODE_DATA(nid);
1197 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1198 struct zone *zone = &pgdat->node_zones[zid];
1200 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1203 __init_single_pfn(pfn, zid, nid);
1206 static inline void init_reserved_page(unsigned long pfn)
1209 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1212 * Initialised pages do not have PageReserved set. This function is
1213 * called for each range allocated by the bootmem allocator and
1214 * marks the pages PageReserved. The remaining valid pages are later
1215 * sent to the buddy page allocator.
1217 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1219 unsigned long start_pfn = PFN_DOWN(start);
1220 unsigned long end_pfn = PFN_UP(end);
1222 for (; start_pfn < end_pfn; start_pfn++) {
1223 if (pfn_valid(start_pfn)) {
1224 struct page *page = pfn_to_page(start_pfn);
1226 init_reserved_page(start_pfn);
1228 /* Avoid false-positive PageTail() */
1229 INIT_LIST_HEAD(&page->lru);
1231 SetPageReserved(page);
1236 static void __free_pages_ok(struct page *page, unsigned int order)
1238 unsigned long flags;
1240 unsigned long pfn = page_to_pfn(page);
1242 if (!free_pages_prepare(page, order, true))
1245 migratetype = get_pfnblock_migratetype(page, pfn);
1246 local_irq_save(flags);
1247 __count_vm_events(PGFREE, 1 << order);
1248 free_one_page(page_zone(page), page, pfn, order, migratetype);
1249 local_irq_restore(flags);
1252 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1254 unsigned int nr_pages = 1 << order;
1255 struct page *p = page;
1259 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1261 __ClearPageReserved(p);
1262 set_page_count(p, 0);
1264 __ClearPageReserved(p);
1265 set_page_count(p, 0);
1267 page_zone(page)->managed_pages += nr_pages;
1268 set_page_refcounted(page);
1269 __free_pages(page, order);
1272 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1273 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1275 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1277 int __meminit early_pfn_to_nid(unsigned long pfn)
1279 static DEFINE_SPINLOCK(early_pfn_lock);
1282 spin_lock(&early_pfn_lock);
1283 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1285 nid = first_online_node;
1286 spin_unlock(&early_pfn_lock);
1292 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1293 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1294 struct mminit_pfnnid_cache *state)
1298 nid = __early_pfn_to_nid(pfn, state);
1299 if (nid >= 0 && nid != node)
1304 /* Only safe to use early in boot when initialisation is single-threaded */
1305 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1307 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1312 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1316 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1317 struct mminit_pfnnid_cache *state)
1324 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1327 if (early_page_uninitialised(pfn))
1329 return __free_pages_boot_core(page, order);
1333 * Check that the whole (or subset of) a pageblock given by the interval of
1334 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1335 * with the migration of free compaction scanner. The scanners then need to
1336 * use only pfn_valid_within() check for arches that allow holes within
1339 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1341 * It's possible on some configurations to have a setup like node0 node1 node0
1342 * i.e. it's possible that all pages within a zones range of pages do not
1343 * belong to a single zone. We assume that a border between node0 and node1
1344 * can occur within a single pageblock, but not a node0 node1 node0
1345 * interleaving within a single pageblock. It is therefore sufficient to check
1346 * the first and last page of a pageblock and avoid checking each individual
1347 * page in a pageblock.
1349 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1350 unsigned long end_pfn, struct zone *zone)
1352 struct page *start_page;
1353 struct page *end_page;
1355 /* end_pfn is one past the range we are checking */
1358 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1361 start_page = pfn_to_page(start_pfn);
1363 if (page_zone(start_page) != zone)
1366 end_page = pfn_to_page(end_pfn);
1368 /* This gives a shorter code than deriving page_zone(end_page) */
1369 if (page_zone_id(start_page) != page_zone_id(end_page))
1375 void set_zone_contiguous(struct zone *zone)
1377 unsigned long block_start_pfn = zone->zone_start_pfn;
1378 unsigned long block_end_pfn;
1380 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1381 for (; block_start_pfn < zone_end_pfn(zone);
1382 block_start_pfn = block_end_pfn,
1383 block_end_pfn += pageblock_nr_pages) {
1385 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1387 if (!__pageblock_pfn_to_page(block_start_pfn,
1388 block_end_pfn, zone))
1392 /* We confirm that there is no hole */
1393 zone->contiguous = true;
1396 void clear_zone_contiguous(struct zone *zone)
1398 zone->contiguous = false;
1401 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1402 static void __init deferred_free_range(struct page *page,
1403 unsigned long pfn, int nr_pages)
1410 /* Free a large naturally-aligned chunk if possible */
1411 if (nr_pages == MAX_ORDER_NR_PAGES &&
1412 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1413 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1414 __free_pages_boot_core(page, MAX_ORDER-1);
1418 for (i = 0; i < nr_pages; i++, page++)
1419 __free_pages_boot_core(page, 0);
1422 /* Completion tracking for deferred_init_memmap() threads */
1423 static atomic_t pgdat_init_n_undone __initdata;
1424 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1426 static inline void __init pgdat_init_report_one_done(void)
1428 if (atomic_dec_and_test(&pgdat_init_n_undone))
1429 complete(&pgdat_init_all_done_comp);
1432 /* Initialise remaining memory on a node */
1433 static int __init deferred_init_memmap(void *data)
1435 pg_data_t *pgdat = data;
1436 int nid = pgdat->node_id;
1437 struct mminit_pfnnid_cache nid_init_state = { };
1438 unsigned long start = jiffies;
1439 unsigned long nr_pages = 0;
1440 unsigned long walk_start, walk_end;
1443 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1444 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1446 if (first_init_pfn == ULONG_MAX) {
1447 pgdat_init_report_one_done();
1451 /* Bind memory initialisation thread to a local node if possible */
1452 if (!cpumask_empty(cpumask))
1453 set_cpus_allowed_ptr(current, cpumask);
1455 /* Sanity check boundaries */
1456 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1457 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1458 pgdat->first_deferred_pfn = ULONG_MAX;
1460 /* Only the highest zone is deferred so find it */
1461 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1462 zone = pgdat->node_zones + zid;
1463 if (first_init_pfn < zone_end_pfn(zone))
1467 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1468 unsigned long pfn, end_pfn;
1469 struct page *page = NULL;
1470 struct page *free_base_page = NULL;
1471 unsigned long free_base_pfn = 0;
1474 end_pfn = min(walk_end, zone_end_pfn(zone));
1475 pfn = first_init_pfn;
1476 if (pfn < walk_start)
1478 if (pfn < zone->zone_start_pfn)
1479 pfn = zone->zone_start_pfn;
1481 for (; pfn < end_pfn; pfn++) {
1482 if (!pfn_valid_within(pfn))
1486 * Ensure pfn_valid is checked every
1487 * MAX_ORDER_NR_PAGES for memory holes
1489 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1490 if (!pfn_valid(pfn)) {
1496 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1501 /* Minimise pfn page lookups and scheduler checks */
1502 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1505 nr_pages += nr_to_free;
1506 deferred_free_range(free_base_page,
1507 free_base_pfn, nr_to_free);
1508 free_base_page = NULL;
1509 free_base_pfn = nr_to_free = 0;
1511 page = pfn_to_page(pfn);
1516 VM_BUG_ON(page_zone(page) != zone);
1520 __init_single_page(page, pfn, zid, nid);
1521 if (!free_base_page) {
1522 free_base_page = page;
1523 free_base_pfn = pfn;
1528 /* Where possible, batch up pages for a single free */
1531 /* Free the current block of pages to allocator */
1532 nr_pages += nr_to_free;
1533 deferred_free_range(free_base_page, free_base_pfn,
1535 free_base_page = NULL;
1536 free_base_pfn = nr_to_free = 0;
1539 first_init_pfn = max(end_pfn, first_init_pfn);
1542 /* Sanity check that the next zone really is unpopulated */
1543 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1545 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1546 jiffies_to_msecs(jiffies - start));
1548 pgdat_init_report_one_done();
1551 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1553 void __init page_alloc_init_late(void)
1557 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1560 /* There will be num_node_state(N_MEMORY) threads */
1561 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1562 for_each_node_state(nid, N_MEMORY) {
1563 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1566 /* Block until all are initialised */
1567 wait_for_completion(&pgdat_init_all_done_comp);
1569 /* Reinit limits that are based on free pages after the kernel is up */
1570 files_maxfiles_init();
1573 for_each_populated_zone(zone)
1574 set_zone_contiguous(zone);
1578 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1579 void __init init_cma_reserved_pageblock(struct page *page)
1581 unsigned i = pageblock_nr_pages;
1582 struct page *p = page;
1585 __ClearPageReserved(p);
1586 set_page_count(p, 0);
1589 set_pageblock_migratetype(page, MIGRATE_CMA);
1591 if (pageblock_order >= MAX_ORDER) {
1592 i = pageblock_nr_pages;
1595 set_page_refcounted(p);
1596 __free_pages(p, MAX_ORDER - 1);
1597 p += MAX_ORDER_NR_PAGES;
1598 } while (i -= MAX_ORDER_NR_PAGES);
1600 set_page_refcounted(page);
1601 __free_pages(page, pageblock_order);
1604 adjust_managed_page_count(page, pageblock_nr_pages);
1609 * The order of subdivision here is critical for the IO subsystem.
1610 * Please do not alter this order without good reasons and regression
1611 * testing. Specifically, as large blocks of memory are subdivided,
1612 * the order in which smaller blocks are delivered depends on the order
1613 * they're subdivided in this function. This is the primary factor
1614 * influencing the order in which pages are delivered to the IO
1615 * subsystem according to empirical testing, and this is also justified
1616 * by considering the behavior of a buddy system containing a single
1617 * large block of memory acted on by a series of small allocations.
1618 * This behavior is a critical factor in sglist merging's success.
1622 static inline void expand(struct zone *zone, struct page *page,
1623 int low, int high, struct free_area *area,
1626 unsigned long size = 1 << high;
1628 while (high > low) {
1632 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1634 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1635 debug_guardpage_enabled() &&
1636 high < debug_guardpage_minorder()) {
1638 * Mark as guard pages (or page), that will allow to
1639 * merge back to allocator when buddy will be freed.
1640 * Corresponding page table entries will not be touched,
1641 * pages will stay not present in virtual address space
1643 set_page_guard(zone, &page[size], high, migratetype);
1646 list_add(&page[size].lru, &area->free_list[migratetype]);
1648 set_page_order(&page[size], high);
1652 static void check_new_page_bad(struct page *page)
1654 const char *bad_reason = NULL;
1655 unsigned long bad_flags = 0;
1657 if (unlikely(atomic_read(&page->_mapcount) != -1))
1658 bad_reason = "nonzero mapcount";
1659 if (unlikely(page->mapping != NULL))
1660 bad_reason = "non-NULL mapping";
1661 if (unlikely(page_ref_count(page) != 0))
1662 bad_reason = "nonzero _count";
1663 if (unlikely(page->flags & __PG_HWPOISON)) {
1664 bad_reason = "HWPoisoned (hardware-corrupted)";
1665 bad_flags = __PG_HWPOISON;
1666 /* Don't complain about hwpoisoned pages */
1667 page_mapcount_reset(page); /* remove PageBuddy */
1670 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1671 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1672 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1675 if (unlikely(page->mem_cgroup))
1676 bad_reason = "page still charged to cgroup";
1678 bad_page(page, bad_reason, bad_flags);
1682 * This page is about to be returned from the page allocator
1684 static inline int check_new_page(struct page *page)
1686 if (likely(page_expected_state(page,
1687 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1690 check_new_page_bad(page);
1694 static inline bool free_pages_prezeroed(bool poisoned)
1696 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1697 page_poisoning_enabled() && poisoned;
1700 #ifdef CONFIG_DEBUG_VM
1701 static bool check_pcp_refill(struct page *page)
1706 static bool check_new_pcp(struct page *page)
1708 return check_new_page(page);
1711 static bool check_pcp_refill(struct page *page)
1713 return check_new_page(page);
1715 static bool check_new_pcp(struct page *page)
1719 #endif /* CONFIG_DEBUG_VM */
1721 static bool check_new_pages(struct page *page, unsigned int order)
1724 for (i = 0; i < (1 << order); i++) {
1725 struct page *p = page + i;
1727 if (unlikely(check_new_page(p)))
1734 inline void post_alloc_hook(struct page *page, unsigned int order,
1737 set_page_private(page, 0);
1738 set_page_refcounted(page);
1740 arch_alloc_page(page, order);
1741 kernel_map_pages(page, 1 << order, 1);
1742 kernel_poison_pages(page, 1 << order, 1);
1743 kasan_alloc_pages(page, order);
1744 set_page_owner(page, order, gfp_flags);
1747 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1748 unsigned int alloc_flags)
1751 bool poisoned = true;
1753 for (i = 0; i < (1 << order); i++) {
1754 struct page *p = page + i;
1756 poisoned &= page_is_poisoned(p);
1759 post_alloc_hook(page, order, gfp_flags);
1761 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1762 for (i = 0; i < (1 << order); i++)
1763 clear_highpage(page + i);
1765 if (order && (gfp_flags & __GFP_COMP))
1766 prep_compound_page(page, order);
1769 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1770 * allocate the page. The expectation is that the caller is taking
1771 * steps that will free more memory. The caller should avoid the page
1772 * being used for !PFMEMALLOC purposes.
1774 if (alloc_flags & ALLOC_NO_WATERMARKS)
1775 set_page_pfmemalloc(page);
1777 clear_page_pfmemalloc(page);
1781 * Go through the free lists for the given migratetype and remove
1782 * the smallest available page from the freelists
1785 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1788 unsigned int current_order;
1789 struct free_area *area;
1792 /* Find a page of the appropriate size in the preferred list */
1793 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1794 area = &(zone->free_area[current_order]);
1795 page = list_first_entry_or_null(&area->free_list[migratetype],
1799 list_del(&page->lru);
1800 rmv_page_order(page);
1802 expand(zone, page, order, current_order, area, migratetype);
1803 set_pcppage_migratetype(page, migratetype);
1812 * This array describes the order lists are fallen back to when
1813 * the free lists for the desirable migrate type are depleted
1815 static int fallbacks[MIGRATE_TYPES][4] = {
1816 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1817 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1818 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1820 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1822 #ifdef CONFIG_MEMORY_ISOLATION
1823 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1828 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1831 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1834 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1835 unsigned int order) { return NULL; }
1839 * Move the free pages in a range to the free lists of the requested type.
1840 * Note that start_page and end_pages are not aligned on a pageblock
1841 * boundary. If alignment is required, use move_freepages_block()
1843 int move_freepages(struct zone *zone,
1844 struct page *start_page, struct page *end_page,
1849 int pages_moved = 0;
1851 #ifndef CONFIG_HOLES_IN_ZONE
1853 * page_zone is not safe to call in this context when
1854 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1855 * anyway as we check zone boundaries in move_freepages_block().
1856 * Remove at a later date when no bug reports exist related to
1857 * grouping pages by mobility
1859 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1862 for (page = start_page; page <= end_page;) {
1863 /* Make sure we are not inadvertently changing nodes */
1864 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1866 if (!pfn_valid_within(page_to_pfn(page))) {
1871 if (!PageBuddy(page)) {
1876 order = page_order(page);
1877 list_move(&page->lru,
1878 &zone->free_area[order].free_list[migratetype]);
1880 pages_moved += 1 << order;
1886 int move_freepages_block(struct zone *zone, struct page *page,
1889 unsigned long start_pfn, end_pfn;
1890 struct page *start_page, *end_page;
1892 start_pfn = page_to_pfn(page);
1893 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1894 start_page = pfn_to_page(start_pfn);
1895 end_page = start_page + pageblock_nr_pages - 1;
1896 end_pfn = start_pfn + pageblock_nr_pages - 1;
1898 /* Do not cross zone boundaries */
1899 if (!zone_spans_pfn(zone, start_pfn))
1901 if (!zone_spans_pfn(zone, end_pfn))
1904 return move_freepages(zone, start_page, end_page, migratetype);
1907 static void change_pageblock_range(struct page *pageblock_page,
1908 int start_order, int migratetype)
1910 int nr_pageblocks = 1 << (start_order - pageblock_order);
1912 while (nr_pageblocks--) {
1913 set_pageblock_migratetype(pageblock_page, migratetype);
1914 pageblock_page += pageblock_nr_pages;
1919 * When we are falling back to another migratetype during allocation, try to
1920 * steal extra free pages from the same pageblocks to satisfy further
1921 * allocations, instead of polluting multiple pageblocks.
1923 * If we are stealing a relatively large buddy page, it is likely there will
1924 * be more free pages in the pageblock, so try to steal them all. For
1925 * reclaimable and unmovable allocations, we steal regardless of page size,
1926 * as fragmentation caused by those allocations polluting movable pageblocks
1927 * is worse than movable allocations stealing from unmovable and reclaimable
1930 static bool can_steal_fallback(unsigned int order, int start_mt)
1933 * Leaving this order check is intended, although there is
1934 * relaxed order check in next check. The reason is that
1935 * we can actually steal whole pageblock if this condition met,
1936 * but, below check doesn't guarantee it and that is just heuristic
1937 * so could be changed anytime.
1939 if (order >= pageblock_order)
1942 if (order >= pageblock_order / 2 ||
1943 start_mt == MIGRATE_RECLAIMABLE ||
1944 start_mt == MIGRATE_UNMOVABLE ||
1945 page_group_by_mobility_disabled)
1952 * This function implements actual steal behaviour. If order is large enough,
1953 * we can steal whole pageblock. If not, we first move freepages in this
1954 * pageblock and check whether half of pages are moved or not. If half of
1955 * pages are moved, we can change migratetype of pageblock and permanently
1956 * use it's pages as requested migratetype in the future.
1958 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1961 unsigned int current_order = page_order(page);
1964 /* Take ownership for orders >= pageblock_order */
1965 if (current_order >= pageblock_order) {
1966 change_pageblock_range(page, current_order, start_type);
1970 pages = move_freepages_block(zone, page, start_type);
1972 /* Claim the whole block if over half of it is free */
1973 if (pages >= (1 << (pageblock_order-1)) ||
1974 page_group_by_mobility_disabled)
1975 set_pageblock_migratetype(page, start_type);
1979 * Check whether there is a suitable fallback freepage with requested order.
1980 * If only_stealable is true, this function returns fallback_mt only if
1981 * we can steal other freepages all together. This would help to reduce
1982 * fragmentation due to mixed migratetype pages in one pageblock.
1984 int find_suitable_fallback(struct free_area *area, unsigned int order,
1985 int migratetype, bool only_stealable, bool *can_steal)
1990 if (area->nr_free == 0)
1995 fallback_mt = fallbacks[migratetype][i];
1996 if (fallback_mt == MIGRATE_TYPES)
1999 if (list_empty(&area->free_list[fallback_mt]))
2002 if (can_steal_fallback(order, migratetype))
2005 if (!only_stealable)
2016 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2017 * there are no empty page blocks that contain a page with a suitable order
2019 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2020 unsigned int alloc_order)
2023 unsigned long max_managed, flags;
2026 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2027 * Check is race-prone but harmless.
2029 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2030 if (zone->nr_reserved_highatomic >= max_managed)
2033 spin_lock_irqsave(&zone->lock, flags);
2035 /* Recheck the nr_reserved_highatomic limit under the lock */
2036 if (zone->nr_reserved_highatomic >= max_managed)
2040 mt = get_pageblock_migratetype(page);
2041 if (mt != MIGRATE_HIGHATOMIC &&
2042 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2043 zone->nr_reserved_highatomic += pageblock_nr_pages;
2044 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2045 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2049 spin_unlock_irqrestore(&zone->lock, flags);
2053 * Used when an allocation is about to fail under memory pressure. This
2054 * potentially hurts the reliability of high-order allocations when under
2055 * intense memory pressure but failed atomic allocations should be easier
2056 * to recover from than an OOM.
2058 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2060 struct zonelist *zonelist = ac->zonelist;
2061 unsigned long flags;
2067 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2069 /* Preserve at least one pageblock */
2070 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2073 spin_lock_irqsave(&zone->lock, flags);
2074 for (order = 0; order < MAX_ORDER; order++) {
2075 struct free_area *area = &(zone->free_area[order]);
2077 page = list_first_entry_or_null(
2078 &area->free_list[MIGRATE_HIGHATOMIC],
2084 * It should never happen but changes to locking could
2085 * inadvertently allow a per-cpu drain to add pages
2086 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2087 * and watch for underflows.
2089 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2090 zone->nr_reserved_highatomic);
2093 * Convert to ac->migratetype and avoid the normal
2094 * pageblock stealing heuristics. Minimally, the caller
2095 * is doing the work and needs the pages. More
2096 * importantly, if the block was always converted to
2097 * MIGRATE_UNMOVABLE or another type then the number
2098 * of pageblocks that cannot be completely freed
2101 set_pageblock_migratetype(page, ac->migratetype);
2102 move_freepages_block(zone, page, ac->migratetype);
2103 spin_unlock_irqrestore(&zone->lock, flags);
2106 spin_unlock_irqrestore(&zone->lock, flags);
2110 /* Remove an element from the buddy allocator from the fallback list */
2111 static inline struct page *
2112 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2114 struct free_area *area;
2115 unsigned int current_order;
2120 /* Find the largest possible block of pages in the other list */
2121 for (current_order = MAX_ORDER-1;
2122 current_order >= order && current_order <= MAX_ORDER-1;
2124 area = &(zone->free_area[current_order]);
2125 fallback_mt = find_suitable_fallback(area, current_order,
2126 start_migratetype, false, &can_steal);
2127 if (fallback_mt == -1)
2130 page = list_first_entry(&area->free_list[fallback_mt],
2133 steal_suitable_fallback(zone, page, start_migratetype);
2135 /* Remove the page from the freelists */
2137 list_del(&page->lru);
2138 rmv_page_order(page);
2140 expand(zone, page, order, current_order, area,
2143 * The pcppage_migratetype may differ from pageblock's
2144 * migratetype depending on the decisions in
2145 * find_suitable_fallback(). This is OK as long as it does not
2146 * differ for MIGRATE_CMA pageblocks. Those can be used as
2147 * fallback only via special __rmqueue_cma_fallback() function
2149 set_pcppage_migratetype(page, start_migratetype);
2151 trace_mm_page_alloc_extfrag(page, order, current_order,
2152 start_migratetype, fallback_mt);
2161 * Do the hard work of removing an element from the buddy allocator.
2162 * Call me with the zone->lock already held.
2164 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2169 page = __rmqueue_smallest(zone, order, migratetype);
2170 if (unlikely(!page)) {
2171 if (migratetype == MIGRATE_MOVABLE)
2172 page = __rmqueue_cma_fallback(zone, order);
2175 page = __rmqueue_fallback(zone, order, migratetype);
2178 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2183 * Obtain a specified number of elements from the buddy allocator, all under
2184 * a single hold of the lock, for efficiency. Add them to the supplied list.
2185 * Returns the number of new pages which were placed at *list.
2187 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2188 unsigned long count, struct list_head *list,
2189 int migratetype, bool cold)
2193 spin_lock(&zone->lock);
2194 for (i = 0; i < count; ++i) {
2195 struct page *page = __rmqueue(zone, order, migratetype);
2196 if (unlikely(page == NULL))
2199 if (unlikely(check_pcp_refill(page)))
2203 * Split buddy pages returned by expand() are received here
2204 * in physical page order. The page is added to the callers and
2205 * list and the list head then moves forward. From the callers
2206 * perspective, the linked list is ordered by page number in
2207 * some conditions. This is useful for IO devices that can
2208 * merge IO requests if the physical pages are ordered
2212 list_add(&page->lru, list);
2214 list_add_tail(&page->lru, list);
2216 if (is_migrate_cma(get_pcppage_migratetype(page)))
2217 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2220 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2221 spin_unlock(&zone->lock);
2227 * Called from the vmstat counter updater to drain pagesets of this
2228 * currently executing processor on remote nodes after they have
2231 * Note that this function must be called with the thread pinned to
2232 * a single processor.
2234 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2236 unsigned long flags;
2237 int to_drain, batch;
2239 local_irq_save(flags);
2240 batch = READ_ONCE(pcp->batch);
2241 to_drain = min(pcp->count, batch);
2243 free_pcppages_bulk(zone, to_drain, pcp);
2244 pcp->count -= to_drain;
2246 local_irq_restore(flags);
2251 * Drain pcplists of the indicated processor and zone.
2253 * The processor must either be the current processor and the
2254 * thread pinned to the current processor or a processor that
2257 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2259 unsigned long flags;
2260 struct per_cpu_pageset *pset;
2261 struct per_cpu_pages *pcp;
2263 local_irq_save(flags);
2264 pset = per_cpu_ptr(zone->pageset, cpu);
2268 free_pcppages_bulk(zone, pcp->count, pcp);
2271 local_irq_restore(flags);
2275 * Drain pcplists of all zones on the indicated processor.
2277 * The processor must either be the current processor and the
2278 * thread pinned to the current processor or a processor that
2281 static void drain_pages(unsigned int cpu)
2285 for_each_populated_zone(zone) {
2286 drain_pages_zone(cpu, zone);
2291 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2293 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2294 * the single zone's pages.
2296 void drain_local_pages(struct zone *zone)
2298 int cpu = smp_processor_id();
2301 drain_pages_zone(cpu, zone);
2307 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2309 * When zone parameter is non-NULL, spill just the single zone's pages.
2311 * Note that this code is protected against sending an IPI to an offline
2312 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2313 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2314 * nothing keeps CPUs from showing up after we populated the cpumask and
2315 * before the call to on_each_cpu_mask().
2317 void drain_all_pages(struct zone *zone)
2322 * Allocate in the BSS so we wont require allocation in
2323 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2325 static cpumask_t cpus_with_pcps;
2328 * We don't care about racing with CPU hotplug event
2329 * as offline notification will cause the notified
2330 * cpu to drain that CPU pcps and on_each_cpu_mask
2331 * disables preemption as part of its processing
2333 for_each_online_cpu(cpu) {
2334 struct per_cpu_pageset *pcp;
2336 bool has_pcps = false;
2339 pcp = per_cpu_ptr(zone->pageset, cpu);
2343 for_each_populated_zone(z) {
2344 pcp = per_cpu_ptr(z->pageset, cpu);
2345 if (pcp->pcp.count) {
2353 cpumask_set_cpu(cpu, &cpus_with_pcps);
2355 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2357 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2361 #ifdef CONFIG_HIBERNATION
2363 void mark_free_pages(struct zone *zone)
2365 unsigned long pfn, max_zone_pfn;
2366 unsigned long flags;
2367 unsigned int order, t;
2370 if (zone_is_empty(zone))
2373 spin_lock_irqsave(&zone->lock, flags);
2375 max_zone_pfn = zone_end_pfn(zone);
2376 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2377 if (pfn_valid(pfn)) {
2378 page = pfn_to_page(pfn);
2380 if (page_zone(page) != zone)
2383 if (!swsusp_page_is_forbidden(page))
2384 swsusp_unset_page_free(page);
2387 for_each_migratetype_order(order, t) {
2388 list_for_each_entry(page,
2389 &zone->free_area[order].free_list[t], lru) {
2392 pfn = page_to_pfn(page);
2393 for (i = 0; i < (1UL << order); i++)
2394 swsusp_set_page_free(pfn_to_page(pfn + i));
2397 spin_unlock_irqrestore(&zone->lock, flags);
2399 #endif /* CONFIG_PM */
2402 * Free a 0-order page
2403 * cold == true ? free a cold page : free a hot page
2405 void free_hot_cold_page(struct page *page, bool cold)
2407 struct zone *zone = page_zone(page);
2408 struct per_cpu_pages *pcp;
2409 unsigned long flags;
2410 unsigned long pfn = page_to_pfn(page);
2413 if (!free_pcp_prepare(page))
2416 migratetype = get_pfnblock_migratetype(page, pfn);
2417 set_pcppage_migratetype(page, migratetype);
2418 local_irq_save(flags);
2419 __count_vm_event(PGFREE);
2422 * We only track unmovable, reclaimable and movable on pcp lists.
2423 * Free ISOLATE pages back to the allocator because they are being
2424 * offlined but treat RESERVE as movable pages so we can get those
2425 * areas back if necessary. Otherwise, we may have to free
2426 * excessively into the page allocator
2428 if (migratetype >= MIGRATE_PCPTYPES) {
2429 if (unlikely(is_migrate_isolate(migratetype))) {
2430 free_one_page(zone, page, pfn, 0, migratetype);
2433 migratetype = MIGRATE_MOVABLE;
2436 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2438 list_add(&page->lru, &pcp->lists[migratetype]);
2440 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2442 if (pcp->count >= pcp->high) {
2443 unsigned long batch = READ_ONCE(pcp->batch);
2444 free_pcppages_bulk(zone, batch, pcp);
2445 pcp->count -= batch;
2449 local_irq_restore(flags);
2453 * Free a list of 0-order pages
2455 void free_hot_cold_page_list(struct list_head *list, bool cold)
2457 struct page *page, *next;
2459 list_for_each_entry_safe(page, next, list, lru) {
2460 trace_mm_page_free_batched(page, cold);
2461 free_hot_cold_page(page, cold);
2466 * split_page takes a non-compound higher-order page, and splits it into
2467 * n (1<<order) sub-pages: page[0..n]
2468 * Each sub-page must be freed individually.
2470 * Note: this is probably too low level an operation for use in drivers.
2471 * Please consult with lkml before using this in your driver.
2473 void split_page(struct page *page, unsigned int order)
2477 VM_BUG_ON_PAGE(PageCompound(page), page);
2478 VM_BUG_ON_PAGE(!page_count(page), page);
2480 #ifdef CONFIG_KMEMCHECK
2482 * Split shadow pages too, because free(page[0]) would
2483 * otherwise free the whole shadow.
2485 if (kmemcheck_page_is_tracked(page))
2486 split_page(virt_to_page(page[0].shadow), order);
2489 for (i = 1; i < (1 << order); i++)
2490 set_page_refcounted(page + i);
2491 split_page_owner(page, order);
2493 EXPORT_SYMBOL_GPL(split_page);
2495 int __isolate_free_page(struct page *page, unsigned int order)
2497 unsigned long watermark;
2501 BUG_ON(!PageBuddy(page));
2503 zone = page_zone(page);
2504 mt = get_pageblock_migratetype(page);
2506 if (!is_migrate_isolate(mt)) {
2507 /* Obey watermarks as if the page was being allocated */
2508 watermark = low_wmark_pages(zone) + (1 << order);
2509 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2512 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2515 /* Remove page from free list */
2516 list_del(&page->lru);
2517 zone->free_area[order].nr_free--;
2518 rmv_page_order(page);
2520 /* Set the pageblock if the isolated page is at least a pageblock */
2521 if (order >= pageblock_order - 1) {
2522 struct page *endpage = page + (1 << order) - 1;
2523 for (; page < endpage; page += pageblock_nr_pages) {
2524 int mt = get_pageblock_migratetype(page);
2525 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2526 set_pageblock_migratetype(page,
2532 return 1UL << order;
2536 * Update NUMA hit/miss statistics
2538 * Must be called with interrupts disabled.
2540 * When __GFP_OTHER_NODE is set assume the node of the preferred
2541 * zone is the local node. This is useful for daemons who allocate
2542 * memory on behalf of other processes.
2544 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2548 int local_nid = numa_node_id();
2549 enum zone_stat_item local_stat = NUMA_LOCAL;
2551 if (unlikely(flags & __GFP_OTHER_NODE)) {
2552 local_stat = NUMA_OTHER;
2553 local_nid = preferred_zone->node;
2556 if (z->node == local_nid) {
2557 __inc_zone_state(z, NUMA_HIT);
2558 __inc_zone_state(z, local_stat);
2560 __inc_zone_state(z, NUMA_MISS);
2561 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2567 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2570 struct page *buffered_rmqueue(struct zone *preferred_zone,
2571 struct zone *zone, unsigned int order,
2572 gfp_t gfp_flags, unsigned int alloc_flags,
2575 unsigned long flags;
2577 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2579 if (likely(order == 0)) {
2580 struct per_cpu_pages *pcp;
2581 struct list_head *list;
2583 local_irq_save(flags);
2585 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2586 list = &pcp->lists[migratetype];
2587 if (list_empty(list)) {
2588 pcp->count += rmqueue_bulk(zone, 0,
2591 if (unlikely(list_empty(list)))
2596 page = list_last_entry(list, struct page, lru);
2598 page = list_first_entry(list, struct page, lru);
2600 __dec_zone_state(zone, NR_ALLOC_BATCH);
2601 list_del(&page->lru);
2604 } while (check_new_pcp(page));
2607 * We most definitely don't want callers attempting to
2608 * allocate greater than order-1 page units with __GFP_NOFAIL.
2610 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2611 spin_lock_irqsave(&zone->lock, flags);
2615 if (alloc_flags & ALLOC_HARDER) {
2616 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2618 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2621 page = __rmqueue(zone, order, migratetype);
2622 } while (page && check_new_pages(page, order));
2623 spin_unlock(&zone->lock);
2626 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2627 __mod_zone_freepage_state(zone, -(1 << order),
2628 get_pcppage_migratetype(page));
2631 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2632 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2633 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2635 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2636 zone_statistics(preferred_zone, zone, gfp_flags);
2637 local_irq_restore(flags);
2639 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2643 local_irq_restore(flags);
2647 #ifdef CONFIG_FAIL_PAGE_ALLOC
2650 struct fault_attr attr;
2652 bool ignore_gfp_highmem;
2653 bool ignore_gfp_reclaim;
2655 } fail_page_alloc = {
2656 .attr = FAULT_ATTR_INITIALIZER,
2657 .ignore_gfp_reclaim = true,
2658 .ignore_gfp_highmem = true,
2662 static int __init setup_fail_page_alloc(char *str)
2664 return setup_fault_attr(&fail_page_alloc.attr, str);
2666 __setup("fail_page_alloc=", setup_fail_page_alloc);
2668 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2670 if (order < fail_page_alloc.min_order)
2672 if (gfp_mask & __GFP_NOFAIL)
2674 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2676 if (fail_page_alloc.ignore_gfp_reclaim &&
2677 (gfp_mask & __GFP_DIRECT_RECLAIM))
2680 return should_fail(&fail_page_alloc.attr, 1 << order);
2683 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2685 static int __init fail_page_alloc_debugfs(void)
2687 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2690 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2691 &fail_page_alloc.attr);
2693 return PTR_ERR(dir);
2695 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2696 &fail_page_alloc.ignore_gfp_reclaim))
2698 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2699 &fail_page_alloc.ignore_gfp_highmem))
2701 if (!debugfs_create_u32("min-order", mode, dir,
2702 &fail_page_alloc.min_order))
2707 debugfs_remove_recursive(dir);
2712 late_initcall(fail_page_alloc_debugfs);
2714 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2716 #else /* CONFIG_FAIL_PAGE_ALLOC */
2718 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2723 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2726 * Return true if free base pages are above 'mark'. For high-order checks it
2727 * will return true of the order-0 watermark is reached and there is at least
2728 * one free page of a suitable size. Checking now avoids taking the zone lock
2729 * to check in the allocation paths if no pages are free.
2731 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2732 int classzone_idx, unsigned int alloc_flags,
2737 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2739 /* free_pages may go negative - that's OK */
2740 free_pages -= (1 << order) - 1;
2742 if (alloc_flags & ALLOC_HIGH)
2746 * If the caller does not have rights to ALLOC_HARDER then subtract
2747 * the high-atomic reserves. This will over-estimate the size of the
2748 * atomic reserve but it avoids a search.
2750 if (likely(!alloc_harder))
2751 free_pages -= z->nr_reserved_highatomic;
2756 /* If allocation can't use CMA areas don't use free CMA pages */
2757 if (!(alloc_flags & ALLOC_CMA))
2758 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2762 * Check watermarks for an order-0 allocation request. If these
2763 * are not met, then a high-order request also cannot go ahead
2764 * even if a suitable page happened to be free.
2766 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2769 /* If this is an order-0 request then the watermark is fine */
2773 /* For a high-order request, check at least one suitable page is free */
2774 for (o = order; o < MAX_ORDER; o++) {
2775 struct free_area *area = &z->free_area[o];
2784 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2785 if (!list_empty(&area->free_list[mt]))
2790 if ((alloc_flags & ALLOC_CMA) &&
2791 !list_empty(&area->free_list[MIGRATE_CMA])) {
2799 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2800 int classzone_idx, unsigned int alloc_flags)
2802 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2803 zone_page_state(z, NR_FREE_PAGES));
2806 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2807 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2809 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2813 /* If allocation can't use CMA areas don't use free CMA pages */
2814 if (!(alloc_flags & ALLOC_CMA))
2815 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2819 * Fast check for order-0 only. If this fails then the reserves
2820 * need to be calculated. There is a corner case where the check
2821 * passes but only the high-order atomic reserve are free. If
2822 * the caller is !atomic then it'll uselessly search the free
2823 * list. That corner case is then slower but it is harmless.
2825 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2828 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2832 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2833 unsigned long mark, int classzone_idx)
2835 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2837 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2838 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2840 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2845 static bool zone_local(struct zone *local_zone, struct zone *zone)
2847 return local_zone->node == zone->node;
2850 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2852 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2855 #else /* CONFIG_NUMA */
2856 static bool zone_local(struct zone *local_zone, struct zone *zone)
2861 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2865 #endif /* CONFIG_NUMA */
2867 static void reset_alloc_batches(struct zone *preferred_zone)
2869 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2872 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2873 high_wmark_pages(zone) - low_wmark_pages(zone) -
2874 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2875 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2876 } while (zone++ != preferred_zone);
2880 * get_page_from_freelist goes through the zonelist trying to allocate
2883 static struct page *
2884 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2885 const struct alloc_context *ac)
2887 struct zoneref *z = ac->preferred_zoneref;
2889 bool fair_skipped = false;
2890 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2894 * Scan zonelist, looking for a zone with enough free.
2895 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2897 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2902 if (cpusets_enabled() &&
2903 (alloc_flags & ALLOC_CPUSET) &&
2904 !__cpuset_zone_allowed(zone, gfp_mask))
2907 * Distribute pages in proportion to the individual
2908 * zone size to ensure fair page aging. The zone a
2909 * page was allocated in should have no effect on the
2910 * time the page has in memory before being reclaimed.
2913 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2914 fair_skipped = true;
2917 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2924 * When allocating a page cache page for writing, we
2925 * want to get it from a zone that is within its dirty
2926 * limit, such that no single zone holds more than its
2927 * proportional share of globally allowed dirty pages.
2928 * The dirty limits take into account the zone's
2929 * lowmem reserves and high watermark so that kswapd
2930 * should be able to balance it without having to
2931 * write pages from its LRU list.
2933 * This may look like it could increase pressure on
2934 * lower zones by failing allocations in higher zones
2935 * before they are full. But the pages that do spill
2936 * over are limited as the lower zones are protected
2937 * by this very same mechanism. It should not become
2938 * a practical burden to them.
2940 * XXX: For now, allow allocations to potentially
2941 * exceed the per-zone dirty limit in the slowpath
2942 * (spread_dirty_pages unset) before going into reclaim,
2943 * which is important when on a NUMA setup the allowed
2944 * zones are together not big enough to reach the
2945 * global limit. The proper fix for these situations
2946 * will require awareness of zones in the
2947 * dirty-throttling and the flusher threads.
2949 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2952 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2953 if (!zone_watermark_fast(zone, order, mark,
2954 ac_classzone_idx(ac), alloc_flags)) {
2957 /* Checked here to keep the fast path fast */
2958 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2959 if (alloc_flags & ALLOC_NO_WATERMARKS)
2962 if (zone_reclaim_mode == 0 ||
2963 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2966 ret = zone_reclaim(zone, gfp_mask, order);
2968 case ZONE_RECLAIM_NOSCAN:
2971 case ZONE_RECLAIM_FULL:
2972 /* scanned but unreclaimable */
2975 /* did we reclaim enough */
2976 if (zone_watermark_ok(zone, order, mark,
2977 ac_classzone_idx(ac), alloc_flags))
2985 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2986 gfp_mask, alloc_flags, ac->migratetype);
2988 prep_new_page(page, order, gfp_mask, alloc_flags);
2991 * If this is a high-order atomic allocation then check
2992 * if the pageblock should be reserved for the future
2994 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2995 reserve_highatomic_pageblock(page, zone, order);
3002 * The first pass makes sure allocations are spread fairly within the
3003 * local node. However, the local node might have free pages left
3004 * after the fairness batches are exhausted, and remote zones haven't
3005 * even been considered yet. Try once more without fairness, and
3006 * include remote zones now, before entering the slowpath and waking
3007 * kswapd: prefer spilling to a remote zone over swapping locally.
3012 fair_skipped = false;
3013 reset_alloc_batches(ac->preferred_zoneref->zone);
3014 z = ac->preferred_zoneref;
3022 * Large machines with many possible nodes should not always dump per-node
3023 * meminfo in irq context.
3025 static inline bool should_suppress_show_mem(void)
3030 ret = in_interrupt();
3035 static DEFINE_RATELIMIT_STATE(nopage_rs,
3036 DEFAULT_RATELIMIT_INTERVAL,
3037 DEFAULT_RATELIMIT_BURST);
3039 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3041 unsigned int filter = SHOW_MEM_FILTER_NODES;
3043 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3044 debug_guardpage_minorder() > 0)
3048 * This documents exceptions given to allocations in certain
3049 * contexts that are allowed to allocate outside current's set
3052 if (!(gfp_mask & __GFP_NOMEMALLOC))
3053 if (test_thread_flag(TIF_MEMDIE) ||
3054 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3055 filter &= ~SHOW_MEM_FILTER_NODES;
3056 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3057 filter &= ~SHOW_MEM_FILTER_NODES;
3060 struct va_format vaf;
3063 va_start(args, fmt);
3068 pr_warn("%pV", &vaf);
3073 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3074 current->comm, order, gfp_mask, &gfp_mask);
3076 if (!should_suppress_show_mem())
3080 static inline struct page *
3081 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3082 const struct alloc_context *ac, unsigned long *did_some_progress)
3084 struct oom_control oc = {
3085 .zonelist = ac->zonelist,
3086 .nodemask = ac->nodemask,
3088 .gfp_mask = gfp_mask,
3093 *did_some_progress = 0;
3096 * Acquire the oom lock. If that fails, somebody else is
3097 * making progress for us.
3099 if (!mutex_trylock(&oom_lock)) {
3100 *did_some_progress = 1;
3101 schedule_timeout_uninterruptible(1);
3106 * Go through the zonelist yet one more time, keep very high watermark
3107 * here, this is only to catch a parallel oom killing, we must fail if
3108 * we're still under heavy pressure.
3110 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3111 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3115 if (!(gfp_mask & __GFP_NOFAIL)) {
3116 /* Coredumps can quickly deplete all memory reserves */
3117 if (current->flags & PF_DUMPCORE)
3119 /* The OOM killer will not help higher order allocs */
3120 if (order > PAGE_ALLOC_COSTLY_ORDER)
3122 /* The OOM killer does not needlessly kill tasks for lowmem */
3123 if (ac->high_zoneidx < ZONE_NORMAL)
3125 if (pm_suspended_storage())
3128 * XXX: GFP_NOFS allocations should rather fail than rely on
3129 * other request to make a forward progress.
3130 * We are in an unfortunate situation where out_of_memory cannot
3131 * do much for this context but let's try it to at least get
3132 * access to memory reserved if the current task is killed (see
3133 * out_of_memory). Once filesystems are ready to handle allocation
3134 * failures more gracefully we should just bail out here.
3137 /* The OOM killer may not free memory on a specific node */
3138 if (gfp_mask & __GFP_THISNODE)
3141 /* Exhausted what can be done so it's blamo time */
3142 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3143 *did_some_progress = 1;
3145 if (gfp_mask & __GFP_NOFAIL) {
3146 page = get_page_from_freelist(gfp_mask, order,
3147 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3149 * fallback to ignore cpuset restriction if our nodes
3153 page = get_page_from_freelist(gfp_mask, order,
3154 ALLOC_NO_WATERMARKS, ac);
3158 mutex_unlock(&oom_lock);
3164 * Maximum number of compaction retries wit a progress before OOM
3165 * killer is consider as the only way to move forward.
3167 #define MAX_COMPACT_RETRIES 16
3169 #ifdef CONFIG_COMPACTION
3170 /* Try memory compaction for high-order allocations before reclaim */
3171 static struct page *
3172 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3173 unsigned int alloc_flags, const struct alloc_context *ac,
3174 enum migrate_mode mode, enum compact_result *compact_result)
3177 int contended_compaction;
3182 current->flags |= PF_MEMALLOC;
3183 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3184 mode, &contended_compaction);
3185 current->flags &= ~PF_MEMALLOC;
3187 if (*compact_result <= COMPACT_INACTIVE)
3191 * At least in one zone compaction wasn't deferred or skipped, so let's
3192 * count a compaction stall
3194 count_vm_event(COMPACTSTALL);
3196 page = get_page_from_freelist(gfp_mask, order,
3197 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3200 struct zone *zone = page_zone(page);
3202 zone->compact_blockskip_flush = false;
3203 compaction_defer_reset(zone, order, true);
3204 count_vm_event(COMPACTSUCCESS);
3209 * It's bad if compaction run occurs and fails. The most likely reason
3210 * is that pages exist, but not enough to satisfy watermarks.
3212 count_vm_event(COMPACTFAIL);
3215 * In all zones where compaction was attempted (and not
3216 * deferred or skipped), lock contention has been detected.
3217 * For THP allocation we do not want to disrupt the others
3218 * so we fallback to base pages instead.
3220 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3221 *compact_result = COMPACT_CONTENDED;
3224 * If compaction was aborted due to need_resched(), we do not
3225 * want to further increase allocation latency, unless it is
3226 * khugepaged trying to collapse.
3228 if (contended_compaction == COMPACT_CONTENDED_SCHED
3229 && !(current->flags & PF_KTHREAD))
3230 *compact_result = COMPACT_CONTENDED;
3238 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3239 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3240 int compaction_retries)
3242 int max_retries = MAX_COMPACT_RETRIES;
3248 * compaction considers all the zone as desperately out of memory
3249 * so it doesn't really make much sense to retry except when the
3250 * failure could be caused by weak migration mode.
3252 if (compaction_failed(compact_result)) {
3253 if (*migrate_mode == MIGRATE_ASYNC) {
3254 *migrate_mode = MIGRATE_SYNC_LIGHT;
3261 * make sure the compaction wasn't deferred or didn't bail out early
3262 * due to locks contention before we declare that we should give up.
3263 * But do not retry if the given zonelist is not suitable for
3266 if (compaction_withdrawn(compact_result))
3267 return compaction_zonelist_suitable(ac, order, alloc_flags);
3270 * !costly requests are much more important than __GFP_REPEAT
3271 * costly ones because they are de facto nofail and invoke OOM
3272 * killer to move on while costly can fail and users are ready
3273 * to cope with that. 1/4 retries is rather arbitrary but we
3274 * would need much more detailed feedback from compaction to
3275 * make a better decision.
3277 if (order > PAGE_ALLOC_COSTLY_ORDER)
3279 if (compaction_retries <= max_retries)
3285 static inline struct page *
3286 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3287 unsigned int alloc_flags, const struct alloc_context *ac,
3288 enum migrate_mode mode, enum compact_result *compact_result)
3290 *compact_result = COMPACT_SKIPPED;
3295 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3296 enum compact_result compact_result,
3297 enum migrate_mode *migrate_mode,
3298 int compaction_retries)
3303 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3307 * There are setups with compaction disabled which would prefer to loop
3308 * inside the allocator rather than hit the oom killer prematurely.
3309 * Let's give them a good hope and keep retrying while the order-0
3310 * watermarks are OK.
3312 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3314 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3315 ac_classzone_idx(ac), alloc_flags))
3320 #endif /* CONFIG_COMPACTION */
3322 /* Perform direct synchronous page reclaim */
3324 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3325 const struct alloc_context *ac)
3327 struct reclaim_state reclaim_state;
3332 /* We now go into synchronous reclaim */
3333 cpuset_memory_pressure_bump();
3334 current->flags |= PF_MEMALLOC;
3335 lockdep_set_current_reclaim_state(gfp_mask);
3336 reclaim_state.reclaimed_slab = 0;
3337 current->reclaim_state = &reclaim_state;
3339 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3342 current->reclaim_state = NULL;
3343 lockdep_clear_current_reclaim_state();
3344 current->flags &= ~PF_MEMALLOC;
3351 /* The really slow allocator path where we enter direct reclaim */
3352 static inline struct page *
3353 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3354 unsigned int alloc_flags, const struct alloc_context *ac,
3355 unsigned long *did_some_progress)
3357 struct page *page = NULL;
3358 bool drained = false;
3360 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3361 if (unlikely(!(*did_some_progress)))
3365 page = get_page_from_freelist(gfp_mask, order,
3366 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3369 * If an allocation failed after direct reclaim, it could be because
3370 * pages are pinned on the per-cpu lists or in high alloc reserves.
3371 * Shrink them them and try again
3373 if (!page && !drained) {
3374 unreserve_highatomic_pageblock(ac);
3375 drain_all_pages(NULL);
3383 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3388 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3389 ac->high_zoneidx, ac->nodemask)
3390 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3393 static inline unsigned int
3394 gfp_to_alloc_flags(gfp_t gfp_mask)
3396 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3398 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3399 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3402 * The caller may dip into page reserves a bit more if the caller
3403 * cannot run direct reclaim, or if the caller has realtime scheduling
3404 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3405 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3407 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3409 if (gfp_mask & __GFP_ATOMIC) {
3411 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3412 * if it can't schedule.
3414 if (!(gfp_mask & __GFP_NOMEMALLOC))
3415 alloc_flags |= ALLOC_HARDER;
3417 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3418 * comment for __cpuset_node_allowed().
3420 alloc_flags &= ~ALLOC_CPUSET;
3421 } else if (unlikely(rt_task(current)) && !in_interrupt())
3422 alloc_flags |= ALLOC_HARDER;
3424 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3425 if (gfp_mask & __GFP_MEMALLOC)
3426 alloc_flags |= ALLOC_NO_WATERMARKS;
3427 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3428 alloc_flags |= ALLOC_NO_WATERMARKS;
3429 else if (!in_interrupt() &&
3430 ((current->flags & PF_MEMALLOC) ||
3431 unlikely(test_thread_flag(TIF_MEMDIE))))
3432 alloc_flags |= ALLOC_NO_WATERMARKS;
3435 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3436 alloc_flags |= ALLOC_CMA;
3441 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3443 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3446 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3448 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3452 * Maximum number of reclaim retries without any progress before OOM killer
3453 * is consider as the only way to move forward.
3455 #define MAX_RECLAIM_RETRIES 16
3458 * Checks whether it makes sense to retry the reclaim to make a forward progress
3459 * for the given allocation request.
3460 * The reclaim feedback represented by did_some_progress (any progress during
3461 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3462 * any progress in a row) is considered as well as the reclaimable pages on the
3463 * applicable zone list (with a backoff mechanism which is a function of
3464 * no_progress_loops).
3466 * Returns true if a retry is viable or false to enter the oom path.
3469 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3470 struct alloc_context *ac, int alloc_flags,
3471 bool did_some_progress, int no_progress_loops)
3477 * Make sure we converge to OOM if we cannot make any progress
3478 * several times in the row.
3480 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3484 * Keep reclaiming pages while there is a chance this will lead somewhere.
3485 * If none of the target zones can satisfy our allocation request even
3486 * if all reclaimable pages are considered then we are screwed and have
3489 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3491 unsigned long available;
3492 unsigned long reclaimable;
3494 available = reclaimable = zone_reclaimable_pages(zone);
3495 available -= DIV_ROUND_UP(no_progress_loops * available,
3496 MAX_RECLAIM_RETRIES);
3497 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3500 * Would the allocation succeed if we reclaimed the whole
3503 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3504 ac_classzone_idx(ac), alloc_flags, available)) {
3506 * If we didn't make any progress and have a lot of
3507 * dirty + writeback pages then we should wait for
3508 * an IO to complete to slow down the reclaim and
3509 * prevent from pre mature OOM
3511 if (!did_some_progress) {
3512 unsigned long writeback;
3513 unsigned long dirty;
3515 writeback = zone_page_state_snapshot(zone,
3517 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3519 if (2*(writeback + dirty) > reclaimable) {
3520 congestion_wait(BLK_RW_ASYNC, HZ/10);
3526 * Memory allocation/reclaim might be called from a WQ
3527 * context and the current implementation of the WQ
3528 * concurrency control doesn't recognize that
3529 * a particular WQ is congested if the worker thread is
3530 * looping without ever sleeping. Therefore we have to
3531 * do a short sleep here rather than calling
3534 if (current->flags & PF_WQ_WORKER)
3535 schedule_timeout_uninterruptible(1);
3546 static inline struct page *
3547 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3548 struct alloc_context *ac)
3550 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3551 struct page *page = NULL;
3552 unsigned int alloc_flags;
3553 unsigned long did_some_progress;
3554 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3555 enum compact_result compact_result;
3556 int compaction_retries = 0;
3557 int no_progress_loops = 0;
3560 * In the slowpath, we sanity check order to avoid ever trying to
3561 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3562 * be using allocators in order of preference for an area that is
3565 if (order >= MAX_ORDER) {
3566 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3571 * We also sanity check to catch abuse of atomic reserves being used by
3572 * callers that are not in atomic context.
3574 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3575 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3576 gfp_mask &= ~__GFP_ATOMIC;
3579 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3580 wake_all_kswapds(order, ac);
3583 * OK, we're below the kswapd watermark and have kicked background
3584 * reclaim. Now things get more complex, so set up alloc_flags according
3585 * to how we want to proceed.
3587 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3590 * Reset the zonelist iterators if memory policies can be ignored.
3591 * These allocations are high priority and system rather than user
3594 if ((alloc_flags & ALLOC_NO_WATERMARKS) || !(alloc_flags & ALLOC_CPUSET)) {
3595 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3596 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3597 ac->high_zoneidx, ac->nodemask);
3600 /* This is the last chance, in general, before the goto nopage. */
3601 page = get_page_from_freelist(gfp_mask, order,
3602 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3606 /* Allocate without watermarks if the context allows */
3607 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3608 page = get_page_from_freelist(gfp_mask, order,
3609 ALLOC_NO_WATERMARKS, ac);
3614 /* Caller is not willing to reclaim, we can't balance anything */
3615 if (!can_direct_reclaim) {
3617 * All existing users of the __GFP_NOFAIL are blockable, so warn
3618 * of any new users that actually allow this type of allocation
3621 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3625 /* Avoid recursion of direct reclaim */
3626 if (current->flags & PF_MEMALLOC) {
3628 * __GFP_NOFAIL request from this context is rather bizarre
3629 * because we cannot reclaim anything and only can loop waiting
3630 * for somebody to do a work for us.
3632 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3639 /* Avoid allocations with no watermarks from looping endlessly */
3640 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3644 * Try direct compaction. The first pass is asynchronous. Subsequent
3645 * attempts after direct reclaim are synchronous
3647 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3653 /* Checks for THP-specific high-order allocations */
3654 if (is_thp_gfp_mask(gfp_mask)) {
3656 * If compaction is deferred for high-order allocations, it is
3657 * because sync compaction recently failed. If this is the case
3658 * and the caller requested a THP allocation, we do not want
3659 * to heavily disrupt the system, so we fail the allocation
3660 * instead of entering direct reclaim.
3662 if (compact_result == COMPACT_DEFERRED)
3666 * Compaction is contended so rather back off than cause
3669 if(compact_result == COMPACT_CONTENDED)
3673 if (order && compaction_made_progress(compact_result))
3674 compaction_retries++;
3676 /* Try direct reclaim and then allocating */
3677 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3678 &did_some_progress);
3682 /* Do not loop if specifically requested */
3683 if (gfp_mask & __GFP_NORETRY)
3687 * Do not retry costly high order allocations unless they are
3690 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3694 * Costly allocations might have made a progress but this doesn't mean
3695 * their order will become available due to high fragmentation so
3696 * always increment the no progress counter for them
3698 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3699 no_progress_loops = 0;
3701 no_progress_loops++;
3703 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3704 did_some_progress > 0, no_progress_loops))
3708 * It doesn't make any sense to retry for the compaction if the order-0
3709 * reclaim is not able to make any progress because the current
3710 * implementation of the compaction depends on the sufficient amount
3711 * of free memory (see __compaction_suitable)
3713 if (did_some_progress > 0 &&
3714 should_compact_retry(ac, order, alloc_flags,
3715 compact_result, &migration_mode,
3716 compaction_retries))
3719 /* Reclaim has failed us, start killing things */
3720 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3724 /* Retry as long as the OOM killer is making progress */
3725 if (did_some_progress) {
3726 no_progress_loops = 0;
3732 * High-order allocations do not necessarily loop after direct reclaim
3733 * and reclaim/compaction depends on compaction being called after
3734 * reclaim so call directly if necessary.
3735 * It can become very expensive to allocate transparent hugepages at
3736 * fault, so use asynchronous memory compaction for THP unless it is
3737 * khugepaged trying to collapse. All other requests should tolerate
3738 * at least light sync migration.
3740 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3741 migration_mode = MIGRATE_ASYNC;
3743 migration_mode = MIGRATE_SYNC_LIGHT;
3744 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3750 warn_alloc_failed(gfp_mask, order, NULL);
3756 * This is the 'heart' of the zoned buddy allocator.
3759 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3760 struct zonelist *zonelist, nodemask_t *nodemask)
3763 unsigned int cpuset_mems_cookie;
3764 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3765 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3766 struct alloc_context ac = {
3767 .high_zoneidx = gfp_zone(gfp_mask),
3768 .zonelist = zonelist,
3769 .nodemask = nodemask,
3770 .migratetype = gfpflags_to_migratetype(gfp_mask),
3773 if (cpusets_enabled()) {
3774 alloc_mask |= __GFP_HARDWALL;
3775 alloc_flags |= ALLOC_CPUSET;
3777 ac.nodemask = &cpuset_current_mems_allowed;
3780 gfp_mask &= gfp_allowed_mask;
3782 lockdep_trace_alloc(gfp_mask);
3784 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3786 if (should_fail_alloc_page(gfp_mask, order))
3790 * Check the zones suitable for the gfp_mask contain at least one
3791 * valid zone. It's possible to have an empty zonelist as a result
3792 * of __GFP_THISNODE and a memoryless node
3794 if (unlikely(!zonelist->_zonerefs->zone))
3797 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3798 alloc_flags |= ALLOC_CMA;
3801 cpuset_mems_cookie = read_mems_allowed_begin();
3803 /* Dirty zone balancing only done in the fast path */
3804 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3807 * The preferred zone is used for statistics but crucially it is
3808 * also used as the starting point for the zonelist iterator. It
3809 * may get reset for allocations that ignore memory policies.
3811 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3812 ac.high_zoneidx, ac.nodemask);
3813 if (!ac.preferred_zoneref) {
3818 /* First allocation attempt */
3819 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3824 * Runtime PM, block IO and its error handling path can deadlock
3825 * because I/O on the device might not complete.
3827 alloc_mask = memalloc_noio_flags(gfp_mask);
3828 ac.spread_dirty_pages = false;
3831 * Restore the original nodemask if it was potentially replaced with
3832 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3834 if (cpusets_enabled())
3835 ac.nodemask = nodemask;
3836 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3840 * When updating a task's mems_allowed, it is possible to race with
3841 * parallel threads in such a way that an allocation can fail while
3842 * the mask is being updated. If a page allocation is about to fail,
3843 * check if the cpuset changed during allocation and if so, retry.
3845 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3846 alloc_mask = gfp_mask;
3851 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page) {
3852 if (unlikely(memcg_kmem_charge(page, gfp_mask, order))) {
3853 __free_pages(page, order);
3856 __SetPageKmemcg(page);
3859 if (kmemcheck_enabled && page)
3860 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3862 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3866 EXPORT_SYMBOL(__alloc_pages_nodemask);
3869 * Common helper functions.
3871 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3876 * __get_free_pages() returns a 32-bit address, which cannot represent
3879 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3881 page = alloc_pages(gfp_mask, order);
3884 return (unsigned long) page_address(page);
3886 EXPORT_SYMBOL(__get_free_pages);
3888 unsigned long get_zeroed_page(gfp_t gfp_mask)
3890 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3892 EXPORT_SYMBOL(get_zeroed_page);
3894 void __free_pages(struct page *page, unsigned int order)
3896 if (put_page_testzero(page)) {
3898 free_hot_cold_page(page, false);
3900 __free_pages_ok(page, order);
3904 EXPORT_SYMBOL(__free_pages);
3906 void free_pages(unsigned long addr, unsigned int order)
3909 VM_BUG_ON(!virt_addr_valid((void *)addr));
3910 __free_pages(virt_to_page((void *)addr), order);
3914 EXPORT_SYMBOL(free_pages);
3918 * An arbitrary-length arbitrary-offset area of memory which resides
3919 * within a 0 or higher order page. Multiple fragments within that page
3920 * are individually refcounted, in the page's reference counter.
3922 * The page_frag functions below provide a simple allocation framework for
3923 * page fragments. This is used by the network stack and network device
3924 * drivers to provide a backing region of memory for use as either an
3925 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3927 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3930 struct page *page = NULL;
3931 gfp_t gfp = gfp_mask;
3933 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3934 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3936 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3937 PAGE_FRAG_CACHE_MAX_ORDER);
3938 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3940 if (unlikely(!page))
3941 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3943 nc->va = page ? page_address(page) : NULL;
3948 void *__alloc_page_frag(struct page_frag_cache *nc,
3949 unsigned int fragsz, gfp_t gfp_mask)
3951 unsigned int size = PAGE_SIZE;
3955 if (unlikely(!nc->va)) {
3957 page = __page_frag_refill(nc, gfp_mask);
3961 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3962 /* if size can vary use size else just use PAGE_SIZE */
3965 /* Even if we own the page, we do not use atomic_set().
3966 * This would break get_page_unless_zero() users.
3968 page_ref_add(page, size - 1);
3970 /* reset page count bias and offset to start of new frag */
3971 nc->pfmemalloc = page_is_pfmemalloc(page);
3972 nc->pagecnt_bias = size;
3976 offset = nc->offset - fragsz;
3977 if (unlikely(offset < 0)) {
3978 page = virt_to_page(nc->va);
3980 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3983 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3984 /* if size can vary use size else just use PAGE_SIZE */
3987 /* OK, page count is 0, we can safely set it */
3988 set_page_count(page, size);
3990 /* reset page count bias and offset to start of new frag */
3991 nc->pagecnt_bias = size;
3992 offset = size - fragsz;
3996 nc->offset = offset;
3998 return nc->va + offset;
4000 EXPORT_SYMBOL(__alloc_page_frag);
4003 * Frees a page fragment allocated out of either a compound or order 0 page.
4005 void __free_page_frag(void *addr)
4007 struct page *page = virt_to_head_page(addr);
4009 if (unlikely(put_page_testzero(page)))
4010 __free_pages_ok(page, compound_order(page));
4012 EXPORT_SYMBOL(__free_page_frag);
4014 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4018 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4019 unsigned long used = addr + PAGE_ALIGN(size);
4021 split_page(virt_to_page((void *)addr), order);
4022 while (used < alloc_end) {
4027 return (void *)addr;
4031 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4032 * @size: the number of bytes to allocate
4033 * @gfp_mask: GFP flags for the allocation
4035 * This function is similar to alloc_pages(), except that it allocates the
4036 * minimum number of pages to satisfy the request. alloc_pages() can only
4037 * allocate memory in power-of-two pages.
4039 * This function is also limited by MAX_ORDER.
4041 * Memory allocated by this function must be released by free_pages_exact().
4043 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4045 unsigned int order = get_order(size);
4048 addr = __get_free_pages(gfp_mask, order);
4049 return make_alloc_exact(addr, order, size);
4051 EXPORT_SYMBOL(alloc_pages_exact);
4054 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4056 * @nid: the preferred node ID where memory should be allocated
4057 * @size: the number of bytes to allocate
4058 * @gfp_mask: GFP flags for the allocation
4060 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4063 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4065 unsigned int order = get_order(size);
4066 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4069 return make_alloc_exact((unsigned long)page_address(p), order, size);
4073 * free_pages_exact - release memory allocated via alloc_pages_exact()
4074 * @virt: the value returned by alloc_pages_exact.
4075 * @size: size of allocation, same value as passed to alloc_pages_exact().
4077 * Release the memory allocated by a previous call to alloc_pages_exact.
4079 void free_pages_exact(void *virt, size_t size)
4081 unsigned long addr = (unsigned long)virt;
4082 unsigned long end = addr + PAGE_ALIGN(size);
4084 while (addr < end) {
4089 EXPORT_SYMBOL(free_pages_exact);
4092 * nr_free_zone_pages - count number of pages beyond high watermark
4093 * @offset: The zone index of the highest zone
4095 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4096 * high watermark within all zones at or below a given zone index. For each
4097 * zone, the number of pages is calculated as:
4098 * managed_pages - high_pages
4100 static unsigned long nr_free_zone_pages(int offset)
4105 /* Just pick one node, since fallback list is circular */
4106 unsigned long sum = 0;
4108 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4110 for_each_zone_zonelist(zone, z, zonelist, offset) {
4111 unsigned long size = zone->managed_pages;
4112 unsigned long high = high_wmark_pages(zone);
4121 * nr_free_buffer_pages - count number of pages beyond high watermark
4123 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4124 * watermark within ZONE_DMA and ZONE_NORMAL.
4126 unsigned long nr_free_buffer_pages(void)
4128 return nr_free_zone_pages(gfp_zone(GFP_USER));
4130 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4133 * nr_free_pagecache_pages - count number of pages beyond high watermark
4135 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4136 * high watermark within all zones.
4138 unsigned long nr_free_pagecache_pages(void)
4140 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4143 static inline void show_node(struct zone *zone)
4145 if (IS_ENABLED(CONFIG_NUMA))
4146 printk("Node %d ", zone_to_nid(zone));
4149 long si_mem_available(void)
4152 unsigned long pagecache;
4153 unsigned long wmark_low = 0;
4154 unsigned long pages[NR_LRU_LISTS];
4158 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4159 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4162 wmark_low += zone->watermark[WMARK_LOW];
4165 * Estimate the amount of memory available for userspace allocations,
4166 * without causing swapping.
4168 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4171 * Not all the page cache can be freed, otherwise the system will
4172 * start swapping. Assume at least half of the page cache, or the
4173 * low watermark worth of cache, needs to stay.
4175 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4176 pagecache -= min(pagecache / 2, wmark_low);
4177 available += pagecache;
4180 * Part of the reclaimable slab consists of items that are in use,
4181 * and cannot be freed. Cap this estimate at the low watermark.
4183 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4184 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4190 EXPORT_SYMBOL_GPL(si_mem_available);
4192 void si_meminfo(struct sysinfo *val)
4194 val->totalram = totalram_pages;
4195 val->sharedram = global_page_state(NR_SHMEM);
4196 val->freeram = global_page_state(NR_FREE_PAGES);
4197 val->bufferram = nr_blockdev_pages();
4198 val->totalhigh = totalhigh_pages;
4199 val->freehigh = nr_free_highpages();
4200 val->mem_unit = PAGE_SIZE;
4203 EXPORT_SYMBOL(si_meminfo);
4206 void si_meminfo_node(struct sysinfo *val, int nid)
4208 int zone_type; /* needs to be signed */
4209 unsigned long managed_pages = 0;
4210 unsigned long managed_highpages = 0;
4211 unsigned long free_highpages = 0;
4212 pg_data_t *pgdat = NODE_DATA(nid);
4214 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4215 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4216 val->totalram = managed_pages;
4217 val->sharedram = node_page_state(nid, NR_SHMEM);
4218 val->freeram = node_page_state(nid, NR_FREE_PAGES);
4219 #ifdef CONFIG_HIGHMEM
4220 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4221 struct zone *zone = &pgdat->node_zones[zone_type];
4223 if (is_highmem(zone)) {
4224 managed_highpages += zone->managed_pages;
4225 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4228 val->totalhigh = managed_highpages;
4229 val->freehigh = free_highpages;
4231 val->totalhigh = managed_highpages;
4232 val->freehigh = free_highpages;
4234 val->mem_unit = PAGE_SIZE;
4239 * Determine whether the node should be displayed or not, depending on whether
4240 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4242 bool skip_free_areas_node(unsigned int flags, int nid)
4245 unsigned int cpuset_mems_cookie;
4247 if (!(flags & SHOW_MEM_FILTER_NODES))
4251 cpuset_mems_cookie = read_mems_allowed_begin();
4252 ret = !node_isset(nid, cpuset_current_mems_allowed);
4253 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4258 #define K(x) ((x) << (PAGE_SHIFT-10))
4260 static void show_migration_types(unsigned char type)
4262 static const char types[MIGRATE_TYPES] = {
4263 [MIGRATE_UNMOVABLE] = 'U',
4264 [MIGRATE_MOVABLE] = 'M',
4265 [MIGRATE_RECLAIMABLE] = 'E',
4266 [MIGRATE_HIGHATOMIC] = 'H',
4268 [MIGRATE_CMA] = 'C',
4270 #ifdef CONFIG_MEMORY_ISOLATION
4271 [MIGRATE_ISOLATE] = 'I',
4274 char tmp[MIGRATE_TYPES + 1];
4278 for (i = 0; i < MIGRATE_TYPES; i++) {
4279 if (type & (1 << i))
4284 printk("(%s) ", tmp);
4288 * Show free area list (used inside shift_scroll-lock stuff)
4289 * We also calculate the percentage fragmentation. We do this by counting the
4290 * memory on each free list with the exception of the first item on the list.
4293 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4296 void show_free_areas(unsigned int filter)
4298 unsigned long free_pcp = 0;
4302 for_each_populated_zone(zone) {
4303 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4306 for_each_online_cpu(cpu)
4307 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4310 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4311 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4312 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4313 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4314 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4315 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4316 " anon_thp: %lu shmem_thp: %lu shmem_pmdmapped: %lu\n"
4318 " free:%lu free_pcp:%lu free_cma:%lu\n",
4319 global_page_state(NR_ACTIVE_ANON),
4320 global_page_state(NR_INACTIVE_ANON),
4321 global_page_state(NR_ISOLATED_ANON),
4322 global_page_state(NR_ACTIVE_FILE),
4323 global_page_state(NR_INACTIVE_FILE),
4324 global_page_state(NR_ISOLATED_FILE),
4325 global_page_state(NR_UNEVICTABLE),
4326 global_page_state(NR_FILE_DIRTY),
4327 global_page_state(NR_WRITEBACK),
4328 global_page_state(NR_UNSTABLE_NFS),
4329 global_page_state(NR_SLAB_RECLAIMABLE),
4330 global_page_state(NR_SLAB_UNRECLAIMABLE),
4331 global_page_state(NR_FILE_MAPPED),
4332 global_page_state(NR_SHMEM),
4333 global_page_state(NR_PAGETABLE),
4334 global_page_state(NR_BOUNCE),
4335 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4336 global_page_state(NR_ANON_THPS) * HPAGE_PMD_NR,
4337 global_page_state(NR_SHMEM_THPS) * HPAGE_PMD_NR,
4338 global_page_state(NR_SHMEM_PMDMAPPED) * HPAGE_PMD_NR,
4340 global_page_state(NR_FREE_PAGES),
4342 global_page_state(NR_FREE_CMA_PAGES));
4344 for_each_populated_zone(zone) {
4347 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4351 for_each_online_cpu(cpu)
4352 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4360 " active_anon:%lukB"
4361 " inactive_anon:%lukB"
4362 " active_file:%lukB"
4363 " inactive_file:%lukB"
4364 " unevictable:%lukB"
4365 " isolated(anon):%lukB"
4366 " isolated(file):%lukB"
4374 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4376 " shmem_pmdmapped: %lukB"
4379 " slab_reclaimable:%lukB"
4380 " slab_unreclaimable:%lukB"
4381 " kernel_stack:%lukB"
4388 " writeback_tmp:%lukB"
4389 " pages_scanned:%lu"
4390 " all_unreclaimable? %s"
4393 K(zone_page_state(zone, NR_FREE_PAGES)),
4394 K(min_wmark_pages(zone)),
4395 K(low_wmark_pages(zone)),
4396 K(high_wmark_pages(zone)),
4397 K(zone_page_state(zone, NR_ACTIVE_ANON)),
4398 K(zone_page_state(zone, NR_INACTIVE_ANON)),
4399 K(zone_page_state(zone, NR_ACTIVE_FILE)),
4400 K(zone_page_state(zone, NR_INACTIVE_FILE)),
4401 K(zone_page_state(zone, NR_UNEVICTABLE)),
4402 K(zone_page_state(zone, NR_ISOLATED_ANON)),
4403 K(zone_page_state(zone, NR_ISOLATED_FILE)),
4404 K(zone->present_pages),
4405 K(zone->managed_pages),
4406 K(zone_page_state(zone, NR_MLOCK)),
4407 K(zone_page_state(zone, NR_FILE_DIRTY)),
4408 K(zone_page_state(zone, NR_WRITEBACK)),
4409 K(zone_page_state(zone, NR_FILE_MAPPED)),
4410 K(zone_page_state(zone, NR_SHMEM)),
4411 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4412 K(zone_page_state(zone, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4413 K(zone_page_state(zone, NR_SHMEM_PMDMAPPED)
4415 K(zone_page_state(zone, NR_ANON_THPS) * HPAGE_PMD_NR),
4417 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4418 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4419 zone_page_state(zone, NR_KERNEL_STACK) *
4421 K(zone_page_state(zone, NR_PAGETABLE)),
4422 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4423 K(zone_page_state(zone, NR_BOUNCE)),
4425 K(this_cpu_read(zone->pageset->pcp.count)),
4426 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4427 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4428 K(zone_page_state(zone, NR_PAGES_SCANNED)),
4429 (!zone_reclaimable(zone) ? "yes" : "no")
4431 printk("lowmem_reserve[]:");
4432 for (i = 0; i < MAX_NR_ZONES; i++)
4433 printk(" %ld", zone->lowmem_reserve[i]);
4437 for_each_populated_zone(zone) {
4439 unsigned long nr[MAX_ORDER], flags, total = 0;
4440 unsigned char types[MAX_ORDER];
4442 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4445 printk("%s: ", zone->name);
4447 spin_lock_irqsave(&zone->lock, flags);
4448 for (order = 0; order < MAX_ORDER; order++) {
4449 struct free_area *area = &zone->free_area[order];
4452 nr[order] = area->nr_free;
4453 total += nr[order] << order;
4456 for (type = 0; type < MIGRATE_TYPES; type++) {
4457 if (!list_empty(&area->free_list[type]))
4458 types[order] |= 1 << type;
4461 spin_unlock_irqrestore(&zone->lock, flags);
4462 for (order = 0; order < MAX_ORDER; order++) {
4463 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4465 show_migration_types(types[order]);
4467 printk("= %lukB\n", K(total));
4470 hugetlb_show_meminfo();
4472 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4474 show_swap_cache_info();
4477 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4479 zoneref->zone = zone;
4480 zoneref->zone_idx = zone_idx(zone);
4484 * Builds allocation fallback zone lists.
4486 * Add all populated zones of a node to the zonelist.
4488 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4492 enum zone_type zone_type = MAX_NR_ZONES;
4496 zone = pgdat->node_zones + zone_type;
4497 if (populated_zone(zone)) {
4498 zoneref_set_zone(zone,
4499 &zonelist->_zonerefs[nr_zones++]);
4500 check_highest_zone(zone_type);
4502 } while (zone_type);
4510 * 0 = automatic detection of better ordering.
4511 * 1 = order by ([node] distance, -zonetype)
4512 * 2 = order by (-zonetype, [node] distance)
4514 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4515 * the same zonelist. So only NUMA can configure this param.
4517 #define ZONELIST_ORDER_DEFAULT 0
4518 #define ZONELIST_ORDER_NODE 1
4519 #define ZONELIST_ORDER_ZONE 2
4521 /* zonelist order in the kernel.
4522 * set_zonelist_order() will set this to NODE or ZONE.
4524 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4525 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4529 /* The value user specified ....changed by config */
4530 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4531 /* string for sysctl */
4532 #define NUMA_ZONELIST_ORDER_LEN 16
4533 char numa_zonelist_order[16] = "default";
4536 * interface for configure zonelist ordering.
4537 * command line option "numa_zonelist_order"
4538 * = "[dD]efault - default, automatic configuration.
4539 * = "[nN]ode - order by node locality, then by zone within node
4540 * = "[zZ]one - order by zone, then by locality within zone
4543 static int __parse_numa_zonelist_order(char *s)
4545 if (*s == 'd' || *s == 'D') {
4546 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4547 } else if (*s == 'n' || *s == 'N') {
4548 user_zonelist_order = ZONELIST_ORDER_NODE;
4549 } else if (*s == 'z' || *s == 'Z') {
4550 user_zonelist_order = ZONELIST_ORDER_ZONE;
4552 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4558 static __init int setup_numa_zonelist_order(char *s)
4565 ret = __parse_numa_zonelist_order(s);
4567 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4571 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4574 * sysctl handler for numa_zonelist_order
4576 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4577 void __user *buffer, size_t *length,
4580 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4582 static DEFINE_MUTEX(zl_order_mutex);
4584 mutex_lock(&zl_order_mutex);
4586 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4590 strcpy(saved_string, (char *)table->data);
4592 ret = proc_dostring(table, write, buffer, length, ppos);
4596 int oldval = user_zonelist_order;
4598 ret = __parse_numa_zonelist_order((char *)table->data);
4601 * bogus value. restore saved string
4603 strncpy((char *)table->data, saved_string,
4604 NUMA_ZONELIST_ORDER_LEN);
4605 user_zonelist_order = oldval;
4606 } else if (oldval != user_zonelist_order) {
4607 mutex_lock(&zonelists_mutex);
4608 build_all_zonelists(NULL, NULL);
4609 mutex_unlock(&zonelists_mutex);
4613 mutex_unlock(&zl_order_mutex);
4618 #define MAX_NODE_LOAD (nr_online_nodes)
4619 static int node_load[MAX_NUMNODES];
4622 * find_next_best_node - find the next node that should appear in a given node's fallback list
4623 * @node: node whose fallback list we're appending
4624 * @used_node_mask: nodemask_t of already used nodes
4626 * We use a number of factors to determine which is the next node that should
4627 * appear on a given node's fallback list. The node should not have appeared
4628 * already in @node's fallback list, and it should be the next closest node
4629 * according to the distance array (which contains arbitrary distance values
4630 * from each node to each node in the system), and should also prefer nodes
4631 * with no CPUs, since presumably they'll have very little allocation pressure
4632 * on them otherwise.
4633 * It returns -1 if no node is found.
4635 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4638 int min_val = INT_MAX;
4639 int best_node = NUMA_NO_NODE;
4640 const struct cpumask *tmp = cpumask_of_node(0);
4642 /* Use the local node if we haven't already */
4643 if (!node_isset(node, *used_node_mask)) {
4644 node_set(node, *used_node_mask);
4648 for_each_node_state(n, N_MEMORY) {
4650 /* Don't want a node to appear more than once */
4651 if (node_isset(n, *used_node_mask))
4654 /* Use the distance array to find the distance */
4655 val = node_distance(node, n);
4657 /* Penalize nodes under us ("prefer the next node") */
4660 /* Give preference to headless and unused nodes */
4661 tmp = cpumask_of_node(n);
4662 if (!cpumask_empty(tmp))
4663 val += PENALTY_FOR_NODE_WITH_CPUS;
4665 /* Slight preference for less loaded node */
4666 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4667 val += node_load[n];
4669 if (val < min_val) {
4676 node_set(best_node, *used_node_mask);
4683 * Build zonelists ordered by node and zones within node.
4684 * This results in maximum locality--normal zone overflows into local
4685 * DMA zone, if any--but risks exhausting DMA zone.
4687 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4690 struct zonelist *zonelist;
4692 zonelist = &pgdat->node_zonelists[0];
4693 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4695 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4696 zonelist->_zonerefs[j].zone = NULL;
4697 zonelist->_zonerefs[j].zone_idx = 0;
4701 * Build gfp_thisnode zonelists
4703 static void build_thisnode_zonelists(pg_data_t *pgdat)
4706 struct zonelist *zonelist;
4708 zonelist = &pgdat->node_zonelists[1];
4709 j = build_zonelists_node(pgdat, zonelist, 0);
4710 zonelist->_zonerefs[j].zone = NULL;
4711 zonelist->_zonerefs[j].zone_idx = 0;
4715 * Build zonelists ordered by zone and nodes within zones.
4716 * This results in conserving DMA zone[s] until all Normal memory is
4717 * exhausted, but results in overflowing to remote node while memory
4718 * may still exist in local DMA zone.
4720 static int node_order[MAX_NUMNODES];
4722 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4725 int zone_type; /* needs to be signed */
4727 struct zonelist *zonelist;
4729 zonelist = &pgdat->node_zonelists[0];
4731 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4732 for (j = 0; j < nr_nodes; j++) {
4733 node = node_order[j];
4734 z = &NODE_DATA(node)->node_zones[zone_type];
4735 if (populated_zone(z)) {
4737 &zonelist->_zonerefs[pos++]);
4738 check_highest_zone(zone_type);
4742 zonelist->_zonerefs[pos].zone = NULL;
4743 zonelist->_zonerefs[pos].zone_idx = 0;
4746 #if defined(CONFIG_64BIT)
4748 * Devices that require DMA32/DMA are relatively rare and do not justify a
4749 * penalty to every machine in case the specialised case applies. Default
4750 * to Node-ordering on 64-bit NUMA machines
4752 static int default_zonelist_order(void)
4754 return ZONELIST_ORDER_NODE;
4758 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4759 * by the kernel. If processes running on node 0 deplete the low memory zone
4760 * then reclaim will occur more frequency increasing stalls and potentially
4761 * be easier to OOM if a large percentage of the zone is under writeback or
4762 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4763 * Hence, default to zone ordering on 32-bit.
4765 static int default_zonelist_order(void)
4767 return ZONELIST_ORDER_ZONE;
4769 #endif /* CONFIG_64BIT */
4771 static void set_zonelist_order(void)
4773 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4774 current_zonelist_order = default_zonelist_order();
4776 current_zonelist_order = user_zonelist_order;
4779 static void build_zonelists(pg_data_t *pgdat)
4782 nodemask_t used_mask;
4783 int local_node, prev_node;
4784 struct zonelist *zonelist;
4785 unsigned int order = current_zonelist_order;
4787 /* initialize zonelists */
4788 for (i = 0; i < MAX_ZONELISTS; i++) {
4789 zonelist = pgdat->node_zonelists + i;
4790 zonelist->_zonerefs[0].zone = NULL;
4791 zonelist->_zonerefs[0].zone_idx = 0;
4794 /* NUMA-aware ordering of nodes */
4795 local_node = pgdat->node_id;
4796 load = nr_online_nodes;
4797 prev_node = local_node;
4798 nodes_clear(used_mask);
4800 memset(node_order, 0, sizeof(node_order));
4803 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4805 * We don't want to pressure a particular node.
4806 * So adding penalty to the first node in same
4807 * distance group to make it round-robin.
4809 if (node_distance(local_node, node) !=
4810 node_distance(local_node, prev_node))
4811 node_load[node] = load;
4815 if (order == ZONELIST_ORDER_NODE)
4816 build_zonelists_in_node_order(pgdat, node);
4818 node_order[i++] = node; /* remember order */
4821 if (order == ZONELIST_ORDER_ZONE) {
4822 /* calculate node order -- i.e., DMA last! */
4823 build_zonelists_in_zone_order(pgdat, i);
4826 build_thisnode_zonelists(pgdat);
4829 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4831 * Return node id of node used for "local" allocations.
4832 * I.e., first node id of first zone in arg node's generic zonelist.
4833 * Used for initializing percpu 'numa_mem', which is used primarily
4834 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4836 int local_memory_node(int node)
4840 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4841 gfp_zone(GFP_KERNEL),
4843 return z->zone->node;
4847 #else /* CONFIG_NUMA */
4849 static void set_zonelist_order(void)
4851 current_zonelist_order = ZONELIST_ORDER_ZONE;
4854 static void build_zonelists(pg_data_t *pgdat)
4856 int node, local_node;
4858 struct zonelist *zonelist;
4860 local_node = pgdat->node_id;
4862 zonelist = &pgdat->node_zonelists[0];
4863 j = build_zonelists_node(pgdat, zonelist, 0);
4866 * Now we build the zonelist so that it contains the zones
4867 * of all the other nodes.
4868 * We don't want to pressure a particular node, so when
4869 * building the zones for node N, we make sure that the
4870 * zones coming right after the local ones are those from
4871 * node N+1 (modulo N)
4873 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4874 if (!node_online(node))
4876 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4878 for (node = 0; node < local_node; node++) {
4879 if (!node_online(node))
4881 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4884 zonelist->_zonerefs[j].zone = NULL;
4885 zonelist->_zonerefs[j].zone_idx = 0;
4888 #endif /* CONFIG_NUMA */
4891 * Boot pageset table. One per cpu which is going to be used for all
4892 * zones and all nodes. The parameters will be set in such a way
4893 * that an item put on a list will immediately be handed over to
4894 * the buddy list. This is safe since pageset manipulation is done
4895 * with interrupts disabled.
4897 * The boot_pagesets must be kept even after bootup is complete for
4898 * unused processors and/or zones. They do play a role for bootstrapping
4899 * hotplugged processors.
4901 * zoneinfo_show() and maybe other functions do
4902 * not check if the processor is online before following the pageset pointer.
4903 * Other parts of the kernel may not check if the zone is available.
4905 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4906 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4907 static void setup_zone_pageset(struct zone *zone);
4910 * Global mutex to protect against size modification of zonelists
4911 * as well as to serialize pageset setup for the new populated zone.
4913 DEFINE_MUTEX(zonelists_mutex);
4915 /* return values int ....just for stop_machine() */
4916 static int __build_all_zonelists(void *data)
4920 pg_data_t *self = data;
4923 memset(node_load, 0, sizeof(node_load));
4926 if (self && !node_online(self->node_id)) {
4927 build_zonelists(self);
4930 for_each_online_node(nid) {
4931 pg_data_t *pgdat = NODE_DATA(nid);
4933 build_zonelists(pgdat);
4937 * Initialize the boot_pagesets that are going to be used
4938 * for bootstrapping processors. The real pagesets for
4939 * each zone will be allocated later when the per cpu
4940 * allocator is available.
4942 * boot_pagesets are used also for bootstrapping offline
4943 * cpus if the system is already booted because the pagesets
4944 * are needed to initialize allocators on a specific cpu too.
4945 * F.e. the percpu allocator needs the page allocator which
4946 * needs the percpu allocator in order to allocate its pagesets
4947 * (a chicken-egg dilemma).
4949 for_each_possible_cpu(cpu) {
4950 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4952 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4954 * We now know the "local memory node" for each node--
4955 * i.e., the node of the first zone in the generic zonelist.
4956 * Set up numa_mem percpu variable for on-line cpus. During
4957 * boot, only the boot cpu should be on-line; we'll init the
4958 * secondary cpus' numa_mem as they come on-line. During
4959 * node/memory hotplug, we'll fixup all on-line cpus.
4961 if (cpu_online(cpu))
4962 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4969 static noinline void __init
4970 build_all_zonelists_init(void)
4972 __build_all_zonelists(NULL);
4973 mminit_verify_zonelist();
4974 cpuset_init_current_mems_allowed();
4978 * Called with zonelists_mutex held always
4979 * unless system_state == SYSTEM_BOOTING.
4981 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4982 * [we're only called with non-NULL zone through __meminit paths] and
4983 * (2) call of __init annotated helper build_all_zonelists_init
4984 * [protected by SYSTEM_BOOTING].
4986 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4988 set_zonelist_order();
4990 if (system_state == SYSTEM_BOOTING) {
4991 build_all_zonelists_init();
4993 #ifdef CONFIG_MEMORY_HOTPLUG
4995 setup_zone_pageset(zone);
4997 /* we have to stop all cpus to guarantee there is no user
4999 stop_machine(__build_all_zonelists, pgdat, NULL);
5000 /* cpuset refresh routine should be here */
5002 vm_total_pages = nr_free_pagecache_pages();
5004 * Disable grouping by mobility if the number of pages in the
5005 * system is too low to allow the mechanism to work. It would be
5006 * more accurate, but expensive to check per-zone. This check is
5007 * made on memory-hotadd so a system can start with mobility
5008 * disabled and enable it later
5010 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5011 page_group_by_mobility_disabled = 1;
5013 page_group_by_mobility_disabled = 0;
5015 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5017 zonelist_order_name[current_zonelist_order],
5018 page_group_by_mobility_disabled ? "off" : "on",
5021 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5026 * Helper functions to size the waitqueue hash table.
5027 * Essentially these want to choose hash table sizes sufficiently
5028 * large so that collisions trying to wait on pages are rare.
5029 * But in fact, the number of active page waitqueues on typical
5030 * systems is ridiculously low, less than 200. So this is even
5031 * conservative, even though it seems large.
5033 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5034 * waitqueues, i.e. the size of the waitq table given the number of pages.
5036 #define PAGES_PER_WAITQUEUE 256
5038 #ifndef CONFIG_MEMORY_HOTPLUG
5039 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5041 unsigned long size = 1;
5043 pages /= PAGES_PER_WAITQUEUE;
5045 while (size < pages)
5049 * Once we have dozens or even hundreds of threads sleeping
5050 * on IO we've got bigger problems than wait queue collision.
5051 * Limit the size of the wait table to a reasonable size.
5053 size = min(size, 4096UL);
5055 return max(size, 4UL);
5059 * A zone's size might be changed by hot-add, so it is not possible to determine
5060 * a suitable size for its wait_table. So we use the maximum size now.
5062 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5064 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5065 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5066 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5068 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5069 * or more by the traditional way. (See above). It equals:
5071 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5072 * ia64(16K page size) : = ( 8G + 4M)byte.
5073 * powerpc (64K page size) : = (32G +16M)byte.
5075 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5082 * This is an integer logarithm so that shifts can be used later
5083 * to extract the more random high bits from the multiplicative
5084 * hash function before the remainder is taken.
5086 static inline unsigned long wait_table_bits(unsigned long size)
5092 * Initially all pages are reserved - free ones are freed
5093 * up by free_all_bootmem() once the early boot process is
5094 * done. Non-atomic initialization, single-pass.
5096 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5097 unsigned long start_pfn, enum memmap_context context)
5099 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5100 unsigned long end_pfn = start_pfn + size;
5101 pg_data_t *pgdat = NODE_DATA(nid);
5103 unsigned long nr_initialised = 0;
5104 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5105 struct memblock_region *r = NULL, *tmp;
5108 if (highest_memmap_pfn < end_pfn - 1)
5109 highest_memmap_pfn = end_pfn - 1;
5112 * Honor reservation requested by the driver for this ZONE_DEVICE
5115 if (altmap && start_pfn == altmap->base_pfn)
5116 start_pfn += altmap->reserve;
5118 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5120 * There can be holes in boot-time mem_map[]s handed to this
5121 * function. They do not exist on hotplugged memory.
5123 if (context != MEMMAP_EARLY)
5126 if (!early_pfn_valid(pfn))
5128 if (!early_pfn_in_nid(pfn, nid))
5130 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5133 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5135 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5136 * from zone_movable_pfn[nid] to end of each node should be
5137 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5139 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5140 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5144 * Check given memblock attribute by firmware which can affect
5145 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5146 * mirrored, it's an overlapped memmap init. skip it.
5148 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5149 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5150 for_each_memblock(memory, tmp)
5151 if (pfn < memblock_region_memory_end_pfn(tmp))
5155 if (pfn >= memblock_region_memory_base_pfn(r) &&
5156 memblock_is_mirror(r)) {
5157 /* already initialized as NORMAL */
5158 pfn = memblock_region_memory_end_pfn(r);
5166 * Mark the block movable so that blocks are reserved for
5167 * movable at startup. This will force kernel allocations
5168 * to reserve their blocks rather than leaking throughout
5169 * the address space during boot when many long-lived
5170 * kernel allocations are made.
5172 * bitmap is created for zone's valid pfn range. but memmap
5173 * can be created for invalid pages (for alignment)
5174 * check here not to call set_pageblock_migratetype() against
5177 if (!(pfn & (pageblock_nr_pages - 1))) {
5178 struct page *page = pfn_to_page(pfn);
5180 __init_single_page(page, pfn, zone, nid);
5181 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5183 __init_single_pfn(pfn, zone, nid);
5188 static void __meminit zone_init_free_lists(struct zone *zone)
5190 unsigned int order, t;
5191 for_each_migratetype_order(order, t) {
5192 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5193 zone->free_area[order].nr_free = 0;
5197 #ifndef __HAVE_ARCH_MEMMAP_INIT
5198 #define memmap_init(size, nid, zone, start_pfn) \
5199 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5202 static int zone_batchsize(struct zone *zone)
5208 * The per-cpu-pages pools are set to around 1000th of the
5209 * size of the zone. But no more than 1/2 of a meg.
5211 * OK, so we don't know how big the cache is. So guess.
5213 batch = zone->managed_pages / 1024;
5214 if (batch * PAGE_SIZE > 512 * 1024)
5215 batch = (512 * 1024) / PAGE_SIZE;
5216 batch /= 4; /* We effectively *= 4 below */
5221 * Clamp the batch to a 2^n - 1 value. Having a power
5222 * of 2 value was found to be more likely to have
5223 * suboptimal cache aliasing properties in some cases.
5225 * For example if 2 tasks are alternately allocating
5226 * batches of pages, one task can end up with a lot
5227 * of pages of one half of the possible page colors
5228 * and the other with pages of the other colors.
5230 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5235 /* The deferral and batching of frees should be suppressed under NOMMU
5238 * The problem is that NOMMU needs to be able to allocate large chunks
5239 * of contiguous memory as there's no hardware page translation to
5240 * assemble apparent contiguous memory from discontiguous pages.
5242 * Queueing large contiguous runs of pages for batching, however,
5243 * causes the pages to actually be freed in smaller chunks. As there
5244 * can be a significant delay between the individual batches being
5245 * recycled, this leads to the once large chunks of space being
5246 * fragmented and becoming unavailable for high-order allocations.
5253 * pcp->high and pcp->batch values are related and dependent on one another:
5254 * ->batch must never be higher then ->high.
5255 * The following function updates them in a safe manner without read side
5258 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5259 * those fields changing asynchronously (acording the the above rule).
5261 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5262 * outside of boot time (or some other assurance that no concurrent updaters
5265 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5266 unsigned long batch)
5268 /* start with a fail safe value for batch */
5272 /* Update high, then batch, in order */
5279 /* a companion to pageset_set_high() */
5280 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5282 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5285 static void pageset_init(struct per_cpu_pageset *p)
5287 struct per_cpu_pages *pcp;
5290 memset(p, 0, sizeof(*p));
5294 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5295 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5298 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5301 pageset_set_batch(p, batch);
5305 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5306 * to the value high for the pageset p.
5308 static void pageset_set_high(struct per_cpu_pageset *p,
5311 unsigned long batch = max(1UL, high / 4);
5312 if ((high / 4) > (PAGE_SHIFT * 8))
5313 batch = PAGE_SHIFT * 8;
5315 pageset_update(&p->pcp, high, batch);
5318 static void pageset_set_high_and_batch(struct zone *zone,
5319 struct per_cpu_pageset *pcp)
5321 if (percpu_pagelist_fraction)
5322 pageset_set_high(pcp,
5323 (zone->managed_pages /
5324 percpu_pagelist_fraction));
5326 pageset_set_batch(pcp, zone_batchsize(zone));
5329 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5331 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5334 pageset_set_high_and_batch(zone, pcp);
5337 static void __meminit setup_zone_pageset(struct zone *zone)
5340 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5341 for_each_possible_cpu(cpu)
5342 zone_pageset_init(zone, cpu);
5346 * Allocate per cpu pagesets and initialize them.
5347 * Before this call only boot pagesets were available.
5349 void __init setup_per_cpu_pageset(void)
5353 for_each_populated_zone(zone)
5354 setup_zone_pageset(zone);
5357 static noinline __init_refok
5358 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5364 * The per-page waitqueue mechanism uses hashed waitqueues
5367 zone->wait_table_hash_nr_entries =
5368 wait_table_hash_nr_entries(zone_size_pages);
5369 zone->wait_table_bits =
5370 wait_table_bits(zone->wait_table_hash_nr_entries);
5371 alloc_size = zone->wait_table_hash_nr_entries
5372 * sizeof(wait_queue_head_t);
5374 if (!slab_is_available()) {
5375 zone->wait_table = (wait_queue_head_t *)
5376 memblock_virt_alloc_node_nopanic(
5377 alloc_size, zone->zone_pgdat->node_id);
5380 * This case means that a zone whose size was 0 gets new memory
5381 * via memory hot-add.
5382 * But it may be the case that a new node was hot-added. In
5383 * this case vmalloc() will not be able to use this new node's
5384 * memory - this wait_table must be initialized to use this new
5385 * node itself as well.
5386 * To use this new node's memory, further consideration will be
5389 zone->wait_table = vmalloc(alloc_size);
5391 if (!zone->wait_table)
5394 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5395 init_waitqueue_head(zone->wait_table + i);
5400 static __meminit void zone_pcp_init(struct zone *zone)
5403 * per cpu subsystem is not up at this point. The following code
5404 * relies on the ability of the linker to provide the
5405 * offset of a (static) per cpu variable into the per cpu area.
5407 zone->pageset = &boot_pageset;
5409 if (populated_zone(zone))
5410 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5411 zone->name, zone->present_pages,
5412 zone_batchsize(zone));
5415 int __meminit init_currently_empty_zone(struct zone *zone,
5416 unsigned long zone_start_pfn,
5419 struct pglist_data *pgdat = zone->zone_pgdat;
5421 ret = zone_wait_table_init(zone, size);
5424 pgdat->nr_zones = zone_idx(zone) + 1;
5426 zone->zone_start_pfn = zone_start_pfn;
5428 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5429 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5431 (unsigned long)zone_idx(zone),
5432 zone_start_pfn, (zone_start_pfn + size));
5434 zone_init_free_lists(zone);
5439 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5440 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5443 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5445 int __meminit __early_pfn_to_nid(unsigned long pfn,
5446 struct mminit_pfnnid_cache *state)
5448 unsigned long start_pfn, end_pfn;
5451 if (state->last_start <= pfn && pfn < state->last_end)
5452 return state->last_nid;
5454 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5456 state->last_start = start_pfn;
5457 state->last_end = end_pfn;
5458 state->last_nid = nid;
5463 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5466 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5467 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5468 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5470 * If an architecture guarantees that all ranges registered contain no holes
5471 * and may be freed, this this function may be used instead of calling
5472 * memblock_free_early_nid() manually.
5474 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5476 unsigned long start_pfn, end_pfn;
5479 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5480 start_pfn = min(start_pfn, max_low_pfn);
5481 end_pfn = min(end_pfn, max_low_pfn);
5483 if (start_pfn < end_pfn)
5484 memblock_free_early_nid(PFN_PHYS(start_pfn),
5485 (end_pfn - start_pfn) << PAGE_SHIFT,
5491 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5492 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5494 * If an architecture guarantees that all ranges registered contain no holes and may
5495 * be freed, this function may be used instead of calling memory_present() manually.
5497 void __init sparse_memory_present_with_active_regions(int nid)
5499 unsigned long start_pfn, end_pfn;
5502 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5503 memory_present(this_nid, start_pfn, end_pfn);
5507 * get_pfn_range_for_nid - Return the start and end page frames for a node
5508 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5509 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5510 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5512 * It returns the start and end page frame of a node based on information
5513 * provided by memblock_set_node(). If called for a node
5514 * with no available memory, a warning is printed and the start and end
5517 void __meminit get_pfn_range_for_nid(unsigned int nid,
5518 unsigned long *start_pfn, unsigned long *end_pfn)
5520 unsigned long this_start_pfn, this_end_pfn;
5526 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5527 *start_pfn = min(*start_pfn, this_start_pfn);
5528 *end_pfn = max(*end_pfn, this_end_pfn);
5531 if (*start_pfn == -1UL)
5536 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5537 * assumption is made that zones within a node are ordered in monotonic
5538 * increasing memory addresses so that the "highest" populated zone is used
5540 static void __init find_usable_zone_for_movable(void)
5543 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5544 if (zone_index == ZONE_MOVABLE)
5547 if (arch_zone_highest_possible_pfn[zone_index] >
5548 arch_zone_lowest_possible_pfn[zone_index])
5552 VM_BUG_ON(zone_index == -1);
5553 movable_zone = zone_index;
5557 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5558 * because it is sized independent of architecture. Unlike the other zones,
5559 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5560 * in each node depending on the size of each node and how evenly kernelcore
5561 * is distributed. This helper function adjusts the zone ranges
5562 * provided by the architecture for a given node by using the end of the
5563 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5564 * zones within a node are in order of monotonic increases memory addresses
5566 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5567 unsigned long zone_type,
5568 unsigned long node_start_pfn,
5569 unsigned long node_end_pfn,
5570 unsigned long *zone_start_pfn,
5571 unsigned long *zone_end_pfn)
5573 /* Only adjust if ZONE_MOVABLE is on this node */
5574 if (zone_movable_pfn[nid]) {
5575 /* Size ZONE_MOVABLE */
5576 if (zone_type == ZONE_MOVABLE) {
5577 *zone_start_pfn = zone_movable_pfn[nid];
5578 *zone_end_pfn = min(node_end_pfn,
5579 arch_zone_highest_possible_pfn[movable_zone]);
5581 /* Check if this whole range is within ZONE_MOVABLE */
5582 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5583 *zone_start_pfn = *zone_end_pfn;
5588 * Return the number of pages a zone spans in a node, including holes
5589 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5591 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5592 unsigned long zone_type,
5593 unsigned long node_start_pfn,
5594 unsigned long node_end_pfn,
5595 unsigned long *zone_start_pfn,
5596 unsigned long *zone_end_pfn,
5597 unsigned long *ignored)
5599 /* When hotadd a new node from cpu_up(), the node should be empty */
5600 if (!node_start_pfn && !node_end_pfn)
5603 /* Get the start and end of the zone */
5604 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5605 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5606 adjust_zone_range_for_zone_movable(nid, zone_type,
5607 node_start_pfn, node_end_pfn,
5608 zone_start_pfn, zone_end_pfn);
5610 /* Check that this node has pages within the zone's required range */
5611 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5614 /* Move the zone boundaries inside the node if necessary */
5615 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5616 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5618 /* Return the spanned pages */
5619 return *zone_end_pfn - *zone_start_pfn;
5623 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5624 * then all holes in the requested range will be accounted for.
5626 unsigned long __meminit __absent_pages_in_range(int nid,
5627 unsigned long range_start_pfn,
5628 unsigned long range_end_pfn)
5630 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5631 unsigned long start_pfn, end_pfn;
5634 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5635 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5636 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5637 nr_absent -= end_pfn - start_pfn;
5643 * absent_pages_in_range - Return number of page frames in holes within a range
5644 * @start_pfn: The start PFN to start searching for holes
5645 * @end_pfn: The end PFN to stop searching for holes
5647 * It returns the number of pages frames in memory holes within a range.
5649 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5650 unsigned long end_pfn)
5652 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5655 /* Return the number of page frames in holes in a zone on a node */
5656 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5657 unsigned long zone_type,
5658 unsigned long node_start_pfn,
5659 unsigned long node_end_pfn,
5660 unsigned long *ignored)
5662 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5663 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5664 unsigned long zone_start_pfn, zone_end_pfn;
5665 unsigned long nr_absent;
5667 /* When hotadd a new node from cpu_up(), the node should be empty */
5668 if (!node_start_pfn && !node_end_pfn)
5671 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5672 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5674 adjust_zone_range_for_zone_movable(nid, zone_type,
5675 node_start_pfn, node_end_pfn,
5676 &zone_start_pfn, &zone_end_pfn);
5677 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5680 * ZONE_MOVABLE handling.
5681 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5684 if (zone_movable_pfn[nid]) {
5685 if (mirrored_kernelcore) {
5686 unsigned long start_pfn, end_pfn;
5687 struct memblock_region *r;
5689 for_each_memblock(memory, r) {
5690 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5691 zone_start_pfn, zone_end_pfn);
5692 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5693 zone_start_pfn, zone_end_pfn);
5695 if (zone_type == ZONE_MOVABLE &&
5696 memblock_is_mirror(r))
5697 nr_absent += end_pfn - start_pfn;
5699 if (zone_type == ZONE_NORMAL &&
5700 !memblock_is_mirror(r))
5701 nr_absent += end_pfn - start_pfn;
5704 if (zone_type == ZONE_NORMAL)
5705 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5712 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5713 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5714 unsigned long zone_type,
5715 unsigned long node_start_pfn,
5716 unsigned long node_end_pfn,
5717 unsigned long *zone_start_pfn,
5718 unsigned long *zone_end_pfn,
5719 unsigned long *zones_size)
5723 *zone_start_pfn = node_start_pfn;
5724 for (zone = 0; zone < zone_type; zone++)
5725 *zone_start_pfn += zones_size[zone];
5727 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5729 return zones_size[zone_type];
5732 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5733 unsigned long zone_type,
5734 unsigned long node_start_pfn,
5735 unsigned long node_end_pfn,
5736 unsigned long *zholes_size)
5741 return zholes_size[zone_type];
5744 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5746 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5747 unsigned long node_start_pfn,
5748 unsigned long node_end_pfn,
5749 unsigned long *zones_size,
5750 unsigned long *zholes_size)
5752 unsigned long realtotalpages = 0, totalpages = 0;
5755 for (i = 0; i < MAX_NR_ZONES; i++) {
5756 struct zone *zone = pgdat->node_zones + i;
5757 unsigned long zone_start_pfn, zone_end_pfn;
5758 unsigned long size, real_size;
5760 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5766 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5767 node_start_pfn, node_end_pfn,
5770 zone->zone_start_pfn = zone_start_pfn;
5772 zone->zone_start_pfn = 0;
5773 zone->spanned_pages = size;
5774 zone->present_pages = real_size;
5777 realtotalpages += real_size;
5780 pgdat->node_spanned_pages = totalpages;
5781 pgdat->node_present_pages = realtotalpages;
5782 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5786 #ifndef CONFIG_SPARSEMEM
5788 * Calculate the size of the zone->blockflags rounded to an unsigned long
5789 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5790 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5791 * round what is now in bits to nearest long in bits, then return it in
5794 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5796 unsigned long usemapsize;
5798 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5799 usemapsize = roundup(zonesize, pageblock_nr_pages);
5800 usemapsize = usemapsize >> pageblock_order;
5801 usemapsize *= NR_PAGEBLOCK_BITS;
5802 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5804 return usemapsize / 8;
5807 static void __init setup_usemap(struct pglist_data *pgdat,
5809 unsigned long zone_start_pfn,
5810 unsigned long zonesize)
5812 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5813 zone->pageblock_flags = NULL;
5815 zone->pageblock_flags =
5816 memblock_virt_alloc_node_nopanic(usemapsize,
5820 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5821 unsigned long zone_start_pfn, unsigned long zonesize) {}
5822 #endif /* CONFIG_SPARSEMEM */
5824 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5826 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5827 void __paginginit set_pageblock_order(void)
5831 /* Check that pageblock_nr_pages has not already been setup */
5832 if (pageblock_order)
5835 if (HPAGE_SHIFT > PAGE_SHIFT)
5836 order = HUGETLB_PAGE_ORDER;
5838 order = MAX_ORDER - 1;
5841 * Assume the largest contiguous order of interest is a huge page.
5842 * This value may be variable depending on boot parameters on IA64 and
5845 pageblock_order = order;
5847 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5850 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5851 * is unused as pageblock_order is set at compile-time. See
5852 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5855 void __paginginit set_pageblock_order(void)
5859 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5861 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5862 unsigned long present_pages)
5864 unsigned long pages = spanned_pages;
5867 * Provide a more accurate estimation if there are holes within
5868 * the zone and SPARSEMEM is in use. If there are holes within the
5869 * zone, each populated memory region may cost us one or two extra
5870 * memmap pages due to alignment because memmap pages for each
5871 * populated regions may not naturally algined on page boundary.
5872 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5874 if (spanned_pages > present_pages + (present_pages >> 4) &&
5875 IS_ENABLED(CONFIG_SPARSEMEM))
5876 pages = present_pages;
5878 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5882 * Set up the zone data structures:
5883 * - mark all pages reserved
5884 * - mark all memory queues empty
5885 * - clear the memory bitmaps
5887 * NOTE: pgdat should get zeroed by caller.
5889 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5892 int nid = pgdat->node_id;
5895 pgdat_resize_init(pgdat);
5896 #ifdef CONFIG_NUMA_BALANCING
5897 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5898 pgdat->numabalancing_migrate_nr_pages = 0;
5899 pgdat->numabalancing_migrate_next_window = jiffies;
5901 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5902 spin_lock_init(&pgdat->split_queue_lock);
5903 INIT_LIST_HEAD(&pgdat->split_queue);
5904 pgdat->split_queue_len = 0;
5906 init_waitqueue_head(&pgdat->kswapd_wait);
5907 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5908 #ifdef CONFIG_COMPACTION
5909 init_waitqueue_head(&pgdat->kcompactd_wait);
5911 pgdat_page_ext_init(pgdat);
5913 for (j = 0; j < MAX_NR_ZONES; j++) {
5914 struct zone *zone = pgdat->node_zones + j;
5915 unsigned long size, realsize, freesize, memmap_pages;
5916 unsigned long zone_start_pfn = zone->zone_start_pfn;
5918 size = zone->spanned_pages;
5919 realsize = freesize = zone->present_pages;
5922 * Adjust freesize so that it accounts for how much memory
5923 * is used by this zone for memmap. This affects the watermark
5924 * and per-cpu initialisations
5926 memmap_pages = calc_memmap_size(size, realsize);
5927 if (!is_highmem_idx(j)) {
5928 if (freesize >= memmap_pages) {
5929 freesize -= memmap_pages;
5932 " %s zone: %lu pages used for memmap\n",
5933 zone_names[j], memmap_pages);
5935 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5936 zone_names[j], memmap_pages, freesize);
5939 /* Account for reserved pages */
5940 if (j == 0 && freesize > dma_reserve) {
5941 freesize -= dma_reserve;
5942 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5943 zone_names[0], dma_reserve);
5946 if (!is_highmem_idx(j))
5947 nr_kernel_pages += freesize;
5948 /* Charge for highmem memmap if there are enough kernel pages */
5949 else if (nr_kernel_pages > memmap_pages * 2)
5950 nr_kernel_pages -= memmap_pages;
5951 nr_all_pages += freesize;
5954 * Set an approximate value for lowmem here, it will be adjusted
5955 * when the bootmem allocator frees pages into the buddy system.
5956 * And all highmem pages will be managed by the buddy system.
5958 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5961 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5963 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5965 zone->name = zone_names[j];
5966 spin_lock_init(&zone->lock);
5967 spin_lock_init(&zone->lru_lock);
5968 zone_seqlock_init(zone);
5969 zone->zone_pgdat = pgdat;
5970 zone_pcp_init(zone);
5972 /* For bootup, initialized properly in watermark setup */
5973 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5975 lruvec_init(&zone->lruvec);
5979 set_pageblock_order();
5980 setup_usemap(pgdat, zone, zone_start_pfn, size);
5981 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5983 memmap_init(size, nid, j, zone_start_pfn);
5987 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5989 unsigned long __maybe_unused start = 0;
5990 unsigned long __maybe_unused offset = 0;
5992 /* Skip empty nodes */
5993 if (!pgdat->node_spanned_pages)
5996 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5997 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5998 offset = pgdat->node_start_pfn - start;
5999 /* ia64 gets its own node_mem_map, before this, without bootmem */
6000 if (!pgdat->node_mem_map) {
6001 unsigned long size, end;
6005 * The zone's endpoints aren't required to be MAX_ORDER
6006 * aligned but the node_mem_map endpoints must be in order
6007 * for the buddy allocator to function correctly.
6009 end = pgdat_end_pfn(pgdat);
6010 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6011 size = (end - start) * sizeof(struct page);
6012 map = alloc_remap(pgdat->node_id, size);
6014 map = memblock_virt_alloc_node_nopanic(size,
6016 pgdat->node_mem_map = map + offset;
6018 #ifndef CONFIG_NEED_MULTIPLE_NODES
6020 * With no DISCONTIG, the global mem_map is just set as node 0's
6022 if (pgdat == NODE_DATA(0)) {
6023 mem_map = NODE_DATA(0)->node_mem_map;
6024 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6025 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6027 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6030 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6033 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6034 unsigned long node_start_pfn, unsigned long *zholes_size)
6036 pg_data_t *pgdat = NODE_DATA(nid);
6037 unsigned long start_pfn = 0;
6038 unsigned long end_pfn = 0;
6040 /* pg_data_t should be reset to zero when it's allocated */
6041 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6043 reset_deferred_meminit(pgdat);
6044 pgdat->node_id = nid;
6045 pgdat->node_start_pfn = node_start_pfn;
6046 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6047 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6048 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6049 (u64)start_pfn << PAGE_SHIFT,
6050 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6052 start_pfn = node_start_pfn;
6054 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6055 zones_size, zholes_size);
6057 alloc_node_mem_map(pgdat);
6058 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6059 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6060 nid, (unsigned long)pgdat,
6061 (unsigned long)pgdat->node_mem_map);
6064 free_area_init_core(pgdat);
6067 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6069 #if MAX_NUMNODES > 1
6071 * Figure out the number of possible node ids.
6073 void __init setup_nr_node_ids(void)
6075 unsigned int highest;
6077 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6078 nr_node_ids = highest + 1;
6083 * node_map_pfn_alignment - determine the maximum internode alignment
6085 * This function should be called after node map is populated and sorted.
6086 * It calculates the maximum power of two alignment which can distinguish
6089 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6090 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6091 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6092 * shifted, 1GiB is enough and this function will indicate so.
6094 * This is used to test whether pfn -> nid mapping of the chosen memory
6095 * model has fine enough granularity to avoid incorrect mapping for the
6096 * populated node map.
6098 * Returns the determined alignment in pfn's. 0 if there is no alignment
6099 * requirement (single node).
6101 unsigned long __init node_map_pfn_alignment(void)
6103 unsigned long accl_mask = 0, last_end = 0;
6104 unsigned long start, end, mask;
6108 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6109 if (!start || last_nid < 0 || last_nid == nid) {
6116 * Start with a mask granular enough to pin-point to the
6117 * start pfn and tick off bits one-by-one until it becomes
6118 * too coarse to separate the current node from the last.
6120 mask = ~((1 << __ffs(start)) - 1);
6121 while (mask && last_end <= (start & (mask << 1)))
6124 /* accumulate all internode masks */
6128 /* convert mask to number of pages */
6129 return ~accl_mask + 1;
6132 /* Find the lowest pfn for a node */
6133 static unsigned long __init find_min_pfn_for_node(int nid)
6135 unsigned long min_pfn = ULONG_MAX;
6136 unsigned long start_pfn;
6139 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6140 min_pfn = min(min_pfn, start_pfn);
6142 if (min_pfn == ULONG_MAX) {
6143 pr_warn("Could not find start_pfn for node %d\n", nid);
6151 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6153 * It returns the minimum PFN based on information provided via
6154 * memblock_set_node().
6156 unsigned long __init find_min_pfn_with_active_regions(void)
6158 return find_min_pfn_for_node(MAX_NUMNODES);
6162 * early_calculate_totalpages()
6163 * Sum pages in active regions for movable zone.
6164 * Populate N_MEMORY for calculating usable_nodes.
6166 static unsigned long __init early_calculate_totalpages(void)
6168 unsigned long totalpages = 0;
6169 unsigned long start_pfn, end_pfn;
6172 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6173 unsigned long pages = end_pfn - start_pfn;
6175 totalpages += pages;
6177 node_set_state(nid, N_MEMORY);
6183 * Find the PFN the Movable zone begins in each node. Kernel memory
6184 * is spread evenly between nodes as long as the nodes have enough
6185 * memory. When they don't, some nodes will have more kernelcore than
6188 static void __init find_zone_movable_pfns_for_nodes(void)
6191 unsigned long usable_startpfn;
6192 unsigned long kernelcore_node, kernelcore_remaining;
6193 /* save the state before borrow the nodemask */
6194 nodemask_t saved_node_state = node_states[N_MEMORY];
6195 unsigned long totalpages = early_calculate_totalpages();
6196 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6197 struct memblock_region *r;
6199 /* Need to find movable_zone earlier when movable_node is specified. */
6200 find_usable_zone_for_movable();
6203 * If movable_node is specified, ignore kernelcore and movablecore
6206 if (movable_node_is_enabled()) {
6207 for_each_memblock(memory, r) {
6208 if (!memblock_is_hotpluggable(r))
6213 usable_startpfn = PFN_DOWN(r->base);
6214 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6215 min(usable_startpfn, zone_movable_pfn[nid]) :
6223 * If kernelcore=mirror is specified, ignore movablecore option
6225 if (mirrored_kernelcore) {
6226 bool mem_below_4gb_not_mirrored = false;
6228 for_each_memblock(memory, r) {
6229 if (memblock_is_mirror(r))
6234 usable_startpfn = memblock_region_memory_base_pfn(r);
6236 if (usable_startpfn < 0x100000) {
6237 mem_below_4gb_not_mirrored = true;
6241 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6242 min(usable_startpfn, zone_movable_pfn[nid]) :
6246 if (mem_below_4gb_not_mirrored)
6247 pr_warn("This configuration results in unmirrored kernel memory.");
6253 * If movablecore=nn[KMG] was specified, calculate what size of
6254 * kernelcore that corresponds so that memory usable for
6255 * any allocation type is evenly spread. If both kernelcore
6256 * and movablecore are specified, then the value of kernelcore
6257 * will be used for required_kernelcore if it's greater than
6258 * what movablecore would have allowed.
6260 if (required_movablecore) {
6261 unsigned long corepages;
6264 * Round-up so that ZONE_MOVABLE is at least as large as what
6265 * was requested by the user
6267 required_movablecore =
6268 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6269 required_movablecore = min(totalpages, required_movablecore);
6270 corepages = totalpages - required_movablecore;
6272 required_kernelcore = max(required_kernelcore, corepages);
6276 * If kernelcore was not specified or kernelcore size is larger
6277 * than totalpages, there is no ZONE_MOVABLE.
6279 if (!required_kernelcore || required_kernelcore >= totalpages)
6282 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6283 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6286 /* Spread kernelcore memory as evenly as possible throughout nodes */
6287 kernelcore_node = required_kernelcore / usable_nodes;
6288 for_each_node_state(nid, N_MEMORY) {
6289 unsigned long start_pfn, end_pfn;
6292 * Recalculate kernelcore_node if the division per node
6293 * now exceeds what is necessary to satisfy the requested
6294 * amount of memory for the kernel
6296 if (required_kernelcore < kernelcore_node)
6297 kernelcore_node = required_kernelcore / usable_nodes;
6300 * As the map is walked, we track how much memory is usable
6301 * by the kernel using kernelcore_remaining. When it is
6302 * 0, the rest of the node is usable by ZONE_MOVABLE
6304 kernelcore_remaining = kernelcore_node;
6306 /* Go through each range of PFNs within this node */
6307 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6308 unsigned long size_pages;
6310 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6311 if (start_pfn >= end_pfn)
6314 /* Account for what is only usable for kernelcore */
6315 if (start_pfn < usable_startpfn) {
6316 unsigned long kernel_pages;
6317 kernel_pages = min(end_pfn, usable_startpfn)
6320 kernelcore_remaining -= min(kernel_pages,
6321 kernelcore_remaining);
6322 required_kernelcore -= min(kernel_pages,
6323 required_kernelcore);
6325 /* Continue if range is now fully accounted */
6326 if (end_pfn <= usable_startpfn) {
6329 * Push zone_movable_pfn to the end so
6330 * that if we have to rebalance
6331 * kernelcore across nodes, we will
6332 * not double account here
6334 zone_movable_pfn[nid] = end_pfn;
6337 start_pfn = usable_startpfn;
6341 * The usable PFN range for ZONE_MOVABLE is from
6342 * start_pfn->end_pfn. Calculate size_pages as the
6343 * number of pages used as kernelcore
6345 size_pages = end_pfn - start_pfn;
6346 if (size_pages > kernelcore_remaining)
6347 size_pages = kernelcore_remaining;
6348 zone_movable_pfn[nid] = start_pfn + size_pages;
6351 * Some kernelcore has been met, update counts and
6352 * break if the kernelcore for this node has been
6355 required_kernelcore -= min(required_kernelcore,
6357 kernelcore_remaining -= size_pages;
6358 if (!kernelcore_remaining)
6364 * If there is still required_kernelcore, we do another pass with one
6365 * less node in the count. This will push zone_movable_pfn[nid] further
6366 * along on the nodes that still have memory until kernelcore is
6370 if (usable_nodes && required_kernelcore > usable_nodes)
6374 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6375 for (nid = 0; nid < MAX_NUMNODES; nid++)
6376 zone_movable_pfn[nid] =
6377 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6380 /* restore the node_state */
6381 node_states[N_MEMORY] = saved_node_state;
6384 /* Any regular or high memory on that node ? */
6385 static void check_for_memory(pg_data_t *pgdat, int nid)
6387 enum zone_type zone_type;
6389 if (N_MEMORY == N_NORMAL_MEMORY)
6392 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6393 struct zone *zone = &pgdat->node_zones[zone_type];
6394 if (populated_zone(zone)) {
6395 node_set_state(nid, N_HIGH_MEMORY);
6396 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6397 zone_type <= ZONE_NORMAL)
6398 node_set_state(nid, N_NORMAL_MEMORY);
6405 * free_area_init_nodes - Initialise all pg_data_t and zone data
6406 * @max_zone_pfn: an array of max PFNs for each zone
6408 * This will call free_area_init_node() for each active node in the system.
6409 * Using the page ranges provided by memblock_set_node(), the size of each
6410 * zone in each node and their holes is calculated. If the maximum PFN
6411 * between two adjacent zones match, it is assumed that the zone is empty.
6412 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6413 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6414 * starts where the previous one ended. For example, ZONE_DMA32 starts
6415 * at arch_max_dma_pfn.
6417 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6419 unsigned long start_pfn, end_pfn;
6422 /* Record where the zone boundaries are */
6423 memset(arch_zone_lowest_possible_pfn, 0,
6424 sizeof(arch_zone_lowest_possible_pfn));
6425 memset(arch_zone_highest_possible_pfn, 0,
6426 sizeof(arch_zone_highest_possible_pfn));
6428 start_pfn = find_min_pfn_with_active_regions();
6430 for (i = 0; i < MAX_NR_ZONES; i++) {
6431 if (i == ZONE_MOVABLE)
6434 end_pfn = max(max_zone_pfn[i], start_pfn);
6435 arch_zone_lowest_possible_pfn[i] = start_pfn;
6436 arch_zone_highest_possible_pfn[i] = end_pfn;
6438 start_pfn = end_pfn;
6440 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6441 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6443 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6444 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6445 find_zone_movable_pfns_for_nodes();
6447 /* Print out the zone ranges */
6448 pr_info("Zone ranges:\n");
6449 for (i = 0; i < MAX_NR_ZONES; i++) {
6450 if (i == ZONE_MOVABLE)
6452 pr_info(" %-8s ", zone_names[i]);
6453 if (arch_zone_lowest_possible_pfn[i] ==
6454 arch_zone_highest_possible_pfn[i])
6457 pr_cont("[mem %#018Lx-%#018Lx]\n",
6458 (u64)arch_zone_lowest_possible_pfn[i]
6460 ((u64)arch_zone_highest_possible_pfn[i]
6461 << PAGE_SHIFT) - 1);
6464 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6465 pr_info("Movable zone start for each node\n");
6466 for (i = 0; i < MAX_NUMNODES; i++) {
6467 if (zone_movable_pfn[i])
6468 pr_info(" Node %d: %#018Lx\n", i,
6469 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6472 /* Print out the early node map */
6473 pr_info("Early memory node ranges\n");
6474 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6475 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6476 (u64)start_pfn << PAGE_SHIFT,
6477 ((u64)end_pfn << PAGE_SHIFT) - 1);
6479 /* Initialise every node */
6480 mminit_verify_pageflags_layout();
6481 setup_nr_node_ids();
6482 for_each_online_node(nid) {
6483 pg_data_t *pgdat = NODE_DATA(nid);
6484 free_area_init_node(nid, NULL,
6485 find_min_pfn_for_node(nid), NULL);
6487 /* Any memory on that node */
6488 if (pgdat->node_present_pages)
6489 node_set_state(nid, N_MEMORY);
6490 check_for_memory(pgdat, nid);
6494 static int __init cmdline_parse_core(char *p, unsigned long *core)
6496 unsigned long long coremem;
6500 coremem = memparse(p, &p);
6501 *core = coremem >> PAGE_SHIFT;
6503 /* Paranoid check that UL is enough for the coremem value */
6504 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6510 * kernelcore=size sets the amount of memory for use for allocations that
6511 * cannot be reclaimed or migrated.
6513 static int __init cmdline_parse_kernelcore(char *p)
6515 /* parse kernelcore=mirror */
6516 if (parse_option_str(p, "mirror")) {
6517 mirrored_kernelcore = true;
6521 return cmdline_parse_core(p, &required_kernelcore);
6525 * movablecore=size sets the amount of memory for use for allocations that
6526 * can be reclaimed or migrated.
6528 static int __init cmdline_parse_movablecore(char *p)
6530 return cmdline_parse_core(p, &required_movablecore);
6533 early_param("kernelcore", cmdline_parse_kernelcore);
6534 early_param("movablecore", cmdline_parse_movablecore);
6536 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6538 void adjust_managed_page_count(struct page *page, long count)
6540 spin_lock(&managed_page_count_lock);
6541 page_zone(page)->managed_pages += count;
6542 totalram_pages += count;
6543 #ifdef CONFIG_HIGHMEM
6544 if (PageHighMem(page))
6545 totalhigh_pages += count;
6547 spin_unlock(&managed_page_count_lock);
6549 EXPORT_SYMBOL(adjust_managed_page_count);
6551 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6554 unsigned long pages = 0;
6556 start = (void *)PAGE_ALIGN((unsigned long)start);
6557 end = (void *)((unsigned long)end & PAGE_MASK);
6558 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6559 if ((unsigned int)poison <= 0xFF)
6560 memset(pos, poison, PAGE_SIZE);
6561 free_reserved_page(virt_to_page(pos));
6565 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6566 s, pages << (PAGE_SHIFT - 10), start, end);
6570 EXPORT_SYMBOL(free_reserved_area);
6572 #ifdef CONFIG_HIGHMEM
6573 void free_highmem_page(struct page *page)
6575 __free_reserved_page(page);
6577 page_zone(page)->managed_pages++;
6583 void __init mem_init_print_info(const char *str)
6585 unsigned long physpages, codesize, datasize, rosize, bss_size;
6586 unsigned long init_code_size, init_data_size;
6588 physpages = get_num_physpages();
6589 codesize = _etext - _stext;
6590 datasize = _edata - _sdata;
6591 rosize = __end_rodata - __start_rodata;
6592 bss_size = __bss_stop - __bss_start;
6593 init_data_size = __init_end - __init_begin;
6594 init_code_size = _einittext - _sinittext;
6597 * Detect special cases and adjust section sizes accordingly:
6598 * 1) .init.* may be embedded into .data sections
6599 * 2) .init.text.* may be out of [__init_begin, __init_end],
6600 * please refer to arch/tile/kernel/vmlinux.lds.S.
6601 * 3) .rodata.* may be embedded into .text or .data sections.
6603 #define adj_init_size(start, end, size, pos, adj) \
6605 if (start <= pos && pos < end && size > adj) \
6609 adj_init_size(__init_begin, __init_end, init_data_size,
6610 _sinittext, init_code_size);
6611 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6612 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6613 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6614 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6616 #undef adj_init_size
6618 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6619 #ifdef CONFIG_HIGHMEM
6623 nr_free_pages() << (PAGE_SHIFT - 10),
6624 physpages << (PAGE_SHIFT - 10),
6625 codesize >> 10, datasize >> 10, rosize >> 10,
6626 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6627 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6628 totalcma_pages << (PAGE_SHIFT - 10),
6629 #ifdef CONFIG_HIGHMEM
6630 totalhigh_pages << (PAGE_SHIFT - 10),
6632 str ? ", " : "", str ? str : "");
6636 * set_dma_reserve - set the specified number of pages reserved in the first zone
6637 * @new_dma_reserve: The number of pages to mark reserved
6639 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6640 * In the DMA zone, a significant percentage may be consumed by kernel image
6641 * and other unfreeable allocations which can skew the watermarks badly. This
6642 * function may optionally be used to account for unfreeable pages in the
6643 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6644 * smaller per-cpu batchsize.
6646 void __init set_dma_reserve(unsigned long new_dma_reserve)
6648 dma_reserve = new_dma_reserve;
6651 void __init free_area_init(unsigned long *zones_size)
6653 free_area_init_node(0, zones_size,
6654 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6657 static int page_alloc_cpu_notify(struct notifier_block *self,
6658 unsigned long action, void *hcpu)
6660 int cpu = (unsigned long)hcpu;
6662 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6663 lru_add_drain_cpu(cpu);
6667 * Spill the event counters of the dead processor
6668 * into the current processors event counters.
6669 * This artificially elevates the count of the current
6672 vm_events_fold_cpu(cpu);
6675 * Zero the differential counters of the dead processor
6676 * so that the vm statistics are consistent.
6678 * This is only okay since the processor is dead and cannot
6679 * race with what we are doing.
6681 cpu_vm_stats_fold(cpu);
6686 void __init page_alloc_init(void)
6688 hotcpu_notifier(page_alloc_cpu_notify, 0);
6692 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6693 * or min_free_kbytes changes.
6695 static void calculate_totalreserve_pages(void)
6697 struct pglist_data *pgdat;
6698 unsigned long reserve_pages = 0;
6699 enum zone_type i, j;
6701 for_each_online_pgdat(pgdat) {
6702 for (i = 0; i < MAX_NR_ZONES; i++) {
6703 struct zone *zone = pgdat->node_zones + i;
6706 /* Find valid and maximum lowmem_reserve in the zone */
6707 for (j = i; j < MAX_NR_ZONES; j++) {
6708 if (zone->lowmem_reserve[j] > max)
6709 max = zone->lowmem_reserve[j];
6712 /* we treat the high watermark as reserved pages. */
6713 max += high_wmark_pages(zone);
6715 if (max > zone->managed_pages)
6716 max = zone->managed_pages;
6718 zone->totalreserve_pages = max;
6720 reserve_pages += max;
6723 totalreserve_pages = reserve_pages;
6727 * setup_per_zone_lowmem_reserve - called whenever
6728 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6729 * has a correct pages reserved value, so an adequate number of
6730 * pages are left in the zone after a successful __alloc_pages().
6732 static void setup_per_zone_lowmem_reserve(void)
6734 struct pglist_data *pgdat;
6735 enum zone_type j, idx;
6737 for_each_online_pgdat(pgdat) {
6738 for (j = 0; j < MAX_NR_ZONES; j++) {
6739 struct zone *zone = pgdat->node_zones + j;
6740 unsigned long managed_pages = zone->managed_pages;
6742 zone->lowmem_reserve[j] = 0;
6746 struct zone *lower_zone;
6750 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6751 sysctl_lowmem_reserve_ratio[idx] = 1;
6753 lower_zone = pgdat->node_zones + idx;
6754 lower_zone->lowmem_reserve[j] = managed_pages /
6755 sysctl_lowmem_reserve_ratio[idx];
6756 managed_pages += lower_zone->managed_pages;
6761 /* update totalreserve_pages */
6762 calculate_totalreserve_pages();
6765 static void __setup_per_zone_wmarks(void)
6767 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6768 unsigned long lowmem_pages = 0;
6770 unsigned long flags;
6772 /* Calculate total number of !ZONE_HIGHMEM pages */
6773 for_each_zone(zone) {
6774 if (!is_highmem(zone))
6775 lowmem_pages += zone->managed_pages;
6778 for_each_zone(zone) {
6781 spin_lock_irqsave(&zone->lock, flags);
6782 tmp = (u64)pages_min * zone->managed_pages;
6783 do_div(tmp, lowmem_pages);
6784 if (is_highmem(zone)) {
6786 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6787 * need highmem pages, so cap pages_min to a small
6790 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6791 * deltas control asynch page reclaim, and so should
6792 * not be capped for highmem.
6794 unsigned long min_pages;
6796 min_pages = zone->managed_pages / 1024;
6797 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6798 zone->watermark[WMARK_MIN] = min_pages;
6801 * If it's a lowmem zone, reserve a number of pages
6802 * proportionate to the zone's size.
6804 zone->watermark[WMARK_MIN] = tmp;
6808 * Set the kswapd watermarks distance according to the
6809 * scale factor in proportion to available memory, but
6810 * ensure a minimum size on small systems.
6812 tmp = max_t(u64, tmp >> 2,
6813 mult_frac(zone->managed_pages,
6814 watermark_scale_factor, 10000));
6816 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6817 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6819 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6820 high_wmark_pages(zone) - low_wmark_pages(zone) -
6821 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6823 spin_unlock_irqrestore(&zone->lock, flags);
6826 /* update totalreserve_pages */
6827 calculate_totalreserve_pages();
6831 * setup_per_zone_wmarks - called when min_free_kbytes changes
6832 * or when memory is hot-{added|removed}
6834 * Ensures that the watermark[min,low,high] values for each zone are set
6835 * correctly with respect to min_free_kbytes.
6837 void setup_per_zone_wmarks(void)
6839 mutex_lock(&zonelists_mutex);
6840 __setup_per_zone_wmarks();
6841 mutex_unlock(&zonelists_mutex);
6845 * Initialise min_free_kbytes.
6847 * For small machines we want it small (128k min). For large machines
6848 * we want it large (64MB max). But it is not linear, because network
6849 * bandwidth does not increase linearly with machine size. We use
6851 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6852 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6868 int __meminit init_per_zone_wmark_min(void)
6870 unsigned long lowmem_kbytes;
6871 int new_min_free_kbytes;
6873 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6874 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6876 if (new_min_free_kbytes > user_min_free_kbytes) {
6877 min_free_kbytes = new_min_free_kbytes;
6878 if (min_free_kbytes < 128)
6879 min_free_kbytes = 128;
6880 if (min_free_kbytes > 65536)
6881 min_free_kbytes = 65536;
6883 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6884 new_min_free_kbytes, user_min_free_kbytes);
6886 setup_per_zone_wmarks();
6887 refresh_zone_stat_thresholds();
6888 setup_per_zone_lowmem_reserve();
6891 core_initcall(init_per_zone_wmark_min)
6894 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6895 * that we can call two helper functions whenever min_free_kbytes
6898 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6899 void __user *buffer, size_t *length, loff_t *ppos)
6903 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6908 user_min_free_kbytes = min_free_kbytes;
6909 setup_per_zone_wmarks();
6914 int watermark_scale_factor_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 setup_per_zone_wmarks();
6930 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6931 void __user *buffer, size_t *length, loff_t *ppos)
6936 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6941 zone->min_unmapped_pages = (zone->managed_pages *
6942 sysctl_min_unmapped_ratio) / 100;
6946 int sysctl_min_slab_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_slab_pages = (zone->managed_pages *
6958 sysctl_min_slab_ratio) / 100;
6964 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6965 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6966 * whenever sysctl_lowmem_reserve_ratio changes.
6968 * The reserve ratio obviously has absolutely no relation with the
6969 * minimum watermarks. The lowmem reserve ratio can only make sense
6970 * if in function of the boot time zone sizes.
6972 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6973 void __user *buffer, size_t *length, loff_t *ppos)
6975 proc_dointvec_minmax(table, write, buffer, length, ppos);
6976 setup_per_zone_lowmem_reserve();
6981 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6982 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6983 * pagelist can have before it gets flushed back to buddy allocator.
6985 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6986 void __user *buffer, size_t *length, loff_t *ppos)
6989 int old_percpu_pagelist_fraction;
6992 mutex_lock(&pcp_batch_high_lock);
6993 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6995 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6996 if (!write || ret < 0)
6999 /* Sanity checking to avoid pcp imbalance */
7000 if (percpu_pagelist_fraction &&
7001 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7002 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7008 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7011 for_each_populated_zone(zone) {
7014 for_each_possible_cpu(cpu)
7015 pageset_set_high_and_batch(zone,
7016 per_cpu_ptr(zone->pageset, cpu));
7019 mutex_unlock(&pcp_batch_high_lock);
7024 int hashdist = HASHDIST_DEFAULT;
7026 static int __init set_hashdist(char *str)
7030 hashdist = simple_strtoul(str, &str, 0);
7033 __setup("hashdist=", set_hashdist);
7037 * allocate a large system hash table from bootmem
7038 * - it is assumed that the hash table must contain an exact power-of-2
7039 * quantity of entries
7040 * - limit is the number of hash buckets, not the total allocation size
7042 void *__init alloc_large_system_hash(const char *tablename,
7043 unsigned long bucketsize,
7044 unsigned long numentries,
7047 unsigned int *_hash_shift,
7048 unsigned int *_hash_mask,
7049 unsigned long low_limit,
7050 unsigned long high_limit)
7052 unsigned long long max = high_limit;
7053 unsigned long log2qty, size;
7056 /* allow the kernel cmdline to have a say */
7058 /* round applicable memory size up to nearest megabyte */
7059 numentries = nr_kernel_pages;
7061 /* It isn't necessary when PAGE_SIZE >= 1MB */
7062 if (PAGE_SHIFT < 20)
7063 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7065 /* limit to 1 bucket per 2^scale bytes of low memory */
7066 if (scale > PAGE_SHIFT)
7067 numentries >>= (scale - PAGE_SHIFT);
7069 numentries <<= (PAGE_SHIFT - scale);
7071 /* Make sure we've got at least a 0-order allocation.. */
7072 if (unlikely(flags & HASH_SMALL)) {
7073 /* Makes no sense without HASH_EARLY */
7074 WARN_ON(!(flags & HASH_EARLY));
7075 if (!(numentries >> *_hash_shift)) {
7076 numentries = 1UL << *_hash_shift;
7077 BUG_ON(!numentries);
7079 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7080 numentries = PAGE_SIZE / bucketsize;
7082 numentries = roundup_pow_of_two(numentries);
7084 /* limit allocation size to 1/16 total memory by default */
7086 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7087 do_div(max, bucketsize);
7089 max = min(max, 0x80000000ULL);
7091 if (numentries < low_limit)
7092 numentries = low_limit;
7093 if (numentries > max)
7096 log2qty = ilog2(numentries);
7099 size = bucketsize << log2qty;
7100 if (flags & HASH_EARLY)
7101 table = memblock_virt_alloc_nopanic(size, 0);
7103 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7106 * If bucketsize is not a power-of-two, we may free
7107 * some pages at the end of hash table which
7108 * alloc_pages_exact() automatically does
7110 if (get_order(size) < MAX_ORDER) {
7111 table = alloc_pages_exact(size, GFP_ATOMIC);
7112 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7115 } while (!table && size > PAGE_SIZE && --log2qty);
7118 panic("Failed to allocate %s hash table\n", tablename);
7120 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7121 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7124 *_hash_shift = log2qty;
7126 *_hash_mask = (1 << log2qty) - 1;
7132 * This function checks whether pageblock includes unmovable pages or not.
7133 * If @count is not zero, it is okay to include less @count unmovable pages
7135 * PageLRU check without isolation or lru_lock could race so that
7136 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7137 * expect this function should be exact.
7139 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7140 bool skip_hwpoisoned_pages)
7142 unsigned long pfn, iter, found;
7146 * For avoiding noise data, lru_add_drain_all() should be called
7147 * If ZONE_MOVABLE, the zone never contains unmovable pages
7149 if (zone_idx(zone) == ZONE_MOVABLE)
7151 mt = get_pageblock_migratetype(page);
7152 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7155 pfn = page_to_pfn(page);
7156 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7157 unsigned long check = pfn + iter;
7159 if (!pfn_valid_within(check))
7162 page = pfn_to_page(check);
7165 * Hugepages are not in LRU lists, but they're movable.
7166 * We need not scan over tail pages bacause we don't
7167 * handle each tail page individually in migration.
7169 if (PageHuge(page)) {
7170 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7175 * We can't use page_count without pin a page
7176 * because another CPU can free compound page.
7177 * This check already skips compound tails of THP
7178 * because their page->_refcount is zero at all time.
7180 if (!page_ref_count(page)) {
7181 if (PageBuddy(page))
7182 iter += (1 << page_order(page)) - 1;
7187 * The HWPoisoned page may be not in buddy system, and
7188 * page_count() is not 0.
7190 if (skip_hwpoisoned_pages && PageHWPoison(page))
7196 * If there are RECLAIMABLE pages, we need to check
7197 * it. But now, memory offline itself doesn't call
7198 * shrink_node_slabs() and it still to be fixed.
7201 * If the page is not RAM, page_count()should be 0.
7202 * we don't need more check. This is an _used_ not-movable page.
7204 * The problematic thing here is PG_reserved pages. PG_reserved
7205 * is set to both of a memory hole page and a _used_ kernel
7214 bool is_pageblock_removable_nolock(struct page *page)
7220 * We have to be careful here because we are iterating over memory
7221 * sections which are not zone aware so we might end up outside of
7222 * the zone but still within the section.
7223 * We have to take care about the node as well. If the node is offline
7224 * its NODE_DATA will be NULL - see page_zone.
7226 if (!node_online(page_to_nid(page)))
7229 zone = page_zone(page);
7230 pfn = page_to_pfn(page);
7231 if (!zone_spans_pfn(zone, pfn))
7234 return !has_unmovable_pages(zone, page, 0, true);
7237 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7239 static unsigned long pfn_max_align_down(unsigned long pfn)
7241 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7242 pageblock_nr_pages) - 1);
7245 static unsigned long pfn_max_align_up(unsigned long pfn)
7247 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7248 pageblock_nr_pages));
7251 /* [start, end) must belong to a single zone. */
7252 static int __alloc_contig_migrate_range(struct compact_control *cc,
7253 unsigned long start, unsigned long end)
7255 /* This function is based on compact_zone() from compaction.c. */
7256 unsigned long nr_reclaimed;
7257 unsigned long pfn = start;
7258 unsigned int tries = 0;
7263 while (pfn < end || !list_empty(&cc->migratepages)) {
7264 if (fatal_signal_pending(current)) {
7269 if (list_empty(&cc->migratepages)) {
7270 cc->nr_migratepages = 0;
7271 pfn = isolate_migratepages_range(cc, pfn, end);
7277 } else if (++tries == 5) {
7278 ret = ret < 0 ? ret : -EBUSY;
7282 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7284 cc->nr_migratepages -= nr_reclaimed;
7286 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7287 NULL, 0, cc->mode, MR_CMA);
7290 putback_movable_pages(&cc->migratepages);
7297 * alloc_contig_range() -- tries to allocate given range of pages
7298 * @start: start PFN to allocate
7299 * @end: one-past-the-last PFN to allocate
7300 * @migratetype: migratetype of the underlaying pageblocks (either
7301 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7302 * in range must have the same migratetype and it must
7303 * be either of the two.
7305 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7306 * aligned, however it's the caller's responsibility to guarantee that
7307 * we are the only thread that changes migrate type of pageblocks the
7310 * The PFN range must belong to a single zone.
7312 * Returns zero on success or negative error code. On success all
7313 * pages which PFN is in [start, end) are allocated for the caller and
7314 * need to be freed with free_contig_range().
7316 int alloc_contig_range(unsigned long start, unsigned long end,
7317 unsigned migratetype)
7319 unsigned long outer_start, outer_end;
7323 struct compact_control cc = {
7324 .nr_migratepages = 0,
7326 .zone = page_zone(pfn_to_page(start)),
7327 .mode = MIGRATE_SYNC,
7328 .ignore_skip_hint = true,
7330 INIT_LIST_HEAD(&cc.migratepages);
7333 * What we do here is we mark all pageblocks in range as
7334 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7335 * have different sizes, and due to the way page allocator
7336 * work, we align the range to biggest of the two pages so
7337 * that page allocator won't try to merge buddies from
7338 * different pageblocks and change MIGRATE_ISOLATE to some
7339 * other migration type.
7341 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7342 * migrate the pages from an unaligned range (ie. pages that
7343 * we are interested in). This will put all the pages in
7344 * range back to page allocator as MIGRATE_ISOLATE.
7346 * When this is done, we take the pages in range from page
7347 * allocator removing them from the buddy system. This way
7348 * page allocator will never consider using them.
7350 * This lets us mark the pageblocks back as
7351 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7352 * aligned range but not in the unaligned, original range are
7353 * put back to page allocator so that buddy can use them.
7356 ret = start_isolate_page_range(pfn_max_align_down(start),
7357 pfn_max_align_up(end), migratetype,
7363 * In case of -EBUSY, we'd like to know which page causes problem.
7364 * So, just fall through. We will check it in test_pages_isolated().
7366 ret = __alloc_contig_migrate_range(&cc, start, end);
7367 if (ret && ret != -EBUSY)
7371 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7372 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7373 * more, all pages in [start, end) are free in page allocator.
7374 * What we are going to do is to allocate all pages from
7375 * [start, end) (that is remove them from page allocator).
7377 * The only problem is that pages at the beginning and at the
7378 * end of interesting range may be not aligned with pages that
7379 * page allocator holds, ie. they can be part of higher order
7380 * pages. Because of this, we reserve the bigger range and
7381 * once this is done free the pages we are not interested in.
7383 * We don't have to hold zone->lock here because the pages are
7384 * isolated thus they won't get removed from buddy.
7387 lru_add_drain_all();
7388 drain_all_pages(cc.zone);
7391 outer_start = start;
7392 while (!PageBuddy(pfn_to_page(outer_start))) {
7393 if (++order >= MAX_ORDER) {
7394 outer_start = start;
7397 outer_start &= ~0UL << order;
7400 if (outer_start != start) {
7401 order = page_order(pfn_to_page(outer_start));
7404 * outer_start page could be small order buddy page and
7405 * it doesn't include start page. Adjust outer_start
7406 * in this case to report failed page properly
7407 * on tracepoint in test_pages_isolated()
7409 if (outer_start + (1UL << order) <= start)
7410 outer_start = start;
7413 /* Make sure the range is really isolated. */
7414 if (test_pages_isolated(outer_start, end, false)) {
7415 pr_info("%s: [%lx, %lx) PFNs busy\n",
7416 __func__, outer_start, end);
7421 /* Grab isolated pages from freelists. */
7422 outer_end = isolate_freepages_range(&cc, outer_start, end);
7428 /* Free head and tail (if any) */
7429 if (start != outer_start)
7430 free_contig_range(outer_start, start - outer_start);
7431 if (end != outer_end)
7432 free_contig_range(end, outer_end - end);
7435 undo_isolate_page_range(pfn_max_align_down(start),
7436 pfn_max_align_up(end), migratetype);
7440 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7442 unsigned int count = 0;
7444 for (; nr_pages--; pfn++) {
7445 struct page *page = pfn_to_page(pfn);
7447 count += page_count(page) != 1;
7450 WARN(count != 0, "%d pages are still in use!\n", count);
7454 #ifdef CONFIG_MEMORY_HOTPLUG
7456 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7457 * page high values need to be recalulated.
7459 void __meminit zone_pcp_update(struct zone *zone)
7462 mutex_lock(&pcp_batch_high_lock);
7463 for_each_possible_cpu(cpu)
7464 pageset_set_high_and_batch(zone,
7465 per_cpu_ptr(zone->pageset, cpu));
7466 mutex_unlock(&pcp_batch_high_lock);
7470 void zone_pcp_reset(struct zone *zone)
7472 unsigned long flags;
7474 struct per_cpu_pageset *pset;
7476 /* avoid races with drain_pages() */
7477 local_irq_save(flags);
7478 if (zone->pageset != &boot_pageset) {
7479 for_each_online_cpu(cpu) {
7480 pset = per_cpu_ptr(zone->pageset, cpu);
7481 drain_zonestat(zone, pset);
7483 free_percpu(zone->pageset);
7484 zone->pageset = &boot_pageset;
7486 local_irq_restore(flags);
7489 #ifdef CONFIG_MEMORY_HOTREMOVE
7491 * All pages in the range must be in a single zone and isolated
7492 * before calling this.
7495 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7499 unsigned int order, i;
7501 unsigned long flags;
7502 /* find the first valid pfn */
7503 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7508 zone = page_zone(pfn_to_page(pfn));
7509 spin_lock_irqsave(&zone->lock, flags);
7511 while (pfn < end_pfn) {
7512 if (!pfn_valid(pfn)) {
7516 page = pfn_to_page(pfn);
7518 * The HWPoisoned page may be not in buddy system, and
7519 * page_count() is not 0.
7521 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7523 SetPageReserved(page);
7527 BUG_ON(page_count(page));
7528 BUG_ON(!PageBuddy(page));
7529 order = page_order(page);
7530 #ifdef CONFIG_DEBUG_VM
7531 pr_info("remove from free list %lx %d %lx\n",
7532 pfn, 1 << order, end_pfn);
7534 list_del(&page->lru);
7535 rmv_page_order(page);
7536 zone->free_area[order].nr_free--;
7537 for (i = 0; i < (1 << order); i++)
7538 SetPageReserved((page+i));
7539 pfn += (1 << order);
7541 spin_unlock_irqrestore(&zone->lock, flags);
7545 bool is_free_buddy_page(struct page *page)
7547 struct zone *zone = page_zone(page);
7548 unsigned long pfn = page_to_pfn(page);
7549 unsigned long flags;
7552 spin_lock_irqsave(&zone->lock, flags);
7553 for (order = 0; order < MAX_ORDER; order++) {
7554 struct page *page_head = page - (pfn & ((1 << order) - 1));
7556 if (PageBuddy(page_head) && page_order(page_head) >= order)
7559 spin_unlock_irqrestore(&zone->lock, flags);
7561 return order < MAX_ORDER;