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)
299 * Returns false when the remaining initialisation should be deferred until
300 * later in the boot cycle when it can be parallelised.
302 static inline bool update_defer_init(pg_data_t *pgdat,
303 unsigned long pfn, unsigned long zone_end,
304 unsigned long *nr_initialised)
306 unsigned long max_initialise;
308 /* Always populate low zones for address-contrained allocations */
309 if (zone_end < pgdat_end_pfn(pgdat))
312 * Initialise at least 2G of a node but also take into account that
313 * two large system hashes that can take up 1GB for 0.25TB/node.
315 max_initialise = max(2UL << (30 - PAGE_SHIFT),
316 (pgdat->node_spanned_pages >> 8));
319 if ((*nr_initialised > max_initialise) &&
320 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
321 pgdat->first_deferred_pfn = pfn;
328 static inline void reset_deferred_meminit(pg_data_t *pgdat)
332 static inline bool early_page_uninitialised(unsigned long pfn)
337 static inline bool update_defer_init(pg_data_t *pgdat,
338 unsigned long pfn, unsigned long zone_end,
339 unsigned long *nr_initialised)
345 /* Return a pointer to the bitmap storing bits affecting a block of pages */
346 static inline unsigned long *get_pageblock_bitmap(struct page *page,
349 #ifdef CONFIG_SPARSEMEM
350 return __pfn_to_section(pfn)->pageblock_flags;
352 return page_zone(page)->pageblock_flags;
353 #endif /* CONFIG_SPARSEMEM */
356 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
358 #ifdef CONFIG_SPARSEMEM
359 pfn &= (PAGES_PER_SECTION-1);
360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
363 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
364 #endif /* CONFIG_SPARSEMEM */
368 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
369 * @page: The page within the block of interest
370 * @pfn: The target page frame number
371 * @end_bitidx: The last bit of interest to retrieve
372 * @mask: mask of bits that the caller is interested in
374 * Return: pageblock_bits flags
376 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
378 unsigned long end_bitidx,
381 unsigned long *bitmap;
382 unsigned long bitidx, word_bitidx;
385 bitmap = get_pageblock_bitmap(page, pfn);
386 bitidx = pfn_to_bitidx(page, pfn);
387 word_bitidx = bitidx / BITS_PER_LONG;
388 bitidx &= (BITS_PER_LONG-1);
390 word = bitmap[word_bitidx];
391 bitidx += end_bitidx;
392 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
395 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
396 unsigned long end_bitidx,
399 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
402 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
404 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
408 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
409 * @page: The page within the block of interest
410 * @flags: The flags to set
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest
413 * @mask: mask of bits that the caller is interested in
415 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
417 unsigned long end_bitidx,
420 unsigned long *bitmap;
421 unsigned long bitidx, word_bitidx;
422 unsigned long old_word, word;
424 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
426 bitmap = get_pageblock_bitmap(page, pfn);
427 bitidx = pfn_to_bitidx(page, pfn);
428 word_bitidx = bitidx / BITS_PER_LONG;
429 bitidx &= (BITS_PER_LONG-1);
431 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
433 bitidx += end_bitidx;
434 mask <<= (BITS_PER_LONG - bitidx - 1);
435 flags <<= (BITS_PER_LONG - bitidx - 1);
437 word = READ_ONCE(bitmap[word_bitidx]);
439 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
440 if (word == old_word)
446 void set_pageblock_migratetype(struct page *page, int migratetype)
448 if (unlikely(page_group_by_mobility_disabled &&
449 migratetype < MIGRATE_PCPTYPES))
450 migratetype = MIGRATE_UNMOVABLE;
452 set_pageblock_flags_group(page, (unsigned long)migratetype,
453 PB_migrate, PB_migrate_end);
456 #ifdef CONFIG_DEBUG_VM
457 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
461 unsigned long pfn = page_to_pfn(page);
462 unsigned long sp, start_pfn;
465 seq = zone_span_seqbegin(zone);
466 start_pfn = zone->zone_start_pfn;
467 sp = zone->spanned_pages;
468 if (!zone_spans_pfn(zone, pfn))
470 } while (zone_span_seqretry(zone, seq));
473 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
474 pfn, zone_to_nid(zone), zone->name,
475 start_pfn, start_pfn + sp);
480 static int page_is_consistent(struct zone *zone, struct page *page)
482 if (!pfn_valid_within(page_to_pfn(page)))
484 if (zone != page_zone(page))
490 * Temporary debugging check for pages not lying within a given zone.
492 static int bad_range(struct zone *zone, struct page *page)
494 if (page_outside_zone_boundaries(zone, page))
496 if (!page_is_consistent(zone, page))
502 static inline int bad_range(struct zone *zone, struct page *page)
508 static void bad_page(struct page *page, const char *reason,
509 unsigned long bad_flags)
511 static unsigned long resume;
512 static unsigned long nr_shown;
513 static unsigned long nr_unshown;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown == 60) {
520 if (time_before(jiffies, resume)) {
526 "BUG: Bad page state: %lu messages suppressed\n",
533 resume = jiffies + 60 * HZ;
535 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
536 current->comm, page_to_pfn(page));
537 __dump_page(page, reason);
538 bad_flags &= page->flags;
540 pr_alert("bad because of flags: %#lx(%pGp)\n",
541 bad_flags, &bad_flags);
542 dump_page_owner(page);
547 /* Leave bad fields for debug, except PageBuddy could make trouble */
548 page_mapcount_reset(page); /* remove PageBuddy */
549 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
553 * Higher-order pages are called "compound pages". They are structured thusly:
555 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
557 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
558 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
560 * The first tail page's ->compound_dtor holds the offset in array of compound
561 * page destructors. See compound_page_dtors.
563 * The first tail page's ->compound_order holds the order of allocation.
564 * This usage means that zero-order pages may not be compound.
567 void free_compound_page(struct page *page)
569 __free_pages_ok(page, compound_order(page));
572 void prep_compound_page(struct page *page, unsigned int order)
575 int nr_pages = 1 << order;
577 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
578 set_compound_order(page, order);
580 for (i = 1; i < nr_pages; i++) {
581 struct page *p = page + i;
582 set_page_count(p, 0);
583 p->mapping = TAIL_MAPPING;
584 set_compound_head(p, page);
586 atomic_set(compound_mapcount_ptr(page), -1);
589 #ifdef CONFIG_DEBUG_PAGEALLOC
590 unsigned int _debug_guardpage_minorder;
591 bool _debug_pagealloc_enabled __read_mostly
592 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
593 EXPORT_SYMBOL(_debug_pagealloc_enabled);
594 bool _debug_guardpage_enabled __read_mostly;
596 static int __init early_debug_pagealloc(char *buf)
600 return kstrtobool(buf, &_debug_pagealloc_enabled);
602 early_param("debug_pagealloc", early_debug_pagealloc);
604 static bool need_debug_guardpage(void)
606 /* If we don't use debug_pagealloc, we don't need guard page */
607 if (!debug_pagealloc_enabled())
613 static void init_debug_guardpage(void)
615 if (!debug_pagealloc_enabled())
618 _debug_guardpage_enabled = true;
621 struct page_ext_operations debug_guardpage_ops = {
622 .need = need_debug_guardpage,
623 .init = init_debug_guardpage,
626 static int __init debug_guardpage_minorder_setup(char *buf)
630 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
631 pr_err("Bad debug_guardpage_minorder value\n");
634 _debug_guardpage_minorder = res;
635 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
638 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
640 static inline void set_page_guard(struct zone *zone, struct page *page,
641 unsigned int order, int migratetype)
643 struct page_ext *page_ext;
645 if (!debug_guardpage_enabled())
648 page_ext = lookup_page_ext(page);
649 if (unlikely(!page_ext))
652 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
654 INIT_LIST_HEAD(&page->lru);
655 set_page_private(page, order);
656 /* Guard pages are not available for any usage */
657 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
660 static inline void clear_page_guard(struct zone *zone, struct page *page,
661 unsigned int order, int migratetype)
663 struct page_ext *page_ext;
665 if (!debug_guardpage_enabled())
668 page_ext = lookup_page_ext(page);
669 if (unlikely(!page_ext))
672 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
674 set_page_private(page, 0);
675 if (!is_migrate_isolate(migratetype))
676 __mod_zone_freepage_state(zone, (1 << order), migratetype);
679 struct page_ext_operations debug_guardpage_ops = { NULL, };
680 static inline void set_page_guard(struct zone *zone, struct page *page,
681 unsigned int order, int migratetype) {}
682 static inline void clear_page_guard(struct zone *zone, struct page *page,
683 unsigned int order, int migratetype) {}
686 static inline void set_page_order(struct page *page, unsigned int order)
688 set_page_private(page, order);
689 __SetPageBuddy(page);
692 static inline void rmv_page_order(struct page *page)
694 __ClearPageBuddy(page);
695 set_page_private(page, 0);
699 * This function checks whether a page is free && is the buddy
700 * we can do coalesce a page and its buddy if
701 * (a) the buddy is not in a hole &&
702 * (b) the buddy is in the buddy system &&
703 * (c) a page and its buddy have the same order &&
704 * (d) a page and its buddy are in the same zone.
706 * For recording whether a page is in the buddy system, we set ->_mapcount
707 * PAGE_BUDDY_MAPCOUNT_VALUE.
708 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
709 * serialized by zone->lock.
711 * For recording page's order, we use page_private(page).
713 static inline int page_is_buddy(struct page *page, struct page *buddy,
716 if (!pfn_valid_within(page_to_pfn(buddy)))
719 if (page_is_guard(buddy) && page_order(buddy) == order) {
720 if (page_zone_id(page) != page_zone_id(buddy))
723 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
728 if (PageBuddy(buddy) && page_order(buddy) == order) {
730 * zone check is done late to avoid uselessly
731 * calculating zone/node ids for pages that could
734 if (page_zone_id(page) != page_zone_id(buddy))
737 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
745 * Freeing function for a buddy system allocator.
747 * The concept of a buddy system is to maintain direct-mapped table
748 * (containing bit values) for memory blocks of various "orders".
749 * The bottom level table contains the map for the smallest allocatable
750 * units of memory (here, pages), and each level above it describes
751 * pairs of units from the levels below, hence, "buddies".
752 * At a high level, all that happens here is marking the table entry
753 * at the bottom level available, and propagating the changes upward
754 * as necessary, plus some accounting needed to play nicely with other
755 * parts of the VM system.
756 * At each level, we keep a list of pages, which are heads of continuous
757 * free pages of length of (1 << order) and marked with _mapcount
758 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
760 * So when we are allocating or freeing one, we can derive the state of the
761 * other. That is, if we allocate a small block, and both were
762 * free, the remainder of the region must be split into blocks.
763 * If a block is freed, and its buddy is also free, then this
764 * triggers coalescing into a block of larger size.
769 static inline void __free_one_page(struct page *page,
771 struct zone *zone, unsigned int order,
774 unsigned long page_idx;
775 unsigned long combined_idx;
776 unsigned long uninitialized_var(buddy_idx);
778 unsigned int max_order;
780 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
782 VM_BUG_ON(!zone_is_initialized(zone));
783 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
785 VM_BUG_ON(migratetype == -1);
786 if (likely(!is_migrate_isolate(migratetype)))
787 __mod_zone_freepage_state(zone, 1 << order, migratetype);
789 page_idx = pfn & ((1 << MAX_ORDER) - 1);
791 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
792 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795 while (order < max_order - 1) {
796 buddy_idx = __find_buddy_index(page_idx, order);
797 buddy = page + (buddy_idx - page_idx);
798 if (!page_is_buddy(page, buddy, order))
801 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
802 * merge with it and move up one order.
804 if (page_is_guard(buddy)) {
805 clear_page_guard(zone, buddy, order, migratetype);
807 list_del(&buddy->lru);
808 zone->free_area[order].nr_free--;
809 rmv_page_order(buddy);
811 combined_idx = buddy_idx & page_idx;
812 page = page + (combined_idx - page_idx);
813 page_idx = combined_idx;
816 if (max_order < MAX_ORDER) {
817 /* If we are here, it means order is >= pageblock_order.
818 * We want to prevent merge between freepages on isolate
819 * pageblock and normal pageblock. Without this, pageblock
820 * isolation could cause incorrect freepage or CMA accounting.
822 * We don't want to hit this code for the more frequent
825 if (unlikely(has_isolate_pageblock(zone))) {
828 buddy_idx = __find_buddy_index(page_idx, order);
829 buddy = page + (buddy_idx - page_idx);
830 buddy_mt = get_pageblock_migratetype(buddy);
832 if (migratetype != buddy_mt
833 && (is_migrate_isolate(migratetype) ||
834 is_migrate_isolate(buddy_mt)))
838 goto continue_merging;
842 set_page_order(page, order);
845 * If this is not the largest possible page, check if the buddy
846 * of the next-highest order is free. If it is, it's possible
847 * that pages are being freed that will coalesce soon. In case,
848 * that is happening, add the free page to the tail of the list
849 * so it's less likely to be used soon and more likely to be merged
850 * as a higher order page
852 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
853 struct page *higher_page, *higher_buddy;
854 combined_idx = buddy_idx & page_idx;
855 higher_page = page + (combined_idx - page_idx);
856 buddy_idx = __find_buddy_index(combined_idx, order + 1);
857 higher_buddy = higher_page + (buddy_idx - combined_idx);
858 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
859 list_add_tail(&page->lru,
860 &zone->free_area[order].free_list[migratetype]);
865 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
867 zone->free_area[order].nr_free++;
871 * A bad page could be due to a number of fields. Instead of multiple branches,
872 * try and check multiple fields with one check. The caller must do a detailed
873 * check if necessary.
875 static inline bool page_expected_state(struct page *page,
876 unsigned long check_flags)
878 if (unlikely(atomic_read(&page->_mapcount) != -1))
881 if (unlikely((unsigned long)page->mapping |
882 page_ref_count(page) |
884 (unsigned long)page->mem_cgroup |
886 (page->flags & check_flags)))
892 static void free_pages_check_bad(struct page *page)
894 const char *bad_reason;
895 unsigned long bad_flags;
900 if (unlikely(atomic_read(&page->_mapcount) != -1))
901 bad_reason = "nonzero mapcount";
902 if (unlikely(page->mapping != NULL))
903 bad_reason = "non-NULL mapping";
904 if (unlikely(page_ref_count(page) != 0))
905 bad_reason = "nonzero _refcount";
906 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
907 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
908 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
911 if (unlikely(page->mem_cgroup))
912 bad_reason = "page still charged to cgroup";
914 bad_page(page, bad_reason, bad_flags);
917 static inline int free_pages_check(struct page *page)
919 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
922 /* Something has gone sideways, find it */
923 free_pages_check_bad(page);
927 static int free_tail_pages_check(struct page *head_page, struct page *page)
932 * We rely page->lru.next never has bit 0 set, unless the page
933 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
935 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
937 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
941 switch (page - head_page) {
943 /* the first tail page: ->mapping is compound_mapcount() */
944 if (unlikely(compound_mapcount(page))) {
945 bad_page(page, "nonzero compound_mapcount", 0);
951 * the second tail page: ->mapping is
952 * page_deferred_list().next -- ignore value.
956 if (page->mapping != TAIL_MAPPING) {
957 bad_page(page, "corrupted mapping in tail page", 0);
962 if (unlikely(!PageTail(page))) {
963 bad_page(page, "PageTail not set", 0);
966 if (unlikely(compound_head(page) != head_page)) {
967 bad_page(page, "compound_head not consistent", 0);
972 page->mapping = NULL;
973 clear_compound_head(page);
977 static __always_inline bool free_pages_prepare(struct page *page,
978 unsigned int order, bool check_free)
982 VM_BUG_ON_PAGE(PageTail(page), page);
984 trace_mm_page_free(page, order);
985 kmemcheck_free_shadow(page, order);
988 * Check tail pages before head page information is cleared to
989 * avoid checking PageCompound for order-0 pages.
991 if (unlikely(order)) {
992 bool compound = PageCompound(page);
995 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
998 ClearPageDoubleMap(page);
999 for (i = 1; i < (1 << order); i++) {
1001 bad += free_tail_pages_check(page, page + i);
1002 if (unlikely(free_pages_check(page + i))) {
1006 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1009 if (PageMappingFlags(page))
1010 page->mapping = NULL;
1011 if (memcg_kmem_enabled() && PageKmemcg(page)) {
1012 memcg_kmem_uncharge(page, order);
1013 __ClearPageKmemcg(page);
1016 bad += free_pages_check(page);
1020 page_cpupid_reset_last(page);
1021 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1022 reset_page_owner(page, order);
1024 if (!PageHighMem(page)) {
1025 debug_check_no_locks_freed(page_address(page),
1026 PAGE_SIZE << order);
1027 debug_check_no_obj_freed(page_address(page),
1028 PAGE_SIZE << order);
1030 arch_free_page(page, order);
1031 kernel_poison_pages(page, 1 << order, 0);
1032 kernel_map_pages(page, 1 << order, 0);
1033 kasan_free_pages(page, order);
1038 #ifdef CONFIG_DEBUG_VM
1039 static inline bool free_pcp_prepare(struct page *page)
1041 return free_pages_prepare(page, 0, true);
1044 static inline bool bulkfree_pcp_prepare(struct page *page)
1049 static bool free_pcp_prepare(struct page *page)
1051 return free_pages_prepare(page, 0, false);
1054 static bool bulkfree_pcp_prepare(struct page *page)
1056 return free_pages_check(page);
1058 #endif /* CONFIG_DEBUG_VM */
1061 * Frees a number of pages from the PCP lists
1062 * Assumes all pages on list are in same zone, and of same order.
1063 * count is the number of pages to free.
1065 * If the zone was previously in an "all pages pinned" state then look to
1066 * see if this freeing clears that state.
1068 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1069 * pinned" detection logic.
1071 static void free_pcppages_bulk(struct zone *zone, int count,
1072 struct per_cpu_pages *pcp)
1074 int migratetype = 0;
1076 unsigned long nr_scanned;
1077 bool isolated_pageblocks;
1079 spin_lock(&zone->lock);
1080 isolated_pageblocks = has_isolate_pageblock(zone);
1081 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1083 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1087 struct list_head *list;
1090 * Remove pages from lists in a round-robin fashion. A
1091 * batch_free count is maintained that is incremented when an
1092 * empty list is encountered. This is so more pages are freed
1093 * off fuller lists instead of spinning excessively around empty
1098 if (++migratetype == MIGRATE_PCPTYPES)
1100 list = &pcp->lists[migratetype];
1101 } while (list_empty(list));
1103 /* This is the only non-empty list. Free them all. */
1104 if (batch_free == MIGRATE_PCPTYPES)
1108 int mt; /* migratetype of the to-be-freed page */
1110 page = list_last_entry(list, struct page, lru);
1111 /* must delete as __free_one_page list manipulates */
1112 list_del(&page->lru);
1114 mt = get_pcppage_migratetype(page);
1115 /* MIGRATE_ISOLATE page should not go to pcplists */
1116 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1117 /* Pageblock could have been isolated meanwhile */
1118 if (unlikely(isolated_pageblocks))
1119 mt = get_pageblock_migratetype(page);
1121 if (bulkfree_pcp_prepare(page))
1124 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1125 trace_mm_page_pcpu_drain(page, 0, mt);
1126 } while (--count && --batch_free && !list_empty(list));
1128 spin_unlock(&zone->lock);
1131 static void free_one_page(struct zone *zone,
1132 struct page *page, unsigned long pfn,
1136 unsigned long nr_scanned;
1137 spin_lock(&zone->lock);
1138 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1140 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1142 if (unlikely(has_isolate_pageblock(zone) ||
1143 is_migrate_isolate(migratetype))) {
1144 migratetype = get_pfnblock_migratetype(page, pfn);
1146 __free_one_page(page, pfn, zone, order, migratetype);
1147 spin_unlock(&zone->lock);
1150 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1151 unsigned long zone, int nid)
1153 set_page_links(page, zone, nid, pfn);
1154 init_page_count(page);
1155 page_mapcount_reset(page);
1156 page_cpupid_reset_last(page);
1158 INIT_LIST_HEAD(&page->lru);
1159 #ifdef WANT_PAGE_VIRTUAL
1160 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1161 if (!is_highmem_idx(zone))
1162 set_page_address(page, __va(pfn << PAGE_SHIFT));
1166 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1169 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1172 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1173 static void init_reserved_page(unsigned long pfn)
1178 if (!early_page_uninitialised(pfn))
1181 nid = early_pfn_to_nid(pfn);
1182 pgdat = NODE_DATA(nid);
1184 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1185 struct zone *zone = &pgdat->node_zones[zid];
1187 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1190 __init_single_pfn(pfn, zid, nid);
1193 static inline void init_reserved_page(unsigned long pfn)
1196 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1199 * Initialised pages do not have PageReserved set. This function is
1200 * called for each range allocated by the bootmem allocator and
1201 * marks the pages PageReserved. The remaining valid pages are later
1202 * sent to the buddy page allocator.
1204 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1206 unsigned long start_pfn = PFN_DOWN(start);
1207 unsigned long end_pfn = PFN_UP(end);
1209 for (; start_pfn < end_pfn; start_pfn++) {
1210 if (pfn_valid(start_pfn)) {
1211 struct page *page = pfn_to_page(start_pfn);
1213 init_reserved_page(start_pfn);
1215 /* Avoid false-positive PageTail() */
1216 INIT_LIST_HEAD(&page->lru);
1218 SetPageReserved(page);
1223 static void __free_pages_ok(struct page *page, unsigned int order)
1225 unsigned long flags;
1227 unsigned long pfn = page_to_pfn(page);
1229 if (!free_pages_prepare(page, order, true))
1232 migratetype = get_pfnblock_migratetype(page, pfn);
1233 local_irq_save(flags);
1234 __count_vm_events(PGFREE, 1 << order);
1235 free_one_page(page_zone(page), page, pfn, order, migratetype);
1236 local_irq_restore(flags);
1239 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1241 unsigned int nr_pages = 1 << order;
1242 struct page *p = page;
1246 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1248 __ClearPageReserved(p);
1249 set_page_count(p, 0);
1251 __ClearPageReserved(p);
1252 set_page_count(p, 0);
1254 page_zone(page)->managed_pages += nr_pages;
1255 set_page_refcounted(page);
1256 __free_pages(page, order);
1259 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1260 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1262 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1264 int __meminit early_pfn_to_nid(unsigned long pfn)
1266 static DEFINE_SPINLOCK(early_pfn_lock);
1269 spin_lock(&early_pfn_lock);
1270 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1272 nid = first_online_node;
1273 spin_unlock(&early_pfn_lock);
1279 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1280 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1281 struct mminit_pfnnid_cache *state)
1285 nid = __early_pfn_to_nid(pfn, state);
1286 if (nid >= 0 && nid != node)
1291 /* Only safe to use early in boot when initialisation is single-threaded */
1292 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1294 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1299 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1303 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1304 struct mminit_pfnnid_cache *state)
1311 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1314 if (early_page_uninitialised(pfn))
1316 return __free_pages_boot_core(page, order);
1320 * Check that the whole (or subset of) a pageblock given by the interval of
1321 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1322 * with the migration of free compaction scanner. The scanners then need to
1323 * use only pfn_valid_within() check for arches that allow holes within
1326 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1328 * It's possible on some configurations to have a setup like node0 node1 node0
1329 * i.e. it's possible that all pages within a zones range of pages do not
1330 * belong to a single zone. We assume that a border between node0 and node1
1331 * can occur within a single pageblock, but not a node0 node1 node0
1332 * interleaving within a single pageblock. It is therefore sufficient to check
1333 * the first and last page of a pageblock and avoid checking each individual
1334 * page in a pageblock.
1336 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1337 unsigned long end_pfn, struct zone *zone)
1339 struct page *start_page;
1340 struct page *end_page;
1342 /* end_pfn is one past the range we are checking */
1345 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1348 start_page = pfn_to_page(start_pfn);
1350 if (page_zone(start_page) != zone)
1353 end_page = pfn_to_page(end_pfn);
1355 /* This gives a shorter code than deriving page_zone(end_page) */
1356 if (page_zone_id(start_page) != page_zone_id(end_page))
1362 void set_zone_contiguous(struct zone *zone)
1364 unsigned long block_start_pfn = zone->zone_start_pfn;
1365 unsigned long block_end_pfn;
1367 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1368 for (; block_start_pfn < zone_end_pfn(zone);
1369 block_start_pfn = block_end_pfn,
1370 block_end_pfn += pageblock_nr_pages) {
1372 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1374 if (!__pageblock_pfn_to_page(block_start_pfn,
1375 block_end_pfn, zone))
1379 /* We confirm that there is no hole */
1380 zone->contiguous = true;
1383 void clear_zone_contiguous(struct zone *zone)
1385 zone->contiguous = false;
1388 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1389 static void __init deferred_free_range(struct page *page,
1390 unsigned long pfn, int nr_pages)
1397 /* Free a large naturally-aligned chunk if possible */
1398 if (nr_pages == MAX_ORDER_NR_PAGES &&
1399 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1400 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1401 __free_pages_boot_core(page, MAX_ORDER-1);
1405 for (i = 0; i < nr_pages; i++, page++)
1406 __free_pages_boot_core(page, 0);
1409 /* Completion tracking for deferred_init_memmap() threads */
1410 static atomic_t pgdat_init_n_undone __initdata;
1411 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1413 static inline void __init pgdat_init_report_one_done(void)
1415 if (atomic_dec_and_test(&pgdat_init_n_undone))
1416 complete(&pgdat_init_all_done_comp);
1419 /* Initialise remaining memory on a node */
1420 static int __init deferred_init_memmap(void *data)
1422 pg_data_t *pgdat = data;
1423 int nid = pgdat->node_id;
1424 struct mminit_pfnnid_cache nid_init_state = { };
1425 unsigned long start = jiffies;
1426 unsigned long nr_pages = 0;
1427 unsigned long walk_start, walk_end;
1430 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1431 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1433 if (first_init_pfn == ULONG_MAX) {
1434 pgdat_init_report_one_done();
1438 /* Bind memory initialisation thread to a local node if possible */
1439 if (!cpumask_empty(cpumask))
1440 set_cpus_allowed_ptr(current, cpumask);
1442 /* Sanity check boundaries */
1443 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1444 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1445 pgdat->first_deferred_pfn = ULONG_MAX;
1447 /* Only the highest zone is deferred so find it */
1448 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1449 zone = pgdat->node_zones + zid;
1450 if (first_init_pfn < zone_end_pfn(zone))
1454 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1455 unsigned long pfn, end_pfn;
1456 struct page *page = NULL;
1457 struct page *free_base_page = NULL;
1458 unsigned long free_base_pfn = 0;
1461 end_pfn = min(walk_end, zone_end_pfn(zone));
1462 pfn = first_init_pfn;
1463 if (pfn < walk_start)
1465 if (pfn < zone->zone_start_pfn)
1466 pfn = zone->zone_start_pfn;
1468 for (; pfn < end_pfn; pfn++) {
1469 if (!pfn_valid_within(pfn))
1473 * Ensure pfn_valid is checked every
1474 * MAX_ORDER_NR_PAGES for memory holes
1476 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1477 if (!pfn_valid(pfn)) {
1483 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1488 /* Minimise pfn page lookups and scheduler checks */
1489 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1492 nr_pages += nr_to_free;
1493 deferred_free_range(free_base_page,
1494 free_base_pfn, nr_to_free);
1495 free_base_page = NULL;
1496 free_base_pfn = nr_to_free = 0;
1498 page = pfn_to_page(pfn);
1503 VM_BUG_ON(page_zone(page) != zone);
1507 __init_single_page(page, pfn, zid, nid);
1508 if (!free_base_page) {
1509 free_base_page = page;
1510 free_base_pfn = pfn;
1515 /* Where possible, batch up pages for a single free */
1518 /* Free the current block of pages to allocator */
1519 nr_pages += nr_to_free;
1520 deferred_free_range(free_base_page, free_base_pfn,
1522 free_base_page = NULL;
1523 free_base_pfn = nr_to_free = 0;
1526 first_init_pfn = max(end_pfn, first_init_pfn);
1529 /* Sanity check that the next zone really is unpopulated */
1530 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1532 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1533 jiffies_to_msecs(jiffies - start));
1535 pgdat_init_report_one_done();
1538 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1540 void __init page_alloc_init_late(void)
1544 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1547 /* There will be num_node_state(N_MEMORY) threads */
1548 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1549 for_each_node_state(nid, N_MEMORY) {
1550 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1553 /* Block until all are initialised */
1554 wait_for_completion(&pgdat_init_all_done_comp);
1556 /* Reinit limits that are based on free pages after the kernel is up */
1557 files_maxfiles_init();
1560 for_each_populated_zone(zone)
1561 set_zone_contiguous(zone);
1565 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1566 void __init init_cma_reserved_pageblock(struct page *page)
1568 unsigned i = pageblock_nr_pages;
1569 struct page *p = page;
1572 __ClearPageReserved(p);
1573 set_page_count(p, 0);
1576 set_pageblock_migratetype(page, MIGRATE_CMA);
1578 if (pageblock_order >= MAX_ORDER) {
1579 i = pageblock_nr_pages;
1582 set_page_refcounted(p);
1583 __free_pages(p, MAX_ORDER - 1);
1584 p += MAX_ORDER_NR_PAGES;
1585 } while (i -= MAX_ORDER_NR_PAGES);
1587 set_page_refcounted(page);
1588 __free_pages(page, pageblock_order);
1591 adjust_managed_page_count(page, pageblock_nr_pages);
1596 * The order of subdivision here is critical for the IO subsystem.
1597 * Please do not alter this order without good reasons and regression
1598 * testing. Specifically, as large blocks of memory are subdivided,
1599 * the order in which smaller blocks are delivered depends on the order
1600 * they're subdivided in this function. This is the primary factor
1601 * influencing the order in which pages are delivered to the IO
1602 * subsystem according to empirical testing, and this is also justified
1603 * by considering the behavior of a buddy system containing a single
1604 * large block of memory acted on by a series of small allocations.
1605 * This behavior is a critical factor in sglist merging's success.
1609 static inline void expand(struct zone *zone, struct page *page,
1610 int low, int high, struct free_area *area,
1613 unsigned long size = 1 << high;
1615 while (high > low) {
1619 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1621 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1622 debug_guardpage_enabled() &&
1623 high < debug_guardpage_minorder()) {
1625 * Mark as guard pages (or page), that will allow to
1626 * merge back to allocator when buddy will be freed.
1627 * Corresponding page table entries will not be touched,
1628 * pages will stay not present in virtual address space
1630 set_page_guard(zone, &page[size], high, migratetype);
1633 list_add(&page[size].lru, &area->free_list[migratetype]);
1635 set_page_order(&page[size], high);
1639 static void check_new_page_bad(struct page *page)
1641 const char *bad_reason = NULL;
1642 unsigned long bad_flags = 0;
1644 if (unlikely(atomic_read(&page->_mapcount) != -1))
1645 bad_reason = "nonzero mapcount";
1646 if (unlikely(page->mapping != NULL))
1647 bad_reason = "non-NULL mapping";
1648 if (unlikely(page_ref_count(page) != 0))
1649 bad_reason = "nonzero _count";
1650 if (unlikely(page->flags & __PG_HWPOISON)) {
1651 bad_reason = "HWPoisoned (hardware-corrupted)";
1652 bad_flags = __PG_HWPOISON;
1653 /* Don't complain about hwpoisoned pages */
1654 page_mapcount_reset(page); /* remove PageBuddy */
1657 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1658 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1659 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1662 if (unlikely(page->mem_cgroup))
1663 bad_reason = "page still charged to cgroup";
1665 bad_page(page, bad_reason, bad_flags);
1669 * This page is about to be returned from the page allocator
1671 static inline int check_new_page(struct page *page)
1673 if (likely(page_expected_state(page,
1674 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1677 check_new_page_bad(page);
1681 static inline bool free_pages_prezeroed(bool poisoned)
1683 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1684 page_poisoning_enabled() && poisoned;
1687 #ifdef CONFIG_DEBUG_VM
1688 static bool check_pcp_refill(struct page *page)
1693 static bool check_new_pcp(struct page *page)
1695 return check_new_page(page);
1698 static bool check_pcp_refill(struct page *page)
1700 return check_new_page(page);
1702 static bool check_new_pcp(struct page *page)
1706 #endif /* CONFIG_DEBUG_VM */
1708 static bool check_new_pages(struct page *page, unsigned int order)
1711 for (i = 0; i < (1 << order); i++) {
1712 struct page *p = page + i;
1714 if (unlikely(check_new_page(p)))
1721 inline void post_alloc_hook(struct page *page, unsigned int order,
1724 set_page_private(page, 0);
1725 set_page_refcounted(page);
1727 arch_alloc_page(page, order);
1728 kernel_map_pages(page, 1 << order, 1);
1729 kernel_poison_pages(page, 1 << order, 1);
1730 kasan_alloc_pages(page, order);
1731 set_page_owner(page, order, gfp_flags);
1734 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1735 unsigned int alloc_flags)
1738 bool poisoned = true;
1740 for (i = 0; i < (1 << order); i++) {
1741 struct page *p = page + i;
1743 poisoned &= page_is_poisoned(p);
1746 post_alloc_hook(page, order, gfp_flags);
1748 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1749 for (i = 0; i < (1 << order); i++)
1750 clear_highpage(page + i);
1752 if (order && (gfp_flags & __GFP_COMP))
1753 prep_compound_page(page, order);
1756 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1757 * allocate the page. The expectation is that the caller is taking
1758 * steps that will free more memory. The caller should avoid the page
1759 * being used for !PFMEMALLOC purposes.
1761 if (alloc_flags & ALLOC_NO_WATERMARKS)
1762 set_page_pfmemalloc(page);
1764 clear_page_pfmemalloc(page);
1768 * Go through the free lists for the given migratetype and remove
1769 * the smallest available page from the freelists
1772 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1775 unsigned int current_order;
1776 struct free_area *area;
1779 /* Find a page of the appropriate size in the preferred list */
1780 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1781 area = &(zone->free_area[current_order]);
1782 page = list_first_entry_or_null(&area->free_list[migratetype],
1786 list_del(&page->lru);
1787 rmv_page_order(page);
1789 expand(zone, page, order, current_order, area, migratetype);
1790 set_pcppage_migratetype(page, migratetype);
1799 * This array describes the order lists are fallen back to when
1800 * the free lists for the desirable migrate type are depleted
1802 static int fallbacks[MIGRATE_TYPES][4] = {
1803 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1804 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1805 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1807 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1809 #ifdef CONFIG_MEMORY_ISOLATION
1810 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1815 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1818 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1821 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1822 unsigned int order) { return NULL; }
1826 * Move the free pages in a range to the free lists of the requested type.
1827 * Note that start_page and end_pages are not aligned on a pageblock
1828 * boundary. If alignment is required, use move_freepages_block()
1830 int move_freepages(struct zone *zone,
1831 struct page *start_page, struct page *end_page,
1836 int pages_moved = 0;
1838 #ifndef CONFIG_HOLES_IN_ZONE
1840 * page_zone is not safe to call in this context when
1841 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1842 * anyway as we check zone boundaries in move_freepages_block().
1843 * Remove at a later date when no bug reports exist related to
1844 * grouping pages by mobility
1846 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1849 for (page = start_page; page <= end_page;) {
1850 /* Make sure we are not inadvertently changing nodes */
1851 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1853 if (!pfn_valid_within(page_to_pfn(page))) {
1858 if (!PageBuddy(page)) {
1863 order = page_order(page);
1864 list_move(&page->lru,
1865 &zone->free_area[order].free_list[migratetype]);
1867 pages_moved += 1 << order;
1873 int move_freepages_block(struct zone *zone, struct page *page,
1876 unsigned long start_pfn, end_pfn;
1877 struct page *start_page, *end_page;
1879 start_pfn = page_to_pfn(page);
1880 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1881 start_page = pfn_to_page(start_pfn);
1882 end_page = start_page + pageblock_nr_pages - 1;
1883 end_pfn = start_pfn + pageblock_nr_pages - 1;
1885 /* Do not cross zone boundaries */
1886 if (!zone_spans_pfn(zone, start_pfn))
1888 if (!zone_spans_pfn(zone, end_pfn))
1891 return move_freepages(zone, start_page, end_page, migratetype);
1894 static void change_pageblock_range(struct page *pageblock_page,
1895 int start_order, int migratetype)
1897 int nr_pageblocks = 1 << (start_order - pageblock_order);
1899 while (nr_pageblocks--) {
1900 set_pageblock_migratetype(pageblock_page, migratetype);
1901 pageblock_page += pageblock_nr_pages;
1906 * When we are falling back to another migratetype during allocation, try to
1907 * steal extra free pages from the same pageblocks to satisfy further
1908 * allocations, instead of polluting multiple pageblocks.
1910 * If we are stealing a relatively large buddy page, it is likely there will
1911 * be more free pages in the pageblock, so try to steal them all. For
1912 * reclaimable and unmovable allocations, we steal regardless of page size,
1913 * as fragmentation caused by those allocations polluting movable pageblocks
1914 * is worse than movable allocations stealing from unmovable and reclaimable
1917 static bool can_steal_fallback(unsigned int order, int start_mt)
1920 * Leaving this order check is intended, although there is
1921 * relaxed order check in next check. The reason is that
1922 * we can actually steal whole pageblock if this condition met,
1923 * but, below check doesn't guarantee it and that is just heuristic
1924 * so could be changed anytime.
1926 if (order >= pageblock_order)
1929 if (order >= pageblock_order / 2 ||
1930 start_mt == MIGRATE_RECLAIMABLE ||
1931 start_mt == MIGRATE_UNMOVABLE ||
1932 page_group_by_mobility_disabled)
1939 * This function implements actual steal behaviour. If order is large enough,
1940 * we can steal whole pageblock. If not, we first move freepages in this
1941 * pageblock and check whether half of pages are moved or not. If half of
1942 * pages are moved, we can change migratetype of pageblock and permanently
1943 * use it's pages as requested migratetype in the future.
1945 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1948 unsigned int current_order = page_order(page);
1951 /* Take ownership for orders >= pageblock_order */
1952 if (current_order >= pageblock_order) {
1953 change_pageblock_range(page, current_order, start_type);
1957 pages = move_freepages_block(zone, page, start_type);
1959 /* Claim the whole block if over half of it is free */
1960 if (pages >= (1 << (pageblock_order-1)) ||
1961 page_group_by_mobility_disabled)
1962 set_pageblock_migratetype(page, start_type);
1966 * Check whether there is a suitable fallback freepage with requested order.
1967 * If only_stealable is true, this function returns fallback_mt only if
1968 * we can steal other freepages all together. This would help to reduce
1969 * fragmentation due to mixed migratetype pages in one pageblock.
1971 int find_suitable_fallback(struct free_area *area, unsigned int order,
1972 int migratetype, bool only_stealable, bool *can_steal)
1977 if (area->nr_free == 0)
1982 fallback_mt = fallbacks[migratetype][i];
1983 if (fallback_mt == MIGRATE_TYPES)
1986 if (list_empty(&area->free_list[fallback_mt]))
1989 if (can_steal_fallback(order, migratetype))
1992 if (!only_stealable)
2003 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2004 * there are no empty page blocks that contain a page with a suitable order
2006 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2007 unsigned int alloc_order)
2010 unsigned long max_managed, flags;
2013 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2014 * Check is race-prone but harmless.
2016 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2017 if (zone->nr_reserved_highatomic >= max_managed)
2020 spin_lock_irqsave(&zone->lock, flags);
2022 /* Recheck the nr_reserved_highatomic limit under the lock */
2023 if (zone->nr_reserved_highatomic >= max_managed)
2027 mt = get_pageblock_migratetype(page);
2028 if (mt != MIGRATE_HIGHATOMIC &&
2029 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2030 zone->nr_reserved_highatomic += pageblock_nr_pages;
2031 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2032 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2036 spin_unlock_irqrestore(&zone->lock, flags);
2040 * Used when an allocation is about to fail under memory pressure. This
2041 * potentially hurts the reliability of high-order allocations when under
2042 * intense memory pressure but failed atomic allocations should be easier
2043 * to recover from than an OOM.
2045 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2047 struct zonelist *zonelist = ac->zonelist;
2048 unsigned long flags;
2054 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2056 /* Preserve at least one pageblock */
2057 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2060 spin_lock_irqsave(&zone->lock, flags);
2061 for (order = 0; order < MAX_ORDER; order++) {
2062 struct free_area *area = &(zone->free_area[order]);
2064 page = list_first_entry_or_null(
2065 &area->free_list[MIGRATE_HIGHATOMIC],
2071 * It should never happen but changes to locking could
2072 * inadvertently allow a per-cpu drain to add pages
2073 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2074 * and watch for underflows.
2076 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2077 zone->nr_reserved_highatomic);
2080 * Convert to ac->migratetype and avoid the normal
2081 * pageblock stealing heuristics. Minimally, the caller
2082 * is doing the work and needs the pages. More
2083 * importantly, if the block was always converted to
2084 * MIGRATE_UNMOVABLE or another type then the number
2085 * of pageblocks that cannot be completely freed
2088 set_pageblock_migratetype(page, ac->migratetype);
2089 move_freepages_block(zone, page, ac->migratetype);
2090 spin_unlock_irqrestore(&zone->lock, flags);
2093 spin_unlock_irqrestore(&zone->lock, flags);
2097 /* Remove an element from the buddy allocator from the fallback list */
2098 static inline struct page *
2099 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2101 struct free_area *area;
2102 unsigned int current_order;
2107 /* Find the largest possible block of pages in the other list */
2108 for (current_order = MAX_ORDER-1;
2109 current_order >= order && current_order <= MAX_ORDER-1;
2111 area = &(zone->free_area[current_order]);
2112 fallback_mt = find_suitable_fallback(area, current_order,
2113 start_migratetype, false, &can_steal);
2114 if (fallback_mt == -1)
2117 page = list_first_entry(&area->free_list[fallback_mt],
2120 steal_suitable_fallback(zone, page, start_migratetype);
2122 /* Remove the page from the freelists */
2124 list_del(&page->lru);
2125 rmv_page_order(page);
2127 expand(zone, page, order, current_order, area,
2130 * The pcppage_migratetype may differ from pageblock's
2131 * migratetype depending on the decisions in
2132 * find_suitable_fallback(). This is OK as long as it does not
2133 * differ for MIGRATE_CMA pageblocks. Those can be used as
2134 * fallback only via special __rmqueue_cma_fallback() function
2136 set_pcppage_migratetype(page, start_migratetype);
2138 trace_mm_page_alloc_extfrag(page, order, current_order,
2139 start_migratetype, fallback_mt);
2148 * Do the hard work of removing an element from the buddy allocator.
2149 * Call me with the zone->lock already held.
2151 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2156 page = __rmqueue_smallest(zone, order, migratetype);
2157 if (unlikely(!page)) {
2158 if (migratetype == MIGRATE_MOVABLE)
2159 page = __rmqueue_cma_fallback(zone, order);
2162 page = __rmqueue_fallback(zone, order, migratetype);
2165 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2170 * Obtain a specified number of elements from the buddy allocator, all under
2171 * a single hold of the lock, for efficiency. Add them to the supplied list.
2172 * Returns the number of new pages which were placed at *list.
2174 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2175 unsigned long count, struct list_head *list,
2176 int migratetype, bool cold)
2180 spin_lock(&zone->lock);
2181 for (i = 0; i < count; ++i) {
2182 struct page *page = __rmqueue(zone, order, migratetype);
2183 if (unlikely(page == NULL))
2186 if (unlikely(check_pcp_refill(page)))
2190 * Split buddy pages returned by expand() are received here
2191 * in physical page order. The page is added to the callers and
2192 * list and the list head then moves forward. From the callers
2193 * perspective, the linked list is ordered by page number in
2194 * some conditions. This is useful for IO devices that can
2195 * merge IO requests if the physical pages are ordered
2199 list_add(&page->lru, list);
2201 list_add_tail(&page->lru, list);
2203 if (is_migrate_cma(get_pcppage_migratetype(page)))
2204 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2207 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2208 spin_unlock(&zone->lock);
2214 * Called from the vmstat counter updater to drain pagesets of this
2215 * currently executing processor on remote nodes after they have
2218 * Note that this function must be called with the thread pinned to
2219 * a single processor.
2221 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2223 unsigned long flags;
2224 int to_drain, batch;
2226 local_irq_save(flags);
2227 batch = READ_ONCE(pcp->batch);
2228 to_drain = min(pcp->count, batch);
2230 free_pcppages_bulk(zone, to_drain, pcp);
2231 pcp->count -= to_drain;
2233 local_irq_restore(flags);
2238 * Drain pcplists of the indicated processor and zone.
2240 * The processor must either be the current processor and the
2241 * thread pinned to the current processor or a processor that
2244 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2246 unsigned long flags;
2247 struct per_cpu_pageset *pset;
2248 struct per_cpu_pages *pcp;
2250 local_irq_save(flags);
2251 pset = per_cpu_ptr(zone->pageset, cpu);
2255 free_pcppages_bulk(zone, pcp->count, pcp);
2258 local_irq_restore(flags);
2262 * Drain pcplists of all zones on the indicated processor.
2264 * The processor must either be the current processor and the
2265 * thread pinned to the current processor or a processor that
2268 static void drain_pages(unsigned int cpu)
2272 for_each_populated_zone(zone) {
2273 drain_pages_zone(cpu, zone);
2278 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2280 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2281 * the single zone's pages.
2283 void drain_local_pages(struct zone *zone)
2285 int cpu = smp_processor_id();
2288 drain_pages_zone(cpu, zone);
2294 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2296 * When zone parameter is non-NULL, spill just the single zone's pages.
2298 * Note that this code is protected against sending an IPI to an offline
2299 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2300 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2301 * nothing keeps CPUs from showing up after we populated the cpumask and
2302 * before the call to on_each_cpu_mask().
2304 void drain_all_pages(struct zone *zone)
2309 * Allocate in the BSS so we wont require allocation in
2310 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2312 static cpumask_t cpus_with_pcps;
2315 * We don't care about racing with CPU hotplug event
2316 * as offline notification will cause the notified
2317 * cpu to drain that CPU pcps and on_each_cpu_mask
2318 * disables preemption as part of its processing
2320 for_each_online_cpu(cpu) {
2321 struct per_cpu_pageset *pcp;
2323 bool has_pcps = false;
2326 pcp = per_cpu_ptr(zone->pageset, cpu);
2330 for_each_populated_zone(z) {
2331 pcp = per_cpu_ptr(z->pageset, cpu);
2332 if (pcp->pcp.count) {
2340 cpumask_set_cpu(cpu, &cpus_with_pcps);
2342 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2344 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2348 #ifdef CONFIG_HIBERNATION
2350 void mark_free_pages(struct zone *zone)
2352 unsigned long pfn, max_zone_pfn;
2353 unsigned long flags;
2354 unsigned int order, t;
2357 if (zone_is_empty(zone))
2360 spin_lock_irqsave(&zone->lock, flags);
2362 max_zone_pfn = zone_end_pfn(zone);
2363 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2364 if (pfn_valid(pfn)) {
2365 page = pfn_to_page(pfn);
2367 if (page_zone(page) != zone)
2370 if (!swsusp_page_is_forbidden(page))
2371 swsusp_unset_page_free(page);
2374 for_each_migratetype_order(order, t) {
2375 list_for_each_entry(page,
2376 &zone->free_area[order].free_list[t], lru) {
2379 pfn = page_to_pfn(page);
2380 for (i = 0; i < (1UL << order); i++)
2381 swsusp_set_page_free(pfn_to_page(pfn + i));
2384 spin_unlock_irqrestore(&zone->lock, flags);
2386 #endif /* CONFIG_PM */
2389 * Free a 0-order page
2390 * cold == true ? free a cold page : free a hot page
2392 void free_hot_cold_page(struct page *page, bool cold)
2394 struct zone *zone = page_zone(page);
2395 struct per_cpu_pages *pcp;
2396 unsigned long flags;
2397 unsigned long pfn = page_to_pfn(page);
2400 if (!free_pcp_prepare(page))
2403 migratetype = get_pfnblock_migratetype(page, pfn);
2404 set_pcppage_migratetype(page, migratetype);
2405 local_irq_save(flags);
2406 __count_vm_event(PGFREE);
2409 * We only track unmovable, reclaimable and movable on pcp lists.
2410 * Free ISOLATE pages back to the allocator because they are being
2411 * offlined but treat RESERVE as movable pages so we can get those
2412 * areas back if necessary. Otherwise, we may have to free
2413 * excessively into the page allocator
2415 if (migratetype >= MIGRATE_PCPTYPES) {
2416 if (unlikely(is_migrate_isolate(migratetype))) {
2417 free_one_page(zone, page, pfn, 0, migratetype);
2420 migratetype = MIGRATE_MOVABLE;
2423 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2425 list_add(&page->lru, &pcp->lists[migratetype]);
2427 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2429 if (pcp->count >= pcp->high) {
2430 unsigned long batch = READ_ONCE(pcp->batch);
2431 free_pcppages_bulk(zone, batch, pcp);
2432 pcp->count -= batch;
2436 local_irq_restore(flags);
2440 * Free a list of 0-order pages
2442 void free_hot_cold_page_list(struct list_head *list, bool cold)
2444 struct page *page, *next;
2446 list_for_each_entry_safe(page, next, list, lru) {
2447 trace_mm_page_free_batched(page, cold);
2448 free_hot_cold_page(page, cold);
2453 * split_page takes a non-compound higher-order page, and splits it into
2454 * n (1<<order) sub-pages: page[0..n]
2455 * Each sub-page must be freed individually.
2457 * Note: this is probably too low level an operation for use in drivers.
2458 * Please consult with lkml before using this in your driver.
2460 void split_page(struct page *page, unsigned int order)
2464 VM_BUG_ON_PAGE(PageCompound(page), page);
2465 VM_BUG_ON_PAGE(!page_count(page), page);
2467 #ifdef CONFIG_KMEMCHECK
2469 * Split shadow pages too, because free(page[0]) would
2470 * otherwise free the whole shadow.
2472 if (kmemcheck_page_is_tracked(page))
2473 split_page(virt_to_page(page[0].shadow), order);
2476 for (i = 1; i < (1 << order); i++)
2477 set_page_refcounted(page + i);
2478 split_page_owner(page, order);
2480 EXPORT_SYMBOL_GPL(split_page);
2482 int __isolate_free_page(struct page *page, unsigned int order)
2484 unsigned long watermark;
2488 BUG_ON(!PageBuddy(page));
2490 zone = page_zone(page);
2491 mt = get_pageblock_migratetype(page);
2493 if (!is_migrate_isolate(mt)) {
2494 /* Obey watermarks as if the page was being allocated */
2495 watermark = low_wmark_pages(zone) + (1 << order);
2496 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2499 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2502 /* Remove page from free list */
2503 list_del(&page->lru);
2504 zone->free_area[order].nr_free--;
2505 rmv_page_order(page);
2508 * Set the pageblock if the isolated page is at least half of a
2511 if (order >= pageblock_order - 1) {
2512 struct page *endpage = page + (1 << order) - 1;
2513 for (; page < endpage; page += pageblock_nr_pages) {
2514 int mt = get_pageblock_migratetype(page);
2515 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2516 set_pageblock_migratetype(page,
2522 return 1UL << order;
2526 * Update NUMA hit/miss statistics
2528 * Must be called with interrupts disabled.
2530 * When __GFP_OTHER_NODE is set assume the node of the preferred
2531 * zone is the local node. This is useful for daemons who allocate
2532 * memory on behalf of other processes.
2534 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2538 int local_nid = numa_node_id();
2539 enum zone_stat_item local_stat = NUMA_LOCAL;
2541 if (unlikely(flags & __GFP_OTHER_NODE)) {
2542 local_stat = NUMA_OTHER;
2543 local_nid = preferred_zone->node;
2546 if (z->node == local_nid) {
2547 __inc_zone_state(z, NUMA_HIT);
2548 __inc_zone_state(z, local_stat);
2550 __inc_zone_state(z, NUMA_MISS);
2551 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2557 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2560 struct page *buffered_rmqueue(struct zone *preferred_zone,
2561 struct zone *zone, unsigned int order,
2562 gfp_t gfp_flags, unsigned int alloc_flags,
2565 unsigned long flags;
2567 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2569 if (likely(order == 0)) {
2570 struct per_cpu_pages *pcp;
2571 struct list_head *list;
2573 local_irq_save(flags);
2575 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2576 list = &pcp->lists[migratetype];
2577 if (list_empty(list)) {
2578 pcp->count += rmqueue_bulk(zone, 0,
2581 if (unlikely(list_empty(list)))
2586 page = list_last_entry(list, struct page, lru);
2588 page = list_first_entry(list, struct page, lru);
2590 __dec_zone_state(zone, NR_ALLOC_BATCH);
2591 list_del(&page->lru);
2594 } while (check_new_pcp(page));
2597 * We most definitely don't want callers attempting to
2598 * allocate greater than order-1 page units with __GFP_NOFAIL.
2600 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2601 spin_lock_irqsave(&zone->lock, flags);
2605 if (alloc_flags & ALLOC_HARDER) {
2606 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2608 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2611 page = __rmqueue(zone, order, migratetype);
2612 } while (page && check_new_pages(page, order));
2613 spin_unlock(&zone->lock);
2616 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2617 __mod_zone_freepage_state(zone, -(1 << order),
2618 get_pcppage_migratetype(page));
2621 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2622 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2623 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2625 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2626 zone_statistics(preferred_zone, zone, gfp_flags);
2627 local_irq_restore(flags);
2629 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2633 local_irq_restore(flags);
2637 #ifdef CONFIG_FAIL_PAGE_ALLOC
2640 struct fault_attr attr;
2642 bool ignore_gfp_highmem;
2643 bool ignore_gfp_reclaim;
2645 } fail_page_alloc = {
2646 .attr = FAULT_ATTR_INITIALIZER,
2647 .ignore_gfp_reclaim = true,
2648 .ignore_gfp_highmem = true,
2652 static int __init setup_fail_page_alloc(char *str)
2654 return setup_fault_attr(&fail_page_alloc.attr, str);
2656 __setup("fail_page_alloc=", setup_fail_page_alloc);
2658 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2660 if (order < fail_page_alloc.min_order)
2662 if (gfp_mask & __GFP_NOFAIL)
2664 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2666 if (fail_page_alloc.ignore_gfp_reclaim &&
2667 (gfp_mask & __GFP_DIRECT_RECLAIM))
2670 return should_fail(&fail_page_alloc.attr, 1 << order);
2673 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2675 static int __init fail_page_alloc_debugfs(void)
2677 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2680 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2681 &fail_page_alloc.attr);
2683 return PTR_ERR(dir);
2685 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2686 &fail_page_alloc.ignore_gfp_reclaim))
2688 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2689 &fail_page_alloc.ignore_gfp_highmem))
2691 if (!debugfs_create_u32("min-order", mode, dir,
2692 &fail_page_alloc.min_order))
2697 debugfs_remove_recursive(dir);
2702 late_initcall(fail_page_alloc_debugfs);
2704 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2706 #else /* CONFIG_FAIL_PAGE_ALLOC */
2708 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2713 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2716 * Return true if free base pages are above 'mark'. For high-order checks it
2717 * will return true of the order-0 watermark is reached and there is at least
2718 * one free page of a suitable size. Checking now avoids taking the zone lock
2719 * to check in the allocation paths if no pages are free.
2721 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2722 int classzone_idx, unsigned int alloc_flags,
2727 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2729 /* free_pages may go negative - that's OK */
2730 free_pages -= (1 << order) - 1;
2732 if (alloc_flags & ALLOC_HIGH)
2736 * If the caller does not have rights to ALLOC_HARDER then subtract
2737 * the high-atomic reserves. This will over-estimate the size of the
2738 * atomic reserve but it avoids a search.
2740 if (likely(!alloc_harder))
2741 free_pages -= z->nr_reserved_highatomic;
2746 /* If allocation can't use CMA areas don't use free CMA pages */
2747 if (!(alloc_flags & ALLOC_CMA))
2748 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2752 * Check watermarks for an order-0 allocation request. If these
2753 * are not met, then a high-order request also cannot go ahead
2754 * even if a suitable page happened to be free.
2756 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2759 /* If this is an order-0 request then the watermark is fine */
2763 /* For a high-order request, check at least one suitable page is free */
2764 for (o = order; o < MAX_ORDER; o++) {
2765 struct free_area *area = &z->free_area[o];
2774 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2775 if (!list_empty(&area->free_list[mt]))
2780 if ((alloc_flags & ALLOC_CMA) &&
2781 !list_empty(&area->free_list[MIGRATE_CMA])) {
2789 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2790 int classzone_idx, unsigned int alloc_flags)
2792 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2793 zone_page_state(z, NR_FREE_PAGES));
2796 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2797 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2799 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2803 /* If allocation can't use CMA areas don't use free CMA pages */
2804 if (!(alloc_flags & ALLOC_CMA))
2805 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2809 * Fast check for order-0 only. If this fails then the reserves
2810 * need to be calculated. There is a corner case where the check
2811 * passes but only the high-order atomic reserve are free. If
2812 * the caller is !atomic then it'll uselessly search the free
2813 * list. That corner case is then slower but it is harmless.
2815 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2818 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2822 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2823 unsigned long mark, int classzone_idx)
2825 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2827 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2828 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2830 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2835 static bool zone_local(struct zone *local_zone, struct zone *zone)
2837 return local_zone->node == zone->node;
2840 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2842 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2845 #else /* CONFIG_NUMA */
2846 static bool zone_local(struct zone *local_zone, struct zone *zone)
2851 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2855 #endif /* CONFIG_NUMA */
2857 static void reset_alloc_batches(struct zone *preferred_zone)
2859 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2862 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2863 high_wmark_pages(zone) - low_wmark_pages(zone) -
2864 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2865 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2866 } while (zone++ != preferred_zone);
2870 * get_page_from_freelist goes through the zonelist trying to allocate
2873 static struct page *
2874 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2875 const struct alloc_context *ac)
2877 struct zoneref *z = ac->preferred_zoneref;
2879 bool fair_skipped = false;
2880 bool apply_fair = (alloc_flags & ALLOC_FAIR);
2884 * Scan zonelist, looking for a zone with enough free.
2885 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2887 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2892 if (cpusets_enabled() &&
2893 (alloc_flags & ALLOC_CPUSET) &&
2894 !__cpuset_zone_allowed(zone, gfp_mask))
2897 * Distribute pages in proportion to the individual
2898 * zone size to ensure fair page aging. The zone a
2899 * page was allocated in should have no effect on the
2900 * time the page has in memory before being reclaimed.
2903 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2904 fair_skipped = true;
2907 if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2914 * When allocating a page cache page for writing, we
2915 * want to get it from a node that is within its dirty
2916 * limit, such that no single node holds more than its
2917 * proportional share of globally allowed dirty pages.
2918 * The dirty limits take into account the node's
2919 * lowmem reserves and high watermark so that kswapd
2920 * should be able to balance it without having to
2921 * write pages from its LRU list.
2923 * XXX: For now, allow allocations to potentially
2924 * exceed the per-node dirty limit in the slowpath
2925 * (spread_dirty_pages unset) before going into reclaim,
2926 * which is important when on a NUMA setup the allowed
2927 * nodes are together not big enough to reach the
2928 * global limit. The proper fix for these situations
2929 * will require awareness of nodes in the
2930 * dirty-throttling and the flusher threads.
2932 if (ac->spread_dirty_pages && !node_dirty_ok(zone->zone_pgdat))
2935 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2936 if (!zone_watermark_fast(zone, order, mark,
2937 ac_classzone_idx(ac), alloc_flags)) {
2940 /* Checked here to keep the fast path fast */
2941 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2942 if (alloc_flags & ALLOC_NO_WATERMARKS)
2945 if (zone_reclaim_mode == 0 ||
2946 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2949 ret = zone_reclaim(zone, gfp_mask, order);
2951 case ZONE_RECLAIM_NOSCAN:
2954 case ZONE_RECLAIM_FULL:
2955 /* scanned but unreclaimable */
2958 /* did we reclaim enough */
2959 if (zone_watermark_ok(zone, order, mark,
2960 ac_classzone_idx(ac), alloc_flags))
2968 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2969 gfp_mask, alloc_flags, ac->migratetype);
2971 prep_new_page(page, order, gfp_mask, alloc_flags);
2974 * If this is a high-order atomic allocation then check
2975 * if the pageblock should be reserved for the future
2977 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2978 reserve_highatomic_pageblock(page, zone, order);
2985 * The first pass makes sure allocations are spread fairly within the
2986 * local node. However, the local node might have free pages left
2987 * after the fairness batches are exhausted, and remote zones haven't
2988 * even been considered yet. Try once more without fairness, and
2989 * include remote zones now, before entering the slowpath and waking
2990 * kswapd: prefer spilling to a remote zone over swapping locally.
2995 fair_skipped = false;
2996 reset_alloc_batches(ac->preferred_zoneref->zone);
2997 z = ac->preferred_zoneref;
3005 * Large machines with many possible nodes should not always dump per-node
3006 * meminfo in irq context.
3008 static inline bool should_suppress_show_mem(void)
3013 ret = in_interrupt();
3018 static DEFINE_RATELIMIT_STATE(nopage_rs,
3019 DEFAULT_RATELIMIT_INTERVAL,
3020 DEFAULT_RATELIMIT_BURST);
3022 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3024 unsigned int filter = SHOW_MEM_FILTER_NODES;
3026 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3027 debug_guardpage_minorder() > 0)
3031 * This documents exceptions given to allocations in certain
3032 * contexts that are allowed to allocate outside current's set
3035 if (!(gfp_mask & __GFP_NOMEMALLOC))
3036 if (test_thread_flag(TIF_MEMDIE) ||
3037 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3038 filter &= ~SHOW_MEM_FILTER_NODES;
3039 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3040 filter &= ~SHOW_MEM_FILTER_NODES;
3043 struct va_format vaf;
3046 va_start(args, fmt);
3051 pr_warn("%pV", &vaf);
3056 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3057 current->comm, order, gfp_mask, &gfp_mask);
3059 if (!should_suppress_show_mem())
3063 static inline struct page *
3064 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3065 const struct alloc_context *ac, unsigned long *did_some_progress)
3067 struct oom_control oc = {
3068 .zonelist = ac->zonelist,
3069 .nodemask = ac->nodemask,
3071 .gfp_mask = gfp_mask,
3076 *did_some_progress = 0;
3079 * Acquire the oom lock. If that fails, somebody else is
3080 * making progress for us.
3082 if (!mutex_trylock(&oom_lock)) {
3083 *did_some_progress = 1;
3084 schedule_timeout_uninterruptible(1);
3089 * Go through the zonelist yet one more time, keep very high watermark
3090 * here, this is only to catch a parallel oom killing, we must fail if
3091 * we're still under heavy pressure.
3093 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3094 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3098 if (!(gfp_mask & __GFP_NOFAIL)) {
3099 /* Coredumps can quickly deplete all memory reserves */
3100 if (current->flags & PF_DUMPCORE)
3102 /* The OOM killer will not help higher order allocs */
3103 if (order > PAGE_ALLOC_COSTLY_ORDER)
3105 /* The OOM killer does not needlessly kill tasks for lowmem */
3106 if (ac->high_zoneidx < ZONE_NORMAL)
3108 if (pm_suspended_storage())
3111 * XXX: GFP_NOFS allocations should rather fail than rely on
3112 * other request to make a forward progress.
3113 * We are in an unfortunate situation where out_of_memory cannot
3114 * do much for this context but let's try it to at least get
3115 * access to memory reserved if the current task is killed (see
3116 * out_of_memory). Once filesystems are ready to handle allocation
3117 * failures more gracefully we should just bail out here.
3120 /* The OOM killer may not free memory on a specific node */
3121 if (gfp_mask & __GFP_THISNODE)
3124 /* Exhausted what can be done so it's blamo time */
3125 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3126 *did_some_progress = 1;
3128 if (gfp_mask & __GFP_NOFAIL) {
3129 page = get_page_from_freelist(gfp_mask, order,
3130 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3132 * fallback to ignore cpuset restriction if our nodes
3136 page = get_page_from_freelist(gfp_mask, order,
3137 ALLOC_NO_WATERMARKS, ac);
3141 mutex_unlock(&oom_lock);
3147 * Maximum number of compaction retries wit a progress before OOM
3148 * killer is consider as the only way to move forward.
3150 #define MAX_COMPACT_RETRIES 16
3152 #ifdef CONFIG_COMPACTION
3153 /* Try memory compaction for high-order allocations before reclaim */
3154 static struct page *
3155 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3156 unsigned int alloc_flags, const struct alloc_context *ac,
3157 enum migrate_mode mode, enum compact_result *compact_result)
3160 int contended_compaction;
3165 current->flags |= PF_MEMALLOC;
3166 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3167 mode, &contended_compaction);
3168 current->flags &= ~PF_MEMALLOC;
3170 if (*compact_result <= COMPACT_INACTIVE)
3174 * At least in one zone compaction wasn't deferred or skipped, so let's
3175 * count a compaction stall
3177 count_vm_event(COMPACTSTALL);
3179 page = get_page_from_freelist(gfp_mask, order,
3180 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3183 struct zone *zone = page_zone(page);
3185 zone->compact_blockskip_flush = false;
3186 compaction_defer_reset(zone, order, true);
3187 count_vm_event(COMPACTSUCCESS);
3192 * It's bad if compaction run occurs and fails. The most likely reason
3193 * is that pages exist, but not enough to satisfy watermarks.
3195 count_vm_event(COMPACTFAIL);
3198 * In all zones where compaction was attempted (and not
3199 * deferred or skipped), lock contention has been detected.
3200 * For THP allocation we do not want to disrupt the others
3201 * so we fallback to base pages instead.
3203 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3204 *compact_result = COMPACT_CONTENDED;
3207 * If compaction was aborted due to need_resched(), we do not
3208 * want to further increase allocation latency, unless it is
3209 * khugepaged trying to collapse.
3211 if (contended_compaction == COMPACT_CONTENDED_SCHED
3212 && !(current->flags & PF_KTHREAD))
3213 *compact_result = COMPACT_CONTENDED;
3221 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3222 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3223 int compaction_retries)
3225 int max_retries = MAX_COMPACT_RETRIES;
3231 * compaction considers all the zone as desperately out of memory
3232 * so it doesn't really make much sense to retry except when the
3233 * failure could be caused by weak migration mode.
3235 if (compaction_failed(compact_result)) {
3236 if (*migrate_mode == MIGRATE_ASYNC) {
3237 *migrate_mode = MIGRATE_SYNC_LIGHT;
3244 * make sure the compaction wasn't deferred or didn't bail out early
3245 * due to locks contention before we declare that we should give up.
3246 * But do not retry if the given zonelist is not suitable for
3249 if (compaction_withdrawn(compact_result))
3250 return compaction_zonelist_suitable(ac, order, alloc_flags);
3253 * !costly requests are much more important than __GFP_REPEAT
3254 * costly ones because they are de facto nofail and invoke OOM
3255 * killer to move on while costly can fail and users are ready
3256 * to cope with that. 1/4 retries is rather arbitrary but we
3257 * would need much more detailed feedback from compaction to
3258 * make a better decision.
3260 if (order > PAGE_ALLOC_COSTLY_ORDER)
3262 if (compaction_retries <= max_retries)
3268 static inline struct page *
3269 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3270 unsigned int alloc_flags, const struct alloc_context *ac,
3271 enum migrate_mode mode, enum compact_result *compact_result)
3273 *compact_result = COMPACT_SKIPPED;
3278 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3279 enum compact_result compact_result,
3280 enum migrate_mode *migrate_mode,
3281 int compaction_retries)
3286 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3290 * There are setups with compaction disabled which would prefer to loop
3291 * inside the allocator rather than hit the oom killer prematurely.
3292 * Let's give them a good hope and keep retrying while the order-0
3293 * watermarks are OK.
3295 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3297 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3298 ac_classzone_idx(ac), alloc_flags))
3303 #endif /* CONFIG_COMPACTION */
3305 /* Perform direct synchronous page reclaim */
3307 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3308 const struct alloc_context *ac)
3310 struct reclaim_state reclaim_state;
3315 /* We now go into synchronous reclaim */
3316 cpuset_memory_pressure_bump();
3317 current->flags |= PF_MEMALLOC;
3318 lockdep_set_current_reclaim_state(gfp_mask);
3319 reclaim_state.reclaimed_slab = 0;
3320 current->reclaim_state = &reclaim_state;
3322 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3325 current->reclaim_state = NULL;
3326 lockdep_clear_current_reclaim_state();
3327 current->flags &= ~PF_MEMALLOC;
3334 /* The really slow allocator path where we enter direct reclaim */
3335 static inline struct page *
3336 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3337 unsigned int alloc_flags, const struct alloc_context *ac,
3338 unsigned long *did_some_progress)
3340 struct page *page = NULL;
3341 bool drained = false;
3343 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3344 if (unlikely(!(*did_some_progress)))
3348 page = get_page_from_freelist(gfp_mask, order,
3349 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3352 * If an allocation failed after direct reclaim, it could be because
3353 * pages are pinned on the per-cpu lists or in high alloc reserves.
3354 * Shrink them them and try again
3356 if (!page && !drained) {
3357 unreserve_highatomic_pageblock(ac);
3358 drain_all_pages(NULL);
3366 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3370 pg_data_t *last_pgdat = NULL;
3372 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3373 ac->high_zoneidx, ac->nodemask) {
3374 if (last_pgdat != zone->zone_pgdat)
3375 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3376 last_pgdat = zone->zone_pgdat;
3380 static inline unsigned int
3381 gfp_to_alloc_flags(gfp_t gfp_mask)
3383 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3385 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3386 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3389 * The caller may dip into page reserves a bit more if the caller
3390 * cannot run direct reclaim, or if the caller has realtime scheduling
3391 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3392 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3394 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3396 if (gfp_mask & __GFP_ATOMIC) {
3398 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3399 * if it can't schedule.
3401 if (!(gfp_mask & __GFP_NOMEMALLOC))
3402 alloc_flags |= ALLOC_HARDER;
3404 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3405 * comment for __cpuset_node_allowed().
3407 alloc_flags &= ~ALLOC_CPUSET;
3408 } else if (unlikely(rt_task(current)) && !in_interrupt())
3409 alloc_flags |= ALLOC_HARDER;
3411 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3412 if (gfp_mask & __GFP_MEMALLOC)
3413 alloc_flags |= ALLOC_NO_WATERMARKS;
3414 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3415 alloc_flags |= ALLOC_NO_WATERMARKS;
3416 else if (!in_interrupt() &&
3417 ((current->flags & PF_MEMALLOC) ||
3418 unlikely(test_thread_flag(TIF_MEMDIE))))
3419 alloc_flags |= ALLOC_NO_WATERMARKS;
3422 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3423 alloc_flags |= ALLOC_CMA;
3428 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3430 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3433 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3435 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3439 * Maximum number of reclaim retries without any progress before OOM killer
3440 * is consider as the only way to move forward.
3442 #define MAX_RECLAIM_RETRIES 16
3445 * Checks whether it makes sense to retry the reclaim to make a forward progress
3446 * for the given allocation request.
3447 * The reclaim feedback represented by did_some_progress (any progress during
3448 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3449 * any progress in a row) is considered as well as the reclaimable pages on the
3450 * applicable zone list (with a backoff mechanism which is a function of
3451 * no_progress_loops).
3453 * Returns true if a retry is viable or false to enter the oom path.
3456 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3457 struct alloc_context *ac, int alloc_flags,
3458 bool did_some_progress, int no_progress_loops)
3464 * Make sure we converge to OOM if we cannot make any progress
3465 * several times in the row.
3467 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3471 * Keep reclaiming pages while there is a chance this will lead somewhere.
3472 * If none of the target zones can satisfy our allocation request even
3473 * if all reclaimable pages are considered then we are screwed and have
3476 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3478 unsigned long available;
3479 unsigned long reclaimable;
3481 available = reclaimable = zone_reclaimable_pages(zone);
3482 available -= DIV_ROUND_UP(no_progress_loops * available,
3483 MAX_RECLAIM_RETRIES);
3484 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3487 * Would the allocation succeed if we reclaimed the whole
3490 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3491 ac_classzone_idx(ac), alloc_flags, available)) {
3493 * If we didn't make any progress and have a lot of
3494 * dirty + writeback pages then we should wait for
3495 * an IO to complete to slow down the reclaim and
3496 * prevent from pre mature OOM
3498 if (!did_some_progress) {
3499 unsigned long write_pending;
3501 write_pending = zone_page_state_snapshot(zone,
3502 NR_ZONE_WRITE_PENDING);
3504 if (2 * write_pending > reclaimable) {
3505 congestion_wait(BLK_RW_ASYNC, HZ/10);
3511 * Memory allocation/reclaim might be called from a WQ
3512 * context and the current implementation of the WQ
3513 * concurrency control doesn't recognize that
3514 * a particular WQ is congested if the worker thread is
3515 * looping without ever sleeping. Therefore we have to
3516 * do a short sleep here rather than calling
3519 if (current->flags & PF_WQ_WORKER)
3520 schedule_timeout_uninterruptible(1);
3531 static inline struct page *
3532 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3533 struct alloc_context *ac)
3535 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3536 struct page *page = NULL;
3537 unsigned int alloc_flags;
3538 unsigned long did_some_progress;
3539 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3540 enum compact_result compact_result;
3541 int compaction_retries = 0;
3542 int no_progress_loops = 0;
3545 * In the slowpath, we sanity check order to avoid ever trying to
3546 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3547 * be using allocators in order of preference for an area that is
3550 if (order >= MAX_ORDER) {
3551 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3556 * We also sanity check to catch abuse of atomic reserves being used by
3557 * callers that are not in atomic context.
3559 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3560 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3561 gfp_mask &= ~__GFP_ATOMIC;
3564 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3565 wake_all_kswapds(order, ac);
3568 * OK, we're below the kswapd watermark and have kicked background
3569 * reclaim. Now things get more complex, so set up alloc_flags according
3570 * to how we want to proceed.
3572 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3575 * Reset the zonelist iterators if memory policies can be ignored.
3576 * These allocations are high priority and system rather than user
3579 if ((alloc_flags & ALLOC_NO_WATERMARKS) || !(alloc_flags & ALLOC_CPUSET)) {
3580 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3581 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3582 ac->high_zoneidx, ac->nodemask);
3585 /* This is the last chance, in general, before the goto nopage. */
3586 page = get_page_from_freelist(gfp_mask, order,
3587 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3591 /* Allocate without watermarks if the context allows */
3592 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3593 page = get_page_from_freelist(gfp_mask, order,
3594 ALLOC_NO_WATERMARKS, ac);
3599 /* Caller is not willing to reclaim, we can't balance anything */
3600 if (!can_direct_reclaim) {
3602 * All existing users of the __GFP_NOFAIL are blockable, so warn
3603 * of any new users that actually allow this type of allocation
3606 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3610 /* Avoid recursion of direct reclaim */
3611 if (current->flags & PF_MEMALLOC) {
3613 * __GFP_NOFAIL request from this context is rather bizarre
3614 * because we cannot reclaim anything and only can loop waiting
3615 * for somebody to do a work for us.
3617 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3624 /* Avoid allocations with no watermarks from looping endlessly */
3625 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3629 * Try direct compaction. The first pass is asynchronous. Subsequent
3630 * attempts after direct reclaim are synchronous
3632 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3638 /* Checks for THP-specific high-order allocations */
3639 if (is_thp_gfp_mask(gfp_mask)) {
3641 * If compaction is deferred for high-order allocations, it is
3642 * because sync compaction recently failed. If this is the case
3643 * and the caller requested a THP allocation, we do not want
3644 * to heavily disrupt the system, so we fail the allocation
3645 * instead of entering direct reclaim.
3647 if (compact_result == COMPACT_DEFERRED)
3651 * Compaction is contended so rather back off than cause
3654 if(compact_result == COMPACT_CONTENDED)
3658 if (order && compaction_made_progress(compact_result))
3659 compaction_retries++;
3661 /* Try direct reclaim and then allocating */
3662 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3663 &did_some_progress);
3667 /* Do not loop if specifically requested */
3668 if (gfp_mask & __GFP_NORETRY)
3672 * Do not retry costly high order allocations unless they are
3675 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3679 * Costly allocations might have made a progress but this doesn't mean
3680 * their order will become available due to high fragmentation so
3681 * always increment the no progress counter for them
3683 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3684 no_progress_loops = 0;
3686 no_progress_loops++;
3688 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3689 did_some_progress > 0, no_progress_loops))
3693 * It doesn't make any sense to retry for the compaction if the order-0
3694 * reclaim is not able to make any progress because the current
3695 * implementation of the compaction depends on the sufficient amount
3696 * of free memory (see __compaction_suitable)
3698 if (did_some_progress > 0 &&
3699 should_compact_retry(ac, order, alloc_flags,
3700 compact_result, &migration_mode,
3701 compaction_retries))
3704 /* Reclaim has failed us, start killing things */
3705 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3709 /* Retry as long as the OOM killer is making progress */
3710 if (did_some_progress) {
3711 no_progress_loops = 0;
3717 * High-order allocations do not necessarily loop after direct reclaim
3718 * and reclaim/compaction depends on compaction being called after
3719 * reclaim so call directly if necessary.
3720 * It can become very expensive to allocate transparent hugepages at
3721 * fault, so use asynchronous memory compaction for THP unless it is
3722 * khugepaged trying to collapse. All other requests should tolerate
3723 * at least light sync migration.
3725 if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3726 migration_mode = MIGRATE_ASYNC;
3728 migration_mode = MIGRATE_SYNC_LIGHT;
3729 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3735 warn_alloc_failed(gfp_mask, order, NULL);
3741 * This is the 'heart' of the zoned buddy allocator.
3744 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3745 struct zonelist *zonelist, nodemask_t *nodemask)
3748 unsigned int cpuset_mems_cookie;
3749 unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3750 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3751 struct alloc_context ac = {
3752 .high_zoneidx = gfp_zone(gfp_mask),
3753 .zonelist = zonelist,
3754 .nodemask = nodemask,
3755 .migratetype = gfpflags_to_migratetype(gfp_mask),
3758 if (cpusets_enabled()) {
3759 alloc_mask |= __GFP_HARDWALL;
3760 alloc_flags |= ALLOC_CPUSET;
3762 ac.nodemask = &cpuset_current_mems_allowed;
3765 gfp_mask &= gfp_allowed_mask;
3767 lockdep_trace_alloc(gfp_mask);
3769 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3771 if (should_fail_alloc_page(gfp_mask, order))
3775 * Check the zones suitable for the gfp_mask contain at least one
3776 * valid zone. It's possible to have an empty zonelist as a result
3777 * of __GFP_THISNODE and a memoryless node
3779 if (unlikely(!zonelist->_zonerefs->zone))
3782 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3783 alloc_flags |= ALLOC_CMA;
3786 cpuset_mems_cookie = read_mems_allowed_begin();
3788 /* Dirty zone balancing only done in the fast path */
3789 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3792 * The preferred zone is used for statistics but crucially it is
3793 * also used as the starting point for the zonelist iterator. It
3794 * may get reset for allocations that ignore memory policies.
3796 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3797 ac.high_zoneidx, ac.nodemask);
3798 if (!ac.preferred_zoneref) {
3803 /* First allocation attempt */
3804 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3809 * Runtime PM, block IO and its error handling path can deadlock
3810 * because I/O on the device might not complete.
3812 alloc_mask = memalloc_noio_flags(gfp_mask);
3813 ac.spread_dirty_pages = false;
3816 * Restore the original nodemask if it was potentially replaced with
3817 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3819 if (cpusets_enabled())
3820 ac.nodemask = nodemask;
3821 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3825 * When updating a task's mems_allowed, it is possible to race with
3826 * parallel threads in such a way that an allocation can fail while
3827 * the mask is being updated. If a page allocation is about to fail,
3828 * check if the cpuset changed during allocation and if so, retry.
3830 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3831 alloc_mask = gfp_mask;
3836 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page) {
3837 if (unlikely(memcg_kmem_charge(page, gfp_mask, order))) {
3838 __free_pages(page, order);
3841 __SetPageKmemcg(page);
3844 if (kmemcheck_enabled && page)
3845 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3847 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3851 EXPORT_SYMBOL(__alloc_pages_nodemask);
3854 * Common helper functions.
3856 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3861 * __get_free_pages() returns a 32-bit address, which cannot represent
3864 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3866 page = alloc_pages(gfp_mask, order);
3869 return (unsigned long) page_address(page);
3871 EXPORT_SYMBOL(__get_free_pages);
3873 unsigned long get_zeroed_page(gfp_t gfp_mask)
3875 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3877 EXPORT_SYMBOL(get_zeroed_page);
3879 void __free_pages(struct page *page, unsigned int order)
3881 if (put_page_testzero(page)) {
3883 free_hot_cold_page(page, false);
3885 __free_pages_ok(page, order);
3889 EXPORT_SYMBOL(__free_pages);
3891 void free_pages(unsigned long addr, unsigned int order)
3894 VM_BUG_ON(!virt_addr_valid((void *)addr));
3895 __free_pages(virt_to_page((void *)addr), order);
3899 EXPORT_SYMBOL(free_pages);
3903 * An arbitrary-length arbitrary-offset area of memory which resides
3904 * within a 0 or higher order page. Multiple fragments within that page
3905 * are individually refcounted, in the page's reference counter.
3907 * The page_frag functions below provide a simple allocation framework for
3908 * page fragments. This is used by the network stack and network device
3909 * drivers to provide a backing region of memory for use as either an
3910 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3912 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3915 struct page *page = NULL;
3916 gfp_t gfp = gfp_mask;
3918 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3919 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3921 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3922 PAGE_FRAG_CACHE_MAX_ORDER);
3923 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3925 if (unlikely(!page))
3926 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3928 nc->va = page ? page_address(page) : NULL;
3933 void *__alloc_page_frag(struct page_frag_cache *nc,
3934 unsigned int fragsz, gfp_t gfp_mask)
3936 unsigned int size = PAGE_SIZE;
3940 if (unlikely(!nc->va)) {
3942 page = __page_frag_refill(nc, gfp_mask);
3946 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3947 /* if size can vary use size else just use PAGE_SIZE */
3950 /* Even if we own the page, we do not use atomic_set().
3951 * This would break get_page_unless_zero() users.
3953 page_ref_add(page, size - 1);
3955 /* reset page count bias and offset to start of new frag */
3956 nc->pfmemalloc = page_is_pfmemalloc(page);
3957 nc->pagecnt_bias = size;
3961 offset = nc->offset - fragsz;
3962 if (unlikely(offset < 0)) {
3963 page = virt_to_page(nc->va);
3965 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3968 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3969 /* if size can vary use size else just use PAGE_SIZE */
3972 /* OK, page count is 0, we can safely set it */
3973 set_page_count(page, size);
3975 /* reset page count bias and offset to start of new frag */
3976 nc->pagecnt_bias = size;
3977 offset = size - fragsz;
3981 nc->offset = offset;
3983 return nc->va + offset;
3985 EXPORT_SYMBOL(__alloc_page_frag);
3988 * Frees a page fragment allocated out of either a compound or order 0 page.
3990 void __free_page_frag(void *addr)
3992 struct page *page = virt_to_head_page(addr);
3994 if (unlikely(put_page_testzero(page)))
3995 __free_pages_ok(page, compound_order(page));
3997 EXPORT_SYMBOL(__free_page_frag);
3999 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4003 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4004 unsigned long used = addr + PAGE_ALIGN(size);
4006 split_page(virt_to_page((void *)addr), order);
4007 while (used < alloc_end) {
4012 return (void *)addr;
4016 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4017 * @size: the number of bytes to allocate
4018 * @gfp_mask: GFP flags for the allocation
4020 * This function is similar to alloc_pages(), except that it allocates the
4021 * minimum number of pages to satisfy the request. alloc_pages() can only
4022 * allocate memory in power-of-two pages.
4024 * This function is also limited by MAX_ORDER.
4026 * Memory allocated by this function must be released by free_pages_exact().
4028 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4030 unsigned int order = get_order(size);
4033 addr = __get_free_pages(gfp_mask, order);
4034 return make_alloc_exact(addr, order, size);
4036 EXPORT_SYMBOL(alloc_pages_exact);
4039 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4041 * @nid: the preferred node ID where memory should be allocated
4042 * @size: the number of bytes to allocate
4043 * @gfp_mask: GFP flags for the allocation
4045 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4048 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4050 unsigned int order = get_order(size);
4051 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4054 return make_alloc_exact((unsigned long)page_address(p), order, size);
4058 * free_pages_exact - release memory allocated via alloc_pages_exact()
4059 * @virt: the value returned by alloc_pages_exact.
4060 * @size: size of allocation, same value as passed to alloc_pages_exact().
4062 * Release the memory allocated by a previous call to alloc_pages_exact.
4064 void free_pages_exact(void *virt, size_t size)
4066 unsigned long addr = (unsigned long)virt;
4067 unsigned long end = addr + PAGE_ALIGN(size);
4069 while (addr < end) {
4074 EXPORT_SYMBOL(free_pages_exact);
4077 * nr_free_zone_pages - count number of pages beyond high watermark
4078 * @offset: The zone index of the highest zone
4080 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4081 * high watermark within all zones at or below a given zone index. For each
4082 * zone, the number of pages is calculated as:
4083 * managed_pages - high_pages
4085 static unsigned long nr_free_zone_pages(int offset)
4090 /* Just pick one node, since fallback list is circular */
4091 unsigned long sum = 0;
4093 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4095 for_each_zone_zonelist(zone, z, zonelist, offset) {
4096 unsigned long size = zone->managed_pages;
4097 unsigned long high = high_wmark_pages(zone);
4106 * nr_free_buffer_pages - count number of pages beyond high watermark
4108 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4109 * watermark within ZONE_DMA and ZONE_NORMAL.
4111 unsigned long nr_free_buffer_pages(void)
4113 return nr_free_zone_pages(gfp_zone(GFP_USER));
4115 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4118 * nr_free_pagecache_pages - count number of pages beyond high watermark
4120 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4121 * high watermark within all zones.
4123 unsigned long nr_free_pagecache_pages(void)
4125 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4128 static inline void show_node(struct zone *zone)
4130 if (IS_ENABLED(CONFIG_NUMA))
4131 printk("Node %d ", zone_to_nid(zone));
4134 long si_mem_available(void)
4137 unsigned long pagecache;
4138 unsigned long wmark_low = 0;
4139 unsigned long pages[NR_LRU_LISTS];
4143 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4144 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4147 wmark_low += zone->watermark[WMARK_LOW];
4150 * Estimate the amount of memory available for userspace allocations,
4151 * without causing swapping.
4153 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4156 * Not all the page cache can be freed, otherwise the system will
4157 * start swapping. Assume at least half of the page cache, or the
4158 * low watermark worth of cache, needs to stay.
4160 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4161 pagecache -= min(pagecache / 2, wmark_low);
4162 available += pagecache;
4165 * Part of the reclaimable slab consists of items that are in use,
4166 * and cannot be freed. Cap this estimate at the low watermark.
4168 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4169 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4175 EXPORT_SYMBOL_GPL(si_mem_available);
4177 void si_meminfo(struct sysinfo *val)
4179 val->totalram = totalram_pages;
4180 val->sharedram = global_node_page_state(NR_SHMEM);
4181 val->freeram = global_page_state(NR_FREE_PAGES);
4182 val->bufferram = nr_blockdev_pages();
4183 val->totalhigh = totalhigh_pages;
4184 val->freehigh = nr_free_highpages();
4185 val->mem_unit = PAGE_SIZE;
4188 EXPORT_SYMBOL(si_meminfo);
4191 void si_meminfo_node(struct sysinfo *val, int nid)
4193 int zone_type; /* needs to be signed */
4194 unsigned long managed_pages = 0;
4195 unsigned long managed_highpages = 0;
4196 unsigned long free_highpages = 0;
4197 pg_data_t *pgdat = NODE_DATA(nid);
4199 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4200 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4201 val->totalram = managed_pages;
4202 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4203 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4204 #ifdef CONFIG_HIGHMEM
4205 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4206 struct zone *zone = &pgdat->node_zones[zone_type];
4208 if (is_highmem(zone)) {
4209 managed_highpages += zone->managed_pages;
4210 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4213 val->totalhigh = managed_highpages;
4214 val->freehigh = free_highpages;
4216 val->totalhigh = managed_highpages;
4217 val->freehigh = free_highpages;
4219 val->mem_unit = PAGE_SIZE;
4224 * Determine whether the node should be displayed or not, depending on whether
4225 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4227 bool skip_free_areas_node(unsigned int flags, int nid)
4230 unsigned int cpuset_mems_cookie;
4232 if (!(flags & SHOW_MEM_FILTER_NODES))
4236 cpuset_mems_cookie = read_mems_allowed_begin();
4237 ret = !node_isset(nid, cpuset_current_mems_allowed);
4238 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4243 #define K(x) ((x) << (PAGE_SHIFT-10))
4245 static void show_migration_types(unsigned char type)
4247 static const char types[MIGRATE_TYPES] = {
4248 [MIGRATE_UNMOVABLE] = 'U',
4249 [MIGRATE_MOVABLE] = 'M',
4250 [MIGRATE_RECLAIMABLE] = 'E',
4251 [MIGRATE_HIGHATOMIC] = 'H',
4253 [MIGRATE_CMA] = 'C',
4255 #ifdef CONFIG_MEMORY_ISOLATION
4256 [MIGRATE_ISOLATE] = 'I',
4259 char tmp[MIGRATE_TYPES + 1];
4263 for (i = 0; i < MIGRATE_TYPES; i++) {
4264 if (type & (1 << i))
4269 printk("(%s) ", tmp);
4273 * Show free area list (used inside shift_scroll-lock stuff)
4274 * We also calculate the percentage fragmentation. We do this by counting the
4275 * memory on each free list with the exception of the first item on the list.
4278 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4281 void show_free_areas(unsigned int filter)
4283 unsigned long free_pcp = 0;
4288 for_each_populated_zone(zone) {
4289 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4292 for_each_online_cpu(cpu)
4293 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4296 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4297 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4298 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4299 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4300 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4301 " free:%lu free_pcp:%lu free_cma:%lu\n",
4302 global_node_page_state(NR_ACTIVE_ANON),
4303 global_node_page_state(NR_INACTIVE_ANON),
4304 global_node_page_state(NR_ISOLATED_ANON),
4305 global_node_page_state(NR_ACTIVE_FILE),
4306 global_node_page_state(NR_INACTIVE_FILE),
4307 global_node_page_state(NR_ISOLATED_FILE),
4308 global_node_page_state(NR_UNEVICTABLE),
4309 global_node_page_state(NR_FILE_DIRTY),
4310 global_node_page_state(NR_WRITEBACK),
4311 global_node_page_state(NR_UNSTABLE_NFS),
4312 global_page_state(NR_SLAB_RECLAIMABLE),
4313 global_page_state(NR_SLAB_UNRECLAIMABLE),
4314 global_node_page_state(NR_FILE_MAPPED),
4315 global_node_page_state(NR_SHMEM),
4316 global_page_state(NR_PAGETABLE),
4317 global_page_state(NR_BOUNCE),
4318 global_page_state(NR_FREE_PAGES),
4320 global_page_state(NR_FREE_CMA_PAGES));
4322 for_each_online_pgdat(pgdat) {
4324 " active_anon:%lukB"
4325 " inactive_anon:%lukB"
4326 " active_file:%lukB"
4327 " inactive_file:%lukB"
4328 " unevictable:%lukB"
4329 " isolated(anon):%lukB"
4330 " isolated(file):%lukB"
4335 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4337 " shmem_pmdmapped: %lukB"
4340 " writeback_tmp:%lukB"
4342 " all_unreclaimable? %s"
4345 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4346 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4347 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4348 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4349 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4350 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4351 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4352 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4353 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4354 K(node_page_state(pgdat, NR_WRITEBACK)),
4355 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4356 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4357 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4359 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4361 K(node_page_state(pgdat, NR_SHMEM)),
4362 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4363 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4364 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4367 for_each_populated_zone(zone) {
4370 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4374 for_each_online_cpu(cpu)
4375 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4386 " slab_reclaimable:%lukB"
4387 " slab_unreclaimable:%lukB"
4388 " kernel_stack:%lukB"
4394 " node_pages_scanned:%lu"
4397 K(zone_page_state(zone, NR_FREE_PAGES)),
4398 K(min_wmark_pages(zone)),
4399 K(low_wmark_pages(zone)),
4400 K(high_wmark_pages(zone)),
4401 K(zone->present_pages),
4402 K(zone->managed_pages),
4403 K(zone_page_state(zone, NR_MLOCK)),
4404 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4405 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4406 zone_page_state(zone, NR_KERNEL_STACK) *
4408 K(zone_page_state(zone, NR_PAGETABLE)),
4409 K(zone_page_state(zone, NR_BOUNCE)),
4411 K(this_cpu_read(zone->pageset->pcp.count)),
4412 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4413 K(node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED)));
4414 printk("lowmem_reserve[]:");
4415 for (i = 0; i < MAX_NR_ZONES; i++)
4416 printk(" %ld", zone->lowmem_reserve[i]);
4420 for_each_populated_zone(zone) {
4422 unsigned long nr[MAX_ORDER], flags, total = 0;
4423 unsigned char types[MAX_ORDER];
4425 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4428 printk("%s: ", zone->name);
4430 spin_lock_irqsave(&zone->lock, flags);
4431 for (order = 0; order < MAX_ORDER; order++) {
4432 struct free_area *area = &zone->free_area[order];
4435 nr[order] = area->nr_free;
4436 total += nr[order] << order;
4439 for (type = 0; type < MIGRATE_TYPES; type++) {
4440 if (!list_empty(&area->free_list[type]))
4441 types[order] |= 1 << type;
4444 spin_unlock_irqrestore(&zone->lock, flags);
4445 for (order = 0; order < MAX_ORDER; order++) {
4446 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4448 show_migration_types(types[order]);
4450 printk("= %lukB\n", K(total));
4453 hugetlb_show_meminfo();
4455 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4457 show_swap_cache_info();
4460 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4462 zoneref->zone = zone;
4463 zoneref->zone_idx = zone_idx(zone);
4467 * Builds allocation fallback zone lists.
4469 * Add all populated zones of a node to the zonelist.
4471 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4475 enum zone_type zone_type = MAX_NR_ZONES;
4479 zone = pgdat->node_zones + zone_type;
4480 if (populated_zone(zone)) {
4481 zoneref_set_zone(zone,
4482 &zonelist->_zonerefs[nr_zones++]);
4483 check_highest_zone(zone_type);
4485 } while (zone_type);
4493 * 0 = automatic detection of better ordering.
4494 * 1 = order by ([node] distance, -zonetype)
4495 * 2 = order by (-zonetype, [node] distance)
4497 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4498 * the same zonelist. So only NUMA can configure this param.
4500 #define ZONELIST_ORDER_DEFAULT 0
4501 #define ZONELIST_ORDER_NODE 1
4502 #define ZONELIST_ORDER_ZONE 2
4504 /* zonelist order in the kernel.
4505 * set_zonelist_order() will set this to NODE or ZONE.
4507 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4508 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4512 /* The value user specified ....changed by config */
4513 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4514 /* string for sysctl */
4515 #define NUMA_ZONELIST_ORDER_LEN 16
4516 char numa_zonelist_order[16] = "default";
4519 * interface for configure zonelist ordering.
4520 * command line option "numa_zonelist_order"
4521 * = "[dD]efault - default, automatic configuration.
4522 * = "[nN]ode - order by node locality, then by zone within node
4523 * = "[zZ]one - order by zone, then by locality within zone
4526 static int __parse_numa_zonelist_order(char *s)
4528 if (*s == 'd' || *s == 'D') {
4529 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4530 } else if (*s == 'n' || *s == 'N') {
4531 user_zonelist_order = ZONELIST_ORDER_NODE;
4532 } else if (*s == 'z' || *s == 'Z') {
4533 user_zonelist_order = ZONELIST_ORDER_ZONE;
4535 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4541 static __init int setup_numa_zonelist_order(char *s)
4548 ret = __parse_numa_zonelist_order(s);
4550 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4554 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4557 * sysctl handler for numa_zonelist_order
4559 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4560 void __user *buffer, size_t *length,
4563 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4565 static DEFINE_MUTEX(zl_order_mutex);
4567 mutex_lock(&zl_order_mutex);
4569 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4573 strcpy(saved_string, (char *)table->data);
4575 ret = proc_dostring(table, write, buffer, length, ppos);
4579 int oldval = user_zonelist_order;
4581 ret = __parse_numa_zonelist_order((char *)table->data);
4584 * bogus value. restore saved string
4586 strncpy((char *)table->data, saved_string,
4587 NUMA_ZONELIST_ORDER_LEN);
4588 user_zonelist_order = oldval;
4589 } else if (oldval != user_zonelist_order) {
4590 mutex_lock(&zonelists_mutex);
4591 build_all_zonelists(NULL, NULL);
4592 mutex_unlock(&zonelists_mutex);
4596 mutex_unlock(&zl_order_mutex);
4601 #define MAX_NODE_LOAD (nr_online_nodes)
4602 static int node_load[MAX_NUMNODES];
4605 * find_next_best_node - find the next node that should appear in a given node's fallback list
4606 * @node: node whose fallback list we're appending
4607 * @used_node_mask: nodemask_t of already used nodes
4609 * We use a number of factors to determine which is the next node that should
4610 * appear on a given node's fallback list. The node should not have appeared
4611 * already in @node's fallback list, and it should be the next closest node
4612 * according to the distance array (which contains arbitrary distance values
4613 * from each node to each node in the system), and should also prefer nodes
4614 * with no CPUs, since presumably they'll have very little allocation pressure
4615 * on them otherwise.
4616 * It returns -1 if no node is found.
4618 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4621 int min_val = INT_MAX;
4622 int best_node = NUMA_NO_NODE;
4623 const struct cpumask *tmp = cpumask_of_node(0);
4625 /* Use the local node if we haven't already */
4626 if (!node_isset(node, *used_node_mask)) {
4627 node_set(node, *used_node_mask);
4631 for_each_node_state(n, N_MEMORY) {
4633 /* Don't want a node to appear more than once */
4634 if (node_isset(n, *used_node_mask))
4637 /* Use the distance array to find the distance */
4638 val = node_distance(node, n);
4640 /* Penalize nodes under us ("prefer the next node") */
4643 /* Give preference to headless and unused nodes */
4644 tmp = cpumask_of_node(n);
4645 if (!cpumask_empty(tmp))
4646 val += PENALTY_FOR_NODE_WITH_CPUS;
4648 /* Slight preference for less loaded node */
4649 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4650 val += node_load[n];
4652 if (val < min_val) {
4659 node_set(best_node, *used_node_mask);
4666 * Build zonelists ordered by node and zones within node.
4667 * This results in maximum locality--normal zone overflows into local
4668 * DMA zone, if any--but risks exhausting DMA zone.
4670 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4673 struct zonelist *zonelist;
4675 zonelist = &pgdat->node_zonelists[0];
4676 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4678 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4679 zonelist->_zonerefs[j].zone = NULL;
4680 zonelist->_zonerefs[j].zone_idx = 0;
4684 * Build gfp_thisnode zonelists
4686 static void build_thisnode_zonelists(pg_data_t *pgdat)
4689 struct zonelist *zonelist;
4691 zonelist = &pgdat->node_zonelists[1];
4692 j = build_zonelists_node(pgdat, zonelist, 0);
4693 zonelist->_zonerefs[j].zone = NULL;
4694 zonelist->_zonerefs[j].zone_idx = 0;
4698 * Build zonelists ordered by zone and nodes within zones.
4699 * This results in conserving DMA zone[s] until all Normal memory is
4700 * exhausted, but results in overflowing to remote node while memory
4701 * may still exist in local DMA zone.
4703 static int node_order[MAX_NUMNODES];
4705 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4708 int zone_type; /* needs to be signed */
4710 struct zonelist *zonelist;
4712 zonelist = &pgdat->node_zonelists[0];
4714 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4715 for (j = 0; j < nr_nodes; j++) {
4716 node = node_order[j];
4717 z = &NODE_DATA(node)->node_zones[zone_type];
4718 if (populated_zone(z)) {
4720 &zonelist->_zonerefs[pos++]);
4721 check_highest_zone(zone_type);
4725 zonelist->_zonerefs[pos].zone = NULL;
4726 zonelist->_zonerefs[pos].zone_idx = 0;
4729 #if defined(CONFIG_64BIT)
4731 * Devices that require DMA32/DMA are relatively rare and do not justify a
4732 * penalty to every machine in case the specialised case applies. Default
4733 * to Node-ordering on 64-bit NUMA machines
4735 static int default_zonelist_order(void)
4737 return ZONELIST_ORDER_NODE;
4741 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4742 * by the kernel. If processes running on node 0 deplete the low memory zone
4743 * then reclaim will occur more frequency increasing stalls and potentially
4744 * be easier to OOM if a large percentage of the zone is under writeback or
4745 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4746 * Hence, default to zone ordering on 32-bit.
4748 static int default_zonelist_order(void)
4750 return ZONELIST_ORDER_ZONE;
4752 #endif /* CONFIG_64BIT */
4754 static void set_zonelist_order(void)
4756 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4757 current_zonelist_order = default_zonelist_order();
4759 current_zonelist_order = user_zonelist_order;
4762 static void build_zonelists(pg_data_t *pgdat)
4765 nodemask_t used_mask;
4766 int local_node, prev_node;
4767 struct zonelist *zonelist;
4768 unsigned int order = current_zonelist_order;
4770 /* initialize zonelists */
4771 for (i = 0; i < MAX_ZONELISTS; i++) {
4772 zonelist = pgdat->node_zonelists + i;
4773 zonelist->_zonerefs[0].zone = NULL;
4774 zonelist->_zonerefs[0].zone_idx = 0;
4777 /* NUMA-aware ordering of nodes */
4778 local_node = pgdat->node_id;
4779 load = nr_online_nodes;
4780 prev_node = local_node;
4781 nodes_clear(used_mask);
4783 memset(node_order, 0, sizeof(node_order));
4786 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4788 * We don't want to pressure a particular node.
4789 * So adding penalty to the first node in same
4790 * distance group to make it round-robin.
4792 if (node_distance(local_node, node) !=
4793 node_distance(local_node, prev_node))
4794 node_load[node] = load;
4798 if (order == ZONELIST_ORDER_NODE)
4799 build_zonelists_in_node_order(pgdat, node);
4801 node_order[i++] = node; /* remember order */
4804 if (order == ZONELIST_ORDER_ZONE) {
4805 /* calculate node order -- i.e., DMA last! */
4806 build_zonelists_in_zone_order(pgdat, i);
4809 build_thisnode_zonelists(pgdat);
4812 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4814 * Return node id of node used for "local" allocations.
4815 * I.e., first node id of first zone in arg node's generic zonelist.
4816 * Used for initializing percpu 'numa_mem', which is used primarily
4817 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4819 int local_memory_node(int node)
4823 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4824 gfp_zone(GFP_KERNEL),
4826 return z->zone->node;
4830 #else /* CONFIG_NUMA */
4832 static void set_zonelist_order(void)
4834 current_zonelist_order = ZONELIST_ORDER_ZONE;
4837 static void build_zonelists(pg_data_t *pgdat)
4839 int node, local_node;
4841 struct zonelist *zonelist;
4843 local_node = pgdat->node_id;
4845 zonelist = &pgdat->node_zonelists[0];
4846 j = build_zonelists_node(pgdat, zonelist, 0);
4849 * Now we build the zonelist so that it contains the zones
4850 * of all the other nodes.
4851 * We don't want to pressure a particular node, so when
4852 * building the zones for node N, we make sure that the
4853 * zones coming right after the local ones are those from
4854 * node N+1 (modulo N)
4856 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4857 if (!node_online(node))
4859 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4861 for (node = 0; node < local_node; node++) {
4862 if (!node_online(node))
4864 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4867 zonelist->_zonerefs[j].zone = NULL;
4868 zonelist->_zonerefs[j].zone_idx = 0;
4871 #endif /* CONFIG_NUMA */
4874 * Boot pageset table. One per cpu which is going to be used for all
4875 * zones and all nodes. The parameters will be set in such a way
4876 * that an item put on a list will immediately be handed over to
4877 * the buddy list. This is safe since pageset manipulation is done
4878 * with interrupts disabled.
4880 * The boot_pagesets must be kept even after bootup is complete for
4881 * unused processors and/or zones. They do play a role for bootstrapping
4882 * hotplugged processors.
4884 * zoneinfo_show() and maybe other functions do
4885 * not check if the processor is online before following the pageset pointer.
4886 * Other parts of the kernel may not check if the zone is available.
4888 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4889 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4890 static void setup_zone_pageset(struct zone *zone);
4893 * Global mutex to protect against size modification of zonelists
4894 * as well as to serialize pageset setup for the new populated zone.
4896 DEFINE_MUTEX(zonelists_mutex);
4898 /* return values int ....just for stop_machine() */
4899 static int __build_all_zonelists(void *data)
4903 pg_data_t *self = data;
4906 memset(node_load, 0, sizeof(node_load));
4909 if (self && !node_online(self->node_id)) {
4910 build_zonelists(self);
4913 for_each_online_node(nid) {
4914 pg_data_t *pgdat = NODE_DATA(nid);
4916 build_zonelists(pgdat);
4920 * Initialize the boot_pagesets that are going to be used
4921 * for bootstrapping processors. The real pagesets for
4922 * each zone will be allocated later when the per cpu
4923 * allocator is available.
4925 * boot_pagesets are used also for bootstrapping offline
4926 * cpus if the system is already booted because the pagesets
4927 * are needed to initialize allocators on a specific cpu too.
4928 * F.e. the percpu allocator needs the page allocator which
4929 * needs the percpu allocator in order to allocate its pagesets
4930 * (a chicken-egg dilemma).
4932 for_each_possible_cpu(cpu) {
4933 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4935 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4937 * We now know the "local memory node" for each node--
4938 * i.e., the node of the first zone in the generic zonelist.
4939 * Set up numa_mem percpu variable for on-line cpus. During
4940 * boot, only the boot cpu should be on-line; we'll init the
4941 * secondary cpus' numa_mem as they come on-line. During
4942 * node/memory hotplug, we'll fixup all on-line cpus.
4944 if (cpu_online(cpu))
4945 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4952 static noinline void __init
4953 build_all_zonelists_init(void)
4955 __build_all_zonelists(NULL);
4956 mminit_verify_zonelist();
4957 cpuset_init_current_mems_allowed();
4961 * Called with zonelists_mutex held always
4962 * unless system_state == SYSTEM_BOOTING.
4964 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4965 * [we're only called with non-NULL zone through __meminit paths] and
4966 * (2) call of __init annotated helper build_all_zonelists_init
4967 * [protected by SYSTEM_BOOTING].
4969 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4971 set_zonelist_order();
4973 if (system_state == SYSTEM_BOOTING) {
4974 build_all_zonelists_init();
4976 #ifdef CONFIG_MEMORY_HOTPLUG
4978 setup_zone_pageset(zone);
4980 /* we have to stop all cpus to guarantee there is no user
4982 stop_machine(__build_all_zonelists, pgdat, NULL);
4983 /* cpuset refresh routine should be here */
4985 vm_total_pages = nr_free_pagecache_pages();
4987 * Disable grouping by mobility if the number of pages in the
4988 * system is too low to allow the mechanism to work. It would be
4989 * more accurate, but expensive to check per-zone. This check is
4990 * made on memory-hotadd so a system can start with mobility
4991 * disabled and enable it later
4993 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4994 page_group_by_mobility_disabled = 1;
4996 page_group_by_mobility_disabled = 0;
4998 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5000 zonelist_order_name[current_zonelist_order],
5001 page_group_by_mobility_disabled ? "off" : "on",
5004 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5009 * Helper functions to size the waitqueue hash table.
5010 * Essentially these want to choose hash table sizes sufficiently
5011 * large so that collisions trying to wait on pages are rare.
5012 * But in fact, the number of active page waitqueues on typical
5013 * systems is ridiculously low, less than 200. So this is even
5014 * conservative, even though it seems large.
5016 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5017 * waitqueues, i.e. the size of the waitq table given the number of pages.
5019 #define PAGES_PER_WAITQUEUE 256
5021 #ifndef CONFIG_MEMORY_HOTPLUG
5022 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5024 unsigned long size = 1;
5026 pages /= PAGES_PER_WAITQUEUE;
5028 while (size < pages)
5032 * Once we have dozens or even hundreds of threads sleeping
5033 * on IO we've got bigger problems than wait queue collision.
5034 * Limit the size of the wait table to a reasonable size.
5036 size = min(size, 4096UL);
5038 return max(size, 4UL);
5042 * A zone's size might be changed by hot-add, so it is not possible to determine
5043 * a suitable size for its wait_table. So we use the maximum size now.
5045 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5047 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5048 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5049 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5051 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5052 * or more by the traditional way. (See above). It equals:
5054 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5055 * ia64(16K page size) : = ( 8G + 4M)byte.
5056 * powerpc (64K page size) : = (32G +16M)byte.
5058 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5065 * This is an integer logarithm so that shifts can be used later
5066 * to extract the more random high bits from the multiplicative
5067 * hash function before the remainder is taken.
5069 static inline unsigned long wait_table_bits(unsigned long size)
5075 * Initially all pages are reserved - free ones are freed
5076 * up by free_all_bootmem() once the early boot process is
5077 * done. Non-atomic initialization, single-pass.
5079 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5080 unsigned long start_pfn, enum memmap_context context)
5082 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5083 unsigned long end_pfn = start_pfn + size;
5084 pg_data_t *pgdat = NODE_DATA(nid);
5086 unsigned long nr_initialised = 0;
5087 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5088 struct memblock_region *r = NULL, *tmp;
5091 if (highest_memmap_pfn < end_pfn - 1)
5092 highest_memmap_pfn = end_pfn - 1;
5095 * Honor reservation requested by the driver for this ZONE_DEVICE
5098 if (altmap && start_pfn == altmap->base_pfn)
5099 start_pfn += altmap->reserve;
5101 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5103 * There can be holes in boot-time mem_map[]s handed to this
5104 * function. They do not exist on hotplugged memory.
5106 if (context != MEMMAP_EARLY)
5109 if (!early_pfn_valid(pfn))
5111 if (!early_pfn_in_nid(pfn, nid))
5113 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5116 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5118 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5119 * from zone_movable_pfn[nid] to end of each node should be
5120 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5122 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5123 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5127 * Check given memblock attribute by firmware which can affect
5128 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5129 * mirrored, it's an overlapped memmap init. skip it.
5131 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5132 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5133 for_each_memblock(memory, tmp)
5134 if (pfn < memblock_region_memory_end_pfn(tmp))
5138 if (pfn >= memblock_region_memory_base_pfn(r) &&
5139 memblock_is_mirror(r)) {
5140 /* already initialized as NORMAL */
5141 pfn = memblock_region_memory_end_pfn(r);
5149 * Mark the block movable so that blocks are reserved for
5150 * movable at startup. This will force kernel allocations
5151 * to reserve their blocks rather than leaking throughout
5152 * the address space during boot when many long-lived
5153 * kernel allocations are made.
5155 * bitmap is created for zone's valid pfn range. but memmap
5156 * can be created for invalid pages (for alignment)
5157 * check here not to call set_pageblock_migratetype() against
5160 if (!(pfn & (pageblock_nr_pages - 1))) {
5161 struct page *page = pfn_to_page(pfn);
5163 __init_single_page(page, pfn, zone, nid);
5164 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5166 __init_single_pfn(pfn, zone, nid);
5171 static void __meminit zone_init_free_lists(struct zone *zone)
5173 unsigned int order, t;
5174 for_each_migratetype_order(order, t) {
5175 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5176 zone->free_area[order].nr_free = 0;
5180 #ifndef __HAVE_ARCH_MEMMAP_INIT
5181 #define memmap_init(size, nid, zone, start_pfn) \
5182 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5185 static int zone_batchsize(struct zone *zone)
5191 * The per-cpu-pages pools are set to around 1000th of the
5192 * size of the zone. But no more than 1/2 of a meg.
5194 * OK, so we don't know how big the cache is. So guess.
5196 batch = zone->managed_pages / 1024;
5197 if (batch * PAGE_SIZE > 512 * 1024)
5198 batch = (512 * 1024) / PAGE_SIZE;
5199 batch /= 4; /* We effectively *= 4 below */
5204 * Clamp the batch to a 2^n - 1 value. Having a power
5205 * of 2 value was found to be more likely to have
5206 * suboptimal cache aliasing properties in some cases.
5208 * For example if 2 tasks are alternately allocating
5209 * batches of pages, one task can end up with a lot
5210 * of pages of one half of the possible page colors
5211 * and the other with pages of the other colors.
5213 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5218 /* The deferral and batching of frees should be suppressed under NOMMU
5221 * The problem is that NOMMU needs to be able to allocate large chunks
5222 * of contiguous memory as there's no hardware page translation to
5223 * assemble apparent contiguous memory from discontiguous pages.
5225 * Queueing large contiguous runs of pages for batching, however,
5226 * causes the pages to actually be freed in smaller chunks. As there
5227 * can be a significant delay between the individual batches being
5228 * recycled, this leads to the once large chunks of space being
5229 * fragmented and becoming unavailable for high-order allocations.
5236 * pcp->high and pcp->batch values are related and dependent on one another:
5237 * ->batch must never be higher then ->high.
5238 * The following function updates them in a safe manner without read side
5241 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5242 * those fields changing asynchronously (acording the the above rule).
5244 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5245 * outside of boot time (or some other assurance that no concurrent updaters
5248 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5249 unsigned long batch)
5251 /* start with a fail safe value for batch */
5255 /* Update high, then batch, in order */
5262 /* a companion to pageset_set_high() */
5263 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5265 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5268 static void pageset_init(struct per_cpu_pageset *p)
5270 struct per_cpu_pages *pcp;
5273 memset(p, 0, sizeof(*p));
5277 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5278 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5281 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5284 pageset_set_batch(p, batch);
5288 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5289 * to the value high for the pageset p.
5291 static void pageset_set_high(struct per_cpu_pageset *p,
5294 unsigned long batch = max(1UL, high / 4);
5295 if ((high / 4) > (PAGE_SHIFT * 8))
5296 batch = PAGE_SHIFT * 8;
5298 pageset_update(&p->pcp, high, batch);
5301 static void pageset_set_high_and_batch(struct zone *zone,
5302 struct per_cpu_pageset *pcp)
5304 if (percpu_pagelist_fraction)
5305 pageset_set_high(pcp,
5306 (zone->managed_pages /
5307 percpu_pagelist_fraction));
5309 pageset_set_batch(pcp, zone_batchsize(zone));
5312 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5314 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5317 pageset_set_high_and_batch(zone, pcp);
5320 static void __meminit setup_zone_pageset(struct zone *zone)
5323 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5324 for_each_possible_cpu(cpu)
5325 zone_pageset_init(zone, cpu);
5327 if (!zone->zone_pgdat->per_cpu_nodestats) {
5328 zone->zone_pgdat->per_cpu_nodestats =
5329 alloc_percpu(struct per_cpu_nodestat);
5334 * Allocate per cpu pagesets and initialize them.
5335 * Before this call only boot pagesets were available.
5337 void __init setup_per_cpu_pageset(void)
5341 for_each_populated_zone(zone)
5342 setup_zone_pageset(zone);
5345 static noinline __init_refok
5346 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5352 * The per-page waitqueue mechanism uses hashed waitqueues
5355 zone->wait_table_hash_nr_entries =
5356 wait_table_hash_nr_entries(zone_size_pages);
5357 zone->wait_table_bits =
5358 wait_table_bits(zone->wait_table_hash_nr_entries);
5359 alloc_size = zone->wait_table_hash_nr_entries
5360 * sizeof(wait_queue_head_t);
5362 if (!slab_is_available()) {
5363 zone->wait_table = (wait_queue_head_t *)
5364 memblock_virt_alloc_node_nopanic(
5365 alloc_size, zone->zone_pgdat->node_id);
5368 * This case means that a zone whose size was 0 gets new memory
5369 * via memory hot-add.
5370 * But it may be the case that a new node was hot-added. In
5371 * this case vmalloc() will not be able to use this new node's
5372 * memory - this wait_table must be initialized to use this new
5373 * node itself as well.
5374 * To use this new node's memory, further consideration will be
5377 zone->wait_table = vmalloc(alloc_size);
5379 if (!zone->wait_table)
5382 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5383 init_waitqueue_head(zone->wait_table + i);
5388 static __meminit void zone_pcp_init(struct zone *zone)
5391 * per cpu subsystem is not up at this point. The following code
5392 * relies on the ability of the linker to provide the
5393 * offset of a (static) per cpu variable into the per cpu area.
5395 zone->pageset = &boot_pageset;
5397 if (populated_zone(zone))
5398 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5399 zone->name, zone->present_pages,
5400 zone_batchsize(zone));
5403 int __meminit init_currently_empty_zone(struct zone *zone,
5404 unsigned long zone_start_pfn,
5407 struct pglist_data *pgdat = zone->zone_pgdat;
5409 ret = zone_wait_table_init(zone, size);
5412 pgdat->nr_zones = zone_idx(zone) + 1;
5414 zone->zone_start_pfn = zone_start_pfn;
5416 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5417 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5419 (unsigned long)zone_idx(zone),
5420 zone_start_pfn, (zone_start_pfn + size));
5422 zone_init_free_lists(zone);
5427 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5428 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5431 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5433 int __meminit __early_pfn_to_nid(unsigned long pfn,
5434 struct mminit_pfnnid_cache *state)
5436 unsigned long start_pfn, end_pfn;
5439 if (state->last_start <= pfn && pfn < state->last_end)
5440 return state->last_nid;
5442 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5444 state->last_start = start_pfn;
5445 state->last_end = end_pfn;
5446 state->last_nid = nid;
5451 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5454 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5455 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5456 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5458 * If an architecture guarantees that all ranges registered contain no holes
5459 * and may be freed, this this function may be used instead of calling
5460 * memblock_free_early_nid() manually.
5462 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5464 unsigned long start_pfn, end_pfn;
5467 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5468 start_pfn = min(start_pfn, max_low_pfn);
5469 end_pfn = min(end_pfn, max_low_pfn);
5471 if (start_pfn < end_pfn)
5472 memblock_free_early_nid(PFN_PHYS(start_pfn),
5473 (end_pfn - start_pfn) << PAGE_SHIFT,
5479 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5480 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5482 * If an architecture guarantees that all ranges registered contain no holes and may
5483 * be freed, this function may be used instead of calling memory_present() manually.
5485 void __init sparse_memory_present_with_active_regions(int nid)
5487 unsigned long start_pfn, end_pfn;
5490 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5491 memory_present(this_nid, start_pfn, end_pfn);
5495 * get_pfn_range_for_nid - Return the start and end page frames for a node
5496 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5497 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5498 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5500 * It returns the start and end page frame of a node based on information
5501 * provided by memblock_set_node(). If called for a node
5502 * with no available memory, a warning is printed and the start and end
5505 void __meminit get_pfn_range_for_nid(unsigned int nid,
5506 unsigned long *start_pfn, unsigned long *end_pfn)
5508 unsigned long this_start_pfn, this_end_pfn;
5514 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5515 *start_pfn = min(*start_pfn, this_start_pfn);
5516 *end_pfn = max(*end_pfn, this_end_pfn);
5519 if (*start_pfn == -1UL)
5524 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5525 * assumption is made that zones within a node are ordered in monotonic
5526 * increasing memory addresses so that the "highest" populated zone is used
5528 static void __init find_usable_zone_for_movable(void)
5531 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5532 if (zone_index == ZONE_MOVABLE)
5535 if (arch_zone_highest_possible_pfn[zone_index] >
5536 arch_zone_lowest_possible_pfn[zone_index])
5540 VM_BUG_ON(zone_index == -1);
5541 movable_zone = zone_index;
5545 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5546 * because it is sized independent of architecture. Unlike the other zones,
5547 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5548 * in each node depending on the size of each node and how evenly kernelcore
5549 * is distributed. This helper function adjusts the zone ranges
5550 * provided by the architecture for a given node by using the end of the
5551 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5552 * zones within a node are in order of monotonic increases memory addresses
5554 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5555 unsigned long zone_type,
5556 unsigned long node_start_pfn,
5557 unsigned long node_end_pfn,
5558 unsigned long *zone_start_pfn,
5559 unsigned long *zone_end_pfn)
5561 /* Only adjust if ZONE_MOVABLE is on this node */
5562 if (zone_movable_pfn[nid]) {
5563 /* Size ZONE_MOVABLE */
5564 if (zone_type == ZONE_MOVABLE) {
5565 *zone_start_pfn = zone_movable_pfn[nid];
5566 *zone_end_pfn = min(node_end_pfn,
5567 arch_zone_highest_possible_pfn[movable_zone]);
5569 /* Check if this whole range is within ZONE_MOVABLE */
5570 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5571 *zone_start_pfn = *zone_end_pfn;
5576 * Return the number of pages a zone spans in a node, including holes
5577 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5579 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5580 unsigned long zone_type,
5581 unsigned long node_start_pfn,
5582 unsigned long node_end_pfn,
5583 unsigned long *zone_start_pfn,
5584 unsigned long *zone_end_pfn,
5585 unsigned long *ignored)
5587 /* When hotadd a new node from cpu_up(), the node should be empty */
5588 if (!node_start_pfn && !node_end_pfn)
5591 /* Get the start and end of the zone */
5592 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5593 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5594 adjust_zone_range_for_zone_movable(nid, zone_type,
5595 node_start_pfn, node_end_pfn,
5596 zone_start_pfn, zone_end_pfn);
5598 /* Check that this node has pages within the zone's required range */
5599 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5602 /* Move the zone boundaries inside the node if necessary */
5603 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5604 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5606 /* Return the spanned pages */
5607 return *zone_end_pfn - *zone_start_pfn;
5611 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5612 * then all holes in the requested range will be accounted for.
5614 unsigned long __meminit __absent_pages_in_range(int nid,
5615 unsigned long range_start_pfn,
5616 unsigned long range_end_pfn)
5618 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5619 unsigned long start_pfn, end_pfn;
5622 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5623 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5624 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5625 nr_absent -= end_pfn - start_pfn;
5631 * absent_pages_in_range - Return number of page frames in holes within a range
5632 * @start_pfn: The start PFN to start searching for holes
5633 * @end_pfn: The end PFN to stop searching for holes
5635 * It returns the number of pages frames in memory holes within a range.
5637 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5638 unsigned long end_pfn)
5640 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5643 /* Return the number of page frames in holes in a zone on a node */
5644 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5645 unsigned long zone_type,
5646 unsigned long node_start_pfn,
5647 unsigned long node_end_pfn,
5648 unsigned long *ignored)
5650 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5651 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5652 unsigned long zone_start_pfn, zone_end_pfn;
5653 unsigned long nr_absent;
5655 /* When hotadd a new node from cpu_up(), the node should be empty */
5656 if (!node_start_pfn && !node_end_pfn)
5659 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5660 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5662 adjust_zone_range_for_zone_movable(nid, zone_type,
5663 node_start_pfn, node_end_pfn,
5664 &zone_start_pfn, &zone_end_pfn);
5665 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5668 * ZONE_MOVABLE handling.
5669 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5672 if (zone_movable_pfn[nid]) {
5673 if (mirrored_kernelcore) {
5674 unsigned long start_pfn, end_pfn;
5675 struct memblock_region *r;
5677 for_each_memblock(memory, r) {
5678 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5679 zone_start_pfn, zone_end_pfn);
5680 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5681 zone_start_pfn, zone_end_pfn);
5683 if (zone_type == ZONE_MOVABLE &&
5684 memblock_is_mirror(r))
5685 nr_absent += end_pfn - start_pfn;
5687 if (zone_type == ZONE_NORMAL &&
5688 !memblock_is_mirror(r))
5689 nr_absent += end_pfn - start_pfn;
5692 if (zone_type == ZONE_NORMAL)
5693 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5700 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5701 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5702 unsigned long zone_type,
5703 unsigned long node_start_pfn,
5704 unsigned long node_end_pfn,
5705 unsigned long *zone_start_pfn,
5706 unsigned long *zone_end_pfn,
5707 unsigned long *zones_size)
5711 *zone_start_pfn = node_start_pfn;
5712 for (zone = 0; zone < zone_type; zone++)
5713 *zone_start_pfn += zones_size[zone];
5715 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5717 return zones_size[zone_type];
5720 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5721 unsigned long zone_type,
5722 unsigned long node_start_pfn,
5723 unsigned long node_end_pfn,
5724 unsigned long *zholes_size)
5729 return zholes_size[zone_type];
5732 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5734 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5735 unsigned long node_start_pfn,
5736 unsigned long node_end_pfn,
5737 unsigned long *zones_size,
5738 unsigned long *zholes_size)
5740 unsigned long realtotalpages = 0, totalpages = 0;
5743 for (i = 0; i < MAX_NR_ZONES; i++) {
5744 struct zone *zone = pgdat->node_zones + i;
5745 unsigned long zone_start_pfn, zone_end_pfn;
5746 unsigned long size, real_size;
5748 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5754 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5755 node_start_pfn, node_end_pfn,
5758 zone->zone_start_pfn = zone_start_pfn;
5760 zone->zone_start_pfn = 0;
5761 zone->spanned_pages = size;
5762 zone->present_pages = real_size;
5765 realtotalpages += real_size;
5768 pgdat->node_spanned_pages = totalpages;
5769 pgdat->node_present_pages = realtotalpages;
5770 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5774 #ifndef CONFIG_SPARSEMEM
5776 * Calculate the size of the zone->blockflags rounded to an unsigned long
5777 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5778 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5779 * round what is now in bits to nearest long in bits, then return it in
5782 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5784 unsigned long usemapsize;
5786 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5787 usemapsize = roundup(zonesize, pageblock_nr_pages);
5788 usemapsize = usemapsize >> pageblock_order;
5789 usemapsize *= NR_PAGEBLOCK_BITS;
5790 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5792 return usemapsize / 8;
5795 static void __init setup_usemap(struct pglist_data *pgdat,
5797 unsigned long zone_start_pfn,
5798 unsigned long zonesize)
5800 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5801 zone->pageblock_flags = NULL;
5803 zone->pageblock_flags =
5804 memblock_virt_alloc_node_nopanic(usemapsize,
5808 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5809 unsigned long zone_start_pfn, unsigned long zonesize) {}
5810 #endif /* CONFIG_SPARSEMEM */
5812 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5814 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5815 void __paginginit set_pageblock_order(void)
5819 /* Check that pageblock_nr_pages has not already been setup */
5820 if (pageblock_order)
5823 if (HPAGE_SHIFT > PAGE_SHIFT)
5824 order = HUGETLB_PAGE_ORDER;
5826 order = MAX_ORDER - 1;
5829 * Assume the largest contiguous order of interest is a huge page.
5830 * This value may be variable depending on boot parameters on IA64 and
5833 pageblock_order = order;
5835 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5838 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5839 * is unused as pageblock_order is set at compile-time. See
5840 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5843 void __paginginit set_pageblock_order(void)
5847 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5849 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5850 unsigned long present_pages)
5852 unsigned long pages = spanned_pages;
5855 * Provide a more accurate estimation if there are holes within
5856 * the zone and SPARSEMEM is in use. If there are holes within the
5857 * zone, each populated memory region may cost us one or two extra
5858 * memmap pages due to alignment because memmap pages for each
5859 * populated regions may not naturally algined on page boundary.
5860 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5862 if (spanned_pages > present_pages + (present_pages >> 4) &&
5863 IS_ENABLED(CONFIG_SPARSEMEM))
5864 pages = present_pages;
5866 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5870 * Set up the zone data structures:
5871 * - mark all pages reserved
5872 * - mark all memory queues empty
5873 * - clear the memory bitmaps
5875 * NOTE: pgdat should get zeroed by caller.
5877 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5880 int nid = pgdat->node_id;
5883 pgdat_resize_init(pgdat);
5884 #ifdef CONFIG_NUMA_BALANCING
5885 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5886 pgdat->numabalancing_migrate_nr_pages = 0;
5887 pgdat->numabalancing_migrate_next_window = jiffies;
5889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5890 spin_lock_init(&pgdat->split_queue_lock);
5891 INIT_LIST_HEAD(&pgdat->split_queue);
5892 pgdat->split_queue_len = 0;
5894 init_waitqueue_head(&pgdat->kswapd_wait);
5895 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5896 #ifdef CONFIG_COMPACTION
5897 init_waitqueue_head(&pgdat->kcompactd_wait);
5899 pgdat_page_ext_init(pgdat);
5900 spin_lock_init(&pgdat->lru_lock);
5901 lruvec_init(node_lruvec(pgdat));
5903 for (j = 0; j < MAX_NR_ZONES; j++) {
5904 struct zone *zone = pgdat->node_zones + j;
5905 unsigned long size, realsize, freesize, memmap_pages;
5906 unsigned long zone_start_pfn = zone->zone_start_pfn;
5908 size = zone->spanned_pages;
5909 realsize = freesize = zone->present_pages;
5912 * Adjust freesize so that it accounts for how much memory
5913 * is used by this zone for memmap. This affects the watermark
5914 * and per-cpu initialisations
5916 memmap_pages = calc_memmap_size(size, realsize);
5917 if (!is_highmem_idx(j)) {
5918 if (freesize >= memmap_pages) {
5919 freesize -= memmap_pages;
5922 " %s zone: %lu pages used for memmap\n",
5923 zone_names[j], memmap_pages);
5925 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5926 zone_names[j], memmap_pages, freesize);
5929 /* Account for reserved pages */
5930 if (j == 0 && freesize > dma_reserve) {
5931 freesize -= dma_reserve;
5932 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5933 zone_names[0], dma_reserve);
5936 if (!is_highmem_idx(j))
5937 nr_kernel_pages += freesize;
5938 /* Charge for highmem memmap if there are enough kernel pages */
5939 else if (nr_kernel_pages > memmap_pages * 2)
5940 nr_kernel_pages -= memmap_pages;
5941 nr_all_pages += freesize;
5944 * Set an approximate value for lowmem here, it will be adjusted
5945 * when the bootmem allocator frees pages into the buddy system.
5946 * And all highmem pages will be managed by the buddy system.
5948 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5951 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5953 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5955 zone->name = zone_names[j];
5956 zone->zone_pgdat = pgdat;
5957 spin_lock_init(&zone->lock);
5958 zone_seqlock_init(zone);
5959 zone_pcp_init(zone);
5961 /* For bootup, initialized properly in watermark setup */
5962 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5967 set_pageblock_order();
5968 setup_usemap(pgdat, zone, zone_start_pfn, size);
5969 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5971 memmap_init(size, nid, j, zone_start_pfn);
5975 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5977 unsigned long __maybe_unused start = 0;
5978 unsigned long __maybe_unused offset = 0;
5980 /* Skip empty nodes */
5981 if (!pgdat->node_spanned_pages)
5984 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5985 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5986 offset = pgdat->node_start_pfn - start;
5987 /* ia64 gets its own node_mem_map, before this, without bootmem */
5988 if (!pgdat->node_mem_map) {
5989 unsigned long size, end;
5993 * The zone's endpoints aren't required to be MAX_ORDER
5994 * aligned but the node_mem_map endpoints must be in order
5995 * for the buddy allocator to function correctly.
5997 end = pgdat_end_pfn(pgdat);
5998 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5999 size = (end - start) * sizeof(struct page);
6000 map = alloc_remap(pgdat->node_id, size);
6002 map = memblock_virt_alloc_node_nopanic(size,
6004 pgdat->node_mem_map = map + offset;
6006 #ifndef CONFIG_NEED_MULTIPLE_NODES
6008 * With no DISCONTIG, the global mem_map is just set as node 0's
6010 if (pgdat == NODE_DATA(0)) {
6011 mem_map = NODE_DATA(0)->node_mem_map;
6012 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6013 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6015 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6018 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6021 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6022 unsigned long node_start_pfn, unsigned long *zholes_size)
6024 pg_data_t *pgdat = NODE_DATA(nid);
6025 unsigned long start_pfn = 0;
6026 unsigned long end_pfn = 0;
6028 /* pg_data_t should be reset to zero when it's allocated */
6029 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6031 reset_deferred_meminit(pgdat);
6032 pgdat->node_id = nid;
6033 pgdat->node_start_pfn = node_start_pfn;
6034 pgdat->per_cpu_nodestats = NULL;
6035 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6036 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6037 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6038 (u64)start_pfn << PAGE_SHIFT,
6039 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6041 start_pfn = node_start_pfn;
6043 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6044 zones_size, zholes_size);
6046 alloc_node_mem_map(pgdat);
6047 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6048 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6049 nid, (unsigned long)pgdat,
6050 (unsigned long)pgdat->node_mem_map);
6053 free_area_init_core(pgdat);
6056 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6058 #if MAX_NUMNODES > 1
6060 * Figure out the number of possible node ids.
6062 void __init setup_nr_node_ids(void)
6064 unsigned int highest;
6066 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6067 nr_node_ids = highest + 1;
6072 * node_map_pfn_alignment - determine the maximum internode alignment
6074 * This function should be called after node map is populated and sorted.
6075 * It calculates the maximum power of two alignment which can distinguish
6078 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6079 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6080 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6081 * shifted, 1GiB is enough and this function will indicate so.
6083 * This is used to test whether pfn -> nid mapping of the chosen memory
6084 * model has fine enough granularity to avoid incorrect mapping for the
6085 * populated node map.
6087 * Returns the determined alignment in pfn's. 0 if there is no alignment
6088 * requirement (single node).
6090 unsigned long __init node_map_pfn_alignment(void)
6092 unsigned long accl_mask = 0, last_end = 0;
6093 unsigned long start, end, mask;
6097 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6098 if (!start || last_nid < 0 || last_nid == nid) {
6105 * Start with a mask granular enough to pin-point to the
6106 * start pfn and tick off bits one-by-one until it becomes
6107 * too coarse to separate the current node from the last.
6109 mask = ~((1 << __ffs(start)) - 1);
6110 while (mask && last_end <= (start & (mask << 1)))
6113 /* accumulate all internode masks */
6117 /* convert mask to number of pages */
6118 return ~accl_mask + 1;
6121 /* Find the lowest pfn for a node */
6122 static unsigned long __init find_min_pfn_for_node(int nid)
6124 unsigned long min_pfn = ULONG_MAX;
6125 unsigned long start_pfn;
6128 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6129 min_pfn = min(min_pfn, start_pfn);
6131 if (min_pfn == ULONG_MAX) {
6132 pr_warn("Could not find start_pfn for node %d\n", nid);
6140 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6142 * It returns the minimum PFN based on information provided via
6143 * memblock_set_node().
6145 unsigned long __init find_min_pfn_with_active_regions(void)
6147 return find_min_pfn_for_node(MAX_NUMNODES);
6151 * early_calculate_totalpages()
6152 * Sum pages in active regions for movable zone.
6153 * Populate N_MEMORY for calculating usable_nodes.
6155 static unsigned long __init early_calculate_totalpages(void)
6157 unsigned long totalpages = 0;
6158 unsigned long start_pfn, end_pfn;
6161 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6162 unsigned long pages = end_pfn - start_pfn;
6164 totalpages += pages;
6166 node_set_state(nid, N_MEMORY);
6172 * Find the PFN the Movable zone begins in each node. Kernel memory
6173 * is spread evenly between nodes as long as the nodes have enough
6174 * memory. When they don't, some nodes will have more kernelcore than
6177 static void __init find_zone_movable_pfns_for_nodes(void)
6180 unsigned long usable_startpfn;
6181 unsigned long kernelcore_node, kernelcore_remaining;
6182 /* save the state before borrow the nodemask */
6183 nodemask_t saved_node_state = node_states[N_MEMORY];
6184 unsigned long totalpages = early_calculate_totalpages();
6185 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6186 struct memblock_region *r;
6188 /* Need to find movable_zone earlier when movable_node is specified. */
6189 find_usable_zone_for_movable();
6192 * If movable_node is specified, ignore kernelcore and movablecore
6195 if (movable_node_is_enabled()) {
6196 for_each_memblock(memory, r) {
6197 if (!memblock_is_hotpluggable(r))
6202 usable_startpfn = PFN_DOWN(r->base);
6203 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6204 min(usable_startpfn, zone_movable_pfn[nid]) :
6212 * If kernelcore=mirror is specified, ignore movablecore option
6214 if (mirrored_kernelcore) {
6215 bool mem_below_4gb_not_mirrored = false;
6217 for_each_memblock(memory, r) {
6218 if (memblock_is_mirror(r))
6223 usable_startpfn = memblock_region_memory_base_pfn(r);
6225 if (usable_startpfn < 0x100000) {
6226 mem_below_4gb_not_mirrored = true;
6230 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6231 min(usable_startpfn, zone_movable_pfn[nid]) :
6235 if (mem_below_4gb_not_mirrored)
6236 pr_warn("This configuration results in unmirrored kernel memory.");
6242 * If movablecore=nn[KMG] was specified, calculate what size of
6243 * kernelcore that corresponds so that memory usable for
6244 * any allocation type is evenly spread. If both kernelcore
6245 * and movablecore are specified, then the value of kernelcore
6246 * will be used for required_kernelcore if it's greater than
6247 * what movablecore would have allowed.
6249 if (required_movablecore) {
6250 unsigned long corepages;
6253 * Round-up so that ZONE_MOVABLE is at least as large as what
6254 * was requested by the user
6256 required_movablecore =
6257 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6258 required_movablecore = min(totalpages, required_movablecore);
6259 corepages = totalpages - required_movablecore;
6261 required_kernelcore = max(required_kernelcore, corepages);
6265 * If kernelcore was not specified or kernelcore size is larger
6266 * than totalpages, there is no ZONE_MOVABLE.
6268 if (!required_kernelcore || required_kernelcore >= totalpages)
6271 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6272 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6275 /* Spread kernelcore memory as evenly as possible throughout nodes */
6276 kernelcore_node = required_kernelcore / usable_nodes;
6277 for_each_node_state(nid, N_MEMORY) {
6278 unsigned long start_pfn, end_pfn;
6281 * Recalculate kernelcore_node if the division per node
6282 * now exceeds what is necessary to satisfy the requested
6283 * amount of memory for the kernel
6285 if (required_kernelcore < kernelcore_node)
6286 kernelcore_node = required_kernelcore / usable_nodes;
6289 * As the map is walked, we track how much memory is usable
6290 * by the kernel using kernelcore_remaining. When it is
6291 * 0, the rest of the node is usable by ZONE_MOVABLE
6293 kernelcore_remaining = kernelcore_node;
6295 /* Go through each range of PFNs within this node */
6296 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6297 unsigned long size_pages;
6299 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6300 if (start_pfn >= end_pfn)
6303 /* Account for what is only usable for kernelcore */
6304 if (start_pfn < usable_startpfn) {
6305 unsigned long kernel_pages;
6306 kernel_pages = min(end_pfn, usable_startpfn)
6309 kernelcore_remaining -= min(kernel_pages,
6310 kernelcore_remaining);
6311 required_kernelcore -= min(kernel_pages,
6312 required_kernelcore);
6314 /* Continue if range is now fully accounted */
6315 if (end_pfn <= usable_startpfn) {
6318 * Push zone_movable_pfn to the end so
6319 * that if we have to rebalance
6320 * kernelcore across nodes, we will
6321 * not double account here
6323 zone_movable_pfn[nid] = end_pfn;
6326 start_pfn = usable_startpfn;
6330 * The usable PFN range for ZONE_MOVABLE is from
6331 * start_pfn->end_pfn. Calculate size_pages as the
6332 * number of pages used as kernelcore
6334 size_pages = end_pfn - start_pfn;
6335 if (size_pages > kernelcore_remaining)
6336 size_pages = kernelcore_remaining;
6337 zone_movable_pfn[nid] = start_pfn + size_pages;
6340 * Some kernelcore has been met, update counts and
6341 * break if the kernelcore for this node has been
6344 required_kernelcore -= min(required_kernelcore,
6346 kernelcore_remaining -= size_pages;
6347 if (!kernelcore_remaining)
6353 * If there is still required_kernelcore, we do another pass with one
6354 * less node in the count. This will push zone_movable_pfn[nid] further
6355 * along on the nodes that still have memory until kernelcore is
6359 if (usable_nodes && required_kernelcore > usable_nodes)
6363 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6364 for (nid = 0; nid < MAX_NUMNODES; nid++)
6365 zone_movable_pfn[nid] =
6366 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6369 /* restore the node_state */
6370 node_states[N_MEMORY] = saved_node_state;
6373 /* Any regular or high memory on that node ? */
6374 static void check_for_memory(pg_data_t *pgdat, int nid)
6376 enum zone_type zone_type;
6378 if (N_MEMORY == N_NORMAL_MEMORY)
6381 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6382 struct zone *zone = &pgdat->node_zones[zone_type];
6383 if (populated_zone(zone)) {
6384 node_set_state(nid, N_HIGH_MEMORY);
6385 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6386 zone_type <= ZONE_NORMAL)
6387 node_set_state(nid, N_NORMAL_MEMORY);
6394 * free_area_init_nodes - Initialise all pg_data_t and zone data
6395 * @max_zone_pfn: an array of max PFNs for each zone
6397 * This will call free_area_init_node() for each active node in the system.
6398 * Using the page ranges provided by memblock_set_node(), the size of each
6399 * zone in each node and their holes is calculated. If the maximum PFN
6400 * between two adjacent zones match, it is assumed that the zone is empty.
6401 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6402 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6403 * starts where the previous one ended. For example, ZONE_DMA32 starts
6404 * at arch_max_dma_pfn.
6406 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6408 unsigned long start_pfn, end_pfn;
6411 /* Record where the zone boundaries are */
6412 memset(arch_zone_lowest_possible_pfn, 0,
6413 sizeof(arch_zone_lowest_possible_pfn));
6414 memset(arch_zone_highest_possible_pfn, 0,
6415 sizeof(arch_zone_highest_possible_pfn));
6417 start_pfn = find_min_pfn_with_active_regions();
6419 for (i = 0; i < MAX_NR_ZONES; i++) {
6420 if (i == ZONE_MOVABLE)
6423 end_pfn = max(max_zone_pfn[i], start_pfn);
6424 arch_zone_lowest_possible_pfn[i] = start_pfn;
6425 arch_zone_highest_possible_pfn[i] = end_pfn;
6427 start_pfn = end_pfn;
6429 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6430 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6432 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6433 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6434 find_zone_movable_pfns_for_nodes();
6436 /* Print out the zone ranges */
6437 pr_info("Zone ranges:\n");
6438 for (i = 0; i < MAX_NR_ZONES; i++) {
6439 if (i == ZONE_MOVABLE)
6441 pr_info(" %-8s ", zone_names[i]);
6442 if (arch_zone_lowest_possible_pfn[i] ==
6443 arch_zone_highest_possible_pfn[i])
6446 pr_cont("[mem %#018Lx-%#018Lx]\n",
6447 (u64)arch_zone_lowest_possible_pfn[i]
6449 ((u64)arch_zone_highest_possible_pfn[i]
6450 << PAGE_SHIFT) - 1);
6453 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6454 pr_info("Movable zone start for each node\n");
6455 for (i = 0; i < MAX_NUMNODES; i++) {
6456 if (zone_movable_pfn[i])
6457 pr_info(" Node %d: %#018Lx\n", i,
6458 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6461 /* Print out the early node map */
6462 pr_info("Early memory node ranges\n");
6463 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6464 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6465 (u64)start_pfn << PAGE_SHIFT,
6466 ((u64)end_pfn << PAGE_SHIFT) - 1);
6468 /* Initialise every node */
6469 mminit_verify_pageflags_layout();
6470 setup_nr_node_ids();
6471 for_each_online_node(nid) {
6472 pg_data_t *pgdat = NODE_DATA(nid);
6473 free_area_init_node(nid, NULL,
6474 find_min_pfn_for_node(nid), NULL);
6476 /* Any memory on that node */
6477 if (pgdat->node_present_pages)
6478 node_set_state(nid, N_MEMORY);
6479 check_for_memory(pgdat, nid);
6483 static int __init cmdline_parse_core(char *p, unsigned long *core)
6485 unsigned long long coremem;
6489 coremem = memparse(p, &p);
6490 *core = coremem >> PAGE_SHIFT;
6492 /* Paranoid check that UL is enough for the coremem value */
6493 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6499 * kernelcore=size sets the amount of memory for use for allocations that
6500 * cannot be reclaimed or migrated.
6502 static int __init cmdline_parse_kernelcore(char *p)
6504 /* parse kernelcore=mirror */
6505 if (parse_option_str(p, "mirror")) {
6506 mirrored_kernelcore = true;
6510 return cmdline_parse_core(p, &required_kernelcore);
6514 * movablecore=size sets the amount of memory for use for allocations that
6515 * can be reclaimed or migrated.
6517 static int __init cmdline_parse_movablecore(char *p)
6519 return cmdline_parse_core(p, &required_movablecore);
6522 early_param("kernelcore", cmdline_parse_kernelcore);
6523 early_param("movablecore", cmdline_parse_movablecore);
6525 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6527 void adjust_managed_page_count(struct page *page, long count)
6529 spin_lock(&managed_page_count_lock);
6530 page_zone(page)->managed_pages += count;
6531 totalram_pages += count;
6532 #ifdef CONFIG_HIGHMEM
6533 if (PageHighMem(page))
6534 totalhigh_pages += count;
6536 spin_unlock(&managed_page_count_lock);
6538 EXPORT_SYMBOL(adjust_managed_page_count);
6540 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6543 unsigned long pages = 0;
6545 start = (void *)PAGE_ALIGN((unsigned long)start);
6546 end = (void *)((unsigned long)end & PAGE_MASK);
6547 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6548 if ((unsigned int)poison <= 0xFF)
6549 memset(pos, poison, PAGE_SIZE);
6550 free_reserved_page(virt_to_page(pos));
6554 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6555 s, pages << (PAGE_SHIFT - 10), start, end);
6559 EXPORT_SYMBOL(free_reserved_area);
6561 #ifdef CONFIG_HIGHMEM
6562 void free_highmem_page(struct page *page)
6564 __free_reserved_page(page);
6566 page_zone(page)->managed_pages++;
6572 void __init mem_init_print_info(const char *str)
6574 unsigned long physpages, codesize, datasize, rosize, bss_size;
6575 unsigned long init_code_size, init_data_size;
6577 physpages = get_num_physpages();
6578 codesize = _etext - _stext;
6579 datasize = _edata - _sdata;
6580 rosize = __end_rodata - __start_rodata;
6581 bss_size = __bss_stop - __bss_start;
6582 init_data_size = __init_end - __init_begin;
6583 init_code_size = _einittext - _sinittext;
6586 * Detect special cases and adjust section sizes accordingly:
6587 * 1) .init.* may be embedded into .data sections
6588 * 2) .init.text.* may be out of [__init_begin, __init_end],
6589 * please refer to arch/tile/kernel/vmlinux.lds.S.
6590 * 3) .rodata.* may be embedded into .text or .data sections.
6592 #define adj_init_size(start, end, size, pos, adj) \
6594 if (start <= pos && pos < end && size > adj) \
6598 adj_init_size(__init_begin, __init_end, init_data_size,
6599 _sinittext, init_code_size);
6600 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6601 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6602 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6603 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6605 #undef adj_init_size
6607 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6608 #ifdef CONFIG_HIGHMEM
6612 nr_free_pages() << (PAGE_SHIFT - 10),
6613 physpages << (PAGE_SHIFT - 10),
6614 codesize >> 10, datasize >> 10, rosize >> 10,
6615 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6616 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6617 totalcma_pages << (PAGE_SHIFT - 10),
6618 #ifdef CONFIG_HIGHMEM
6619 totalhigh_pages << (PAGE_SHIFT - 10),
6621 str ? ", " : "", str ? str : "");
6625 * set_dma_reserve - set the specified number of pages reserved in the first zone
6626 * @new_dma_reserve: The number of pages to mark reserved
6628 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6629 * In the DMA zone, a significant percentage may be consumed by kernel image
6630 * and other unfreeable allocations which can skew the watermarks badly. This
6631 * function may optionally be used to account for unfreeable pages in the
6632 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6633 * smaller per-cpu batchsize.
6635 void __init set_dma_reserve(unsigned long new_dma_reserve)
6637 dma_reserve = new_dma_reserve;
6640 void __init free_area_init(unsigned long *zones_size)
6642 free_area_init_node(0, zones_size,
6643 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6646 static int page_alloc_cpu_notify(struct notifier_block *self,
6647 unsigned long action, void *hcpu)
6649 int cpu = (unsigned long)hcpu;
6651 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6652 lru_add_drain_cpu(cpu);
6656 * Spill the event counters of the dead processor
6657 * into the current processors event counters.
6658 * This artificially elevates the count of the current
6661 vm_events_fold_cpu(cpu);
6664 * Zero the differential counters of the dead processor
6665 * so that the vm statistics are consistent.
6667 * This is only okay since the processor is dead and cannot
6668 * race with what we are doing.
6670 cpu_vm_stats_fold(cpu);
6675 void __init page_alloc_init(void)
6677 hotcpu_notifier(page_alloc_cpu_notify, 0);
6681 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6682 * or min_free_kbytes changes.
6684 static void calculate_totalreserve_pages(void)
6686 struct pglist_data *pgdat;
6687 unsigned long reserve_pages = 0;
6688 enum zone_type i, j;
6690 for_each_online_pgdat(pgdat) {
6692 pgdat->totalreserve_pages = 0;
6694 for (i = 0; i < MAX_NR_ZONES; i++) {
6695 struct zone *zone = pgdat->node_zones + i;
6698 /* Find valid and maximum lowmem_reserve in the zone */
6699 for (j = i; j < MAX_NR_ZONES; j++) {
6700 if (zone->lowmem_reserve[j] > max)
6701 max = zone->lowmem_reserve[j];
6704 /* we treat the high watermark as reserved pages. */
6705 max += high_wmark_pages(zone);
6707 if (max > zone->managed_pages)
6708 max = zone->managed_pages;
6710 pgdat->totalreserve_pages += max;
6712 reserve_pages += max;
6715 totalreserve_pages = reserve_pages;
6719 * setup_per_zone_lowmem_reserve - called whenever
6720 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6721 * has a correct pages reserved value, so an adequate number of
6722 * pages are left in the zone after a successful __alloc_pages().
6724 static void setup_per_zone_lowmem_reserve(void)
6726 struct pglist_data *pgdat;
6727 enum zone_type j, idx;
6729 for_each_online_pgdat(pgdat) {
6730 for (j = 0; j < MAX_NR_ZONES; j++) {
6731 struct zone *zone = pgdat->node_zones + j;
6732 unsigned long managed_pages = zone->managed_pages;
6734 zone->lowmem_reserve[j] = 0;
6738 struct zone *lower_zone;
6742 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6743 sysctl_lowmem_reserve_ratio[idx] = 1;
6745 lower_zone = pgdat->node_zones + idx;
6746 lower_zone->lowmem_reserve[j] = managed_pages /
6747 sysctl_lowmem_reserve_ratio[idx];
6748 managed_pages += lower_zone->managed_pages;
6753 /* update totalreserve_pages */
6754 calculate_totalreserve_pages();
6757 static void __setup_per_zone_wmarks(void)
6759 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6760 unsigned long lowmem_pages = 0;
6762 unsigned long flags;
6764 /* Calculate total number of !ZONE_HIGHMEM pages */
6765 for_each_zone(zone) {
6766 if (!is_highmem(zone))
6767 lowmem_pages += zone->managed_pages;
6770 for_each_zone(zone) {
6773 spin_lock_irqsave(&zone->lock, flags);
6774 tmp = (u64)pages_min * zone->managed_pages;
6775 do_div(tmp, lowmem_pages);
6776 if (is_highmem(zone)) {
6778 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6779 * need highmem pages, so cap pages_min to a small
6782 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6783 * deltas control asynch page reclaim, and so should
6784 * not be capped for highmem.
6786 unsigned long min_pages;
6788 min_pages = zone->managed_pages / 1024;
6789 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6790 zone->watermark[WMARK_MIN] = min_pages;
6793 * If it's a lowmem zone, reserve a number of pages
6794 * proportionate to the zone's size.
6796 zone->watermark[WMARK_MIN] = tmp;
6800 * Set the kswapd watermarks distance according to the
6801 * scale factor in proportion to available memory, but
6802 * ensure a minimum size on small systems.
6804 tmp = max_t(u64, tmp >> 2,
6805 mult_frac(zone->managed_pages,
6806 watermark_scale_factor, 10000));
6808 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6809 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6811 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6812 high_wmark_pages(zone) - low_wmark_pages(zone) -
6813 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6815 spin_unlock_irqrestore(&zone->lock, flags);
6818 /* update totalreserve_pages */
6819 calculate_totalreserve_pages();
6823 * setup_per_zone_wmarks - called when min_free_kbytes changes
6824 * or when memory is hot-{added|removed}
6826 * Ensures that the watermark[min,low,high] values for each zone are set
6827 * correctly with respect to min_free_kbytes.
6829 void setup_per_zone_wmarks(void)
6831 mutex_lock(&zonelists_mutex);
6832 __setup_per_zone_wmarks();
6833 mutex_unlock(&zonelists_mutex);
6837 * Initialise min_free_kbytes.
6839 * For small machines we want it small (128k min). For large machines
6840 * we want it large (64MB max). But it is not linear, because network
6841 * bandwidth does not increase linearly with machine size. We use
6843 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6844 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6860 int __meminit init_per_zone_wmark_min(void)
6862 unsigned long lowmem_kbytes;
6863 int new_min_free_kbytes;
6865 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6866 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6868 if (new_min_free_kbytes > user_min_free_kbytes) {
6869 min_free_kbytes = new_min_free_kbytes;
6870 if (min_free_kbytes < 128)
6871 min_free_kbytes = 128;
6872 if (min_free_kbytes > 65536)
6873 min_free_kbytes = 65536;
6875 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6876 new_min_free_kbytes, user_min_free_kbytes);
6878 setup_per_zone_wmarks();
6879 refresh_zone_stat_thresholds();
6880 setup_per_zone_lowmem_reserve();
6883 core_initcall(init_per_zone_wmark_min)
6886 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6887 * that we can call two helper functions whenever min_free_kbytes
6890 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6891 void __user *buffer, size_t *length, loff_t *ppos)
6895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6900 user_min_free_kbytes = min_free_kbytes;
6901 setup_per_zone_wmarks();
6906 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6907 void __user *buffer, size_t *length, loff_t *ppos)
6911 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6916 setup_per_zone_wmarks();
6922 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6923 void __user *buffer, size_t *length, loff_t *ppos)
6928 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6933 zone->min_unmapped_pages = (zone->managed_pages *
6934 sysctl_min_unmapped_ratio) / 100;
6938 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6939 void __user *buffer, size_t *length, loff_t *ppos)
6944 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6949 zone->min_slab_pages = (zone->managed_pages *
6950 sysctl_min_slab_ratio) / 100;
6956 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6957 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6958 * whenever sysctl_lowmem_reserve_ratio changes.
6960 * The reserve ratio obviously has absolutely no relation with the
6961 * minimum watermarks. The lowmem reserve ratio can only make sense
6962 * if in function of the boot time zone sizes.
6964 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6965 void __user *buffer, size_t *length, loff_t *ppos)
6967 proc_dointvec_minmax(table, write, buffer, length, ppos);
6968 setup_per_zone_lowmem_reserve();
6973 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6974 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6975 * pagelist can have before it gets flushed back to buddy allocator.
6977 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6978 void __user *buffer, size_t *length, loff_t *ppos)
6981 int old_percpu_pagelist_fraction;
6984 mutex_lock(&pcp_batch_high_lock);
6985 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6987 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6988 if (!write || ret < 0)
6991 /* Sanity checking to avoid pcp imbalance */
6992 if (percpu_pagelist_fraction &&
6993 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6994 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7000 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7003 for_each_populated_zone(zone) {
7006 for_each_possible_cpu(cpu)
7007 pageset_set_high_and_batch(zone,
7008 per_cpu_ptr(zone->pageset, cpu));
7011 mutex_unlock(&pcp_batch_high_lock);
7016 int hashdist = HASHDIST_DEFAULT;
7018 static int __init set_hashdist(char *str)
7022 hashdist = simple_strtoul(str, &str, 0);
7025 __setup("hashdist=", set_hashdist);
7029 * allocate a large system hash table from bootmem
7030 * - it is assumed that the hash table must contain an exact power-of-2
7031 * quantity of entries
7032 * - limit is the number of hash buckets, not the total allocation size
7034 void *__init alloc_large_system_hash(const char *tablename,
7035 unsigned long bucketsize,
7036 unsigned long numentries,
7039 unsigned int *_hash_shift,
7040 unsigned int *_hash_mask,
7041 unsigned long low_limit,
7042 unsigned long high_limit)
7044 unsigned long long max = high_limit;
7045 unsigned long log2qty, size;
7048 /* allow the kernel cmdline to have a say */
7050 /* round applicable memory size up to nearest megabyte */
7051 numentries = nr_kernel_pages;
7053 /* It isn't necessary when PAGE_SIZE >= 1MB */
7054 if (PAGE_SHIFT < 20)
7055 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7057 /* limit to 1 bucket per 2^scale bytes of low memory */
7058 if (scale > PAGE_SHIFT)
7059 numentries >>= (scale - PAGE_SHIFT);
7061 numentries <<= (PAGE_SHIFT - scale);
7063 /* Make sure we've got at least a 0-order allocation.. */
7064 if (unlikely(flags & HASH_SMALL)) {
7065 /* Makes no sense without HASH_EARLY */
7066 WARN_ON(!(flags & HASH_EARLY));
7067 if (!(numentries >> *_hash_shift)) {
7068 numentries = 1UL << *_hash_shift;
7069 BUG_ON(!numentries);
7071 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7072 numentries = PAGE_SIZE / bucketsize;
7074 numentries = roundup_pow_of_two(numentries);
7076 /* limit allocation size to 1/16 total memory by default */
7078 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7079 do_div(max, bucketsize);
7081 max = min(max, 0x80000000ULL);
7083 if (numentries < low_limit)
7084 numentries = low_limit;
7085 if (numentries > max)
7088 log2qty = ilog2(numentries);
7091 size = bucketsize << log2qty;
7092 if (flags & HASH_EARLY)
7093 table = memblock_virt_alloc_nopanic(size, 0);
7095 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7098 * If bucketsize is not a power-of-two, we may free
7099 * some pages at the end of hash table which
7100 * alloc_pages_exact() automatically does
7102 if (get_order(size) < MAX_ORDER) {
7103 table = alloc_pages_exact(size, GFP_ATOMIC);
7104 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7107 } while (!table && size > PAGE_SIZE && --log2qty);
7110 panic("Failed to allocate %s hash table\n", tablename);
7112 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7113 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7116 *_hash_shift = log2qty;
7118 *_hash_mask = (1 << log2qty) - 1;
7124 * This function checks whether pageblock includes unmovable pages or not.
7125 * If @count is not zero, it is okay to include less @count unmovable pages
7127 * PageLRU check without isolation or lru_lock could race so that
7128 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7129 * expect this function should be exact.
7131 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7132 bool skip_hwpoisoned_pages)
7134 unsigned long pfn, iter, found;
7138 * For avoiding noise data, lru_add_drain_all() should be called
7139 * If ZONE_MOVABLE, the zone never contains unmovable pages
7141 if (zone_idx(zone) == ZONE_MOVABLE)
7143 mt = get_pageblock_migratetype(page);
7144 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7147 pfn = page_to_pfn(page);
7148 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7149 unsigned long check = pfn + iter;
7151 if (!pfn_valid_within(check))
7154 page = pfn_to_page(check);
7157 * Hugepages are not in LRU lists, but they're movable.
7158 * We need not scan over tail pages bacause we don't
7159 * handle each tail page individually in migration.
7161 if (PageHuge(page)) {
7162 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7167 * We can't use page_count without pin a page
7168 * because another CPU can free compound page.
7169 * This check already skips compound tails of THP
7170 * because their page->_refcount is zero at all time.
7172 if (!page_ref_count(page)) {
7173 if (PageBuddy(page))
7174 iter += (1 << page_order(page)) - 1;
7179 * The HWPoisoned page may be not in buddy system, and
7180 * page_count() is not 0.
7182 if (skip_hwpoisoned_pages && PageHWPoison(page))
7188 * If there are RECLAIMABLE pages, we need to check
7189 * it. But now, memory offline itself doesn't call
7190 * shrink_node_slabs() and it still to be fixed.
7193 * If the page is not RAM, page_count()should be 0.
7194 * we don't need more check. This is an _used_ not-movable page.
7196 * The problematic thing here is PG_reserved pages. PG_reserved
7197 * is set to both of a memory hole page and a _used_ kernel
7206 bool is_pageblock_removable_nolock(struct page *page)
7212 * We have to be careful here because we are iterating over memory
7213 * sections which are not zone aware so we might end up outside of
7214 * the zone but still within the section.
7215 * We have to take care about the node as well. If the node is offline
7216 * its NODE_DATA will be NULL - see page_zone.
7218 if (!node_online(page_to_nid(page)))
7221 zone = page_zone(page);
7222 pfn = page_to_pfn(page);
7223 if (!zone_spans_pfn(zone, pfn))
7226 return !has_unmovable_pages(zone, page, 0, true);
7229 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7231 static unsigned long pfn_max_align_down(unsigned long pfn)
7233 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7234 pageblock_nr_pages) - 1);
7237 static unsigned long pfn_max_align_up(unsigned long pfn)
7239 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7240 pageblock_nr_pages));
7243 /* [start, end) must belong to a single zone. */
7244 static int __alloc_contig_migrate_range(struct compact_control *cc,
7245 unsigned long start, unsigned long end)
7247 /* This function is based on compact_zone() from compaction.c. */
7248 unsigned long nr_reclaimed;
7249 unsigned long pfn = start;
7250 unsigned int tries = 0;
7255 while (pfn < end || !list_empty(&cc->migratepages)) {
7256 if (fatal_signal_pending(current)) {
7261 if (list_empty(&cc->migratepages)) {
7262 cc->nr_migratepages = 0;
7263 pfn = isolate_migratepages_range(cc, pfn, end);
7269 } else if (++tries == 5) {
7270 ret = ret < 0 ? ret : -EBUSY;
7274 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7276 cc->nr_migratepages -= nr_reclaimed;
7278 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7279 NULL, 0, cc->mode, MR_CMA);
7282 putback_movable_pages(&cc->migratepages);
7289 * alloc_contig_range() -- tries to allocate given range of pages
7290 * @start: start PFN to allocate
7291 * @end: one-past-the-last PFN to allocate
7292 * @migratetype: migratetype of the underlaying pageblocks (either
7293 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7294 * in range must have the same migratetype and it must
7295 * be either of the two.
7297 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7298 * aligned, however it's the caller's responsibility to guarantee that
7299 * we are the only thread that changes migrate type of pageblocks the
7302 * The PFN range must belong to a single zone.
7304 * Returns zero on success or negative error code. On success all
7305 * pages which PFN is in [start, end) are allocated for the caller and
7306 * need to be freed with free_contig_range().
7308 int alloc_contig_range(unsigned long start, unsigned long end,
7309 unsigned migratetype)
7311 unsigned long outer_start, outer_end;
7315 struct compact_control cc = {
7316 .nr_migratepages = 0,
7318 .zone = page_zone(pfn_to_page(start)),
7319 .mode = MIGRATE_SYNC,
7320 .ignore_skip_hint = true,
7322 INIT_LIST_HEAD(&cc.migratepages);
7325 * What we do here is we mark all pageblocks in range as
7326 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7327 * have different sizes, and due to the way page allocator
7328 * work, we align the range to biggest of the two pages so
7329 * that page allocator won't try to merge buddies from
7330 * different pageblocks and change MIGRATE_ISOLATE to some
7331 * other migration type.
7333 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7334 * migrate the pages from an unaligned range (ie. pages that
7335 * we are interested in). This will put all the pages in
7336 * range back to page allocator as MIGRATE_ISOLATE.
7338 * When this is done, we take the pages in range from page
7339 * allocator removing them from the buddy system. This way
7340 * page allocator will never consider using them.
7342 * This lets us mark the pageblocks back as
7343 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7344 * aligned range but not in the unaligned, original range are
7345 * put back to page allocator so that buddy can use them.
7348 ret = start_isolate_page_range(pfn_max_align_down(start),
7349 pfn_max_align_up(end), migratetype,
7355 * In case of -EBUSY, we'd like to know which page causes problem.
7356 * So, just fall through. We will check it in test_pages_isolated().
7358 ret = __alloc_contig_migrate_range(&cc, start, end);
7359 if (ret && ret != -EBUSY)
7363 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7364 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7365 * more, all pages in [start, end) are free in page allocator.
7366 * What we are going to do is to allocate all pages from
7367 * [start, end) (that is remove them from page allocator).
7369 * The only problem is that pages at the beginning and at the
7370 * end of interesting range may be not aligned with pages that
7371 * page allocator holds, ie. they can be part of higher order
7372 * pages. Because of this, we reserve the bigger range and
7373 * once this is done free the pages we are not interested in.
7375 * We don't have to hold zone->lock here because the pages are
7376 * isolated thus they won't get removed from buddy.
7379 lru_add_drain_all();
7380 drain_all_pages(cc.zone);
7383 outer_start = start;
7384 while (!PageBuddy(pfn_to_page(outer_start))) {
7385 if (++order >= MAX_ORDER) {
7386 outer_start = start;
7389 outer_start &= ~0UL << order;
7392 if (outer_start != start) {
7393 order = page_order(pfn_to_page(outer_start));
7396 * outer_start page could be small order buddy page and
7397 * it doesn't include start page. Adjust outer_start
7398 * in this case to report failed page properly
7399 * on tracepoint in test_pages_isolated()
7401 if (outer_start + (1UL << order) <= start)
7402 outer_start = start;
7405 /* Make sure the range is really isolated. */
7406 if (test_pages_isolated(outer_start, end, false)) {
7407 pr_info("%s: [%lx, %lx) PFNs busy\n",
7408 __func__, outer_start, end);
7413 /* Grab isolated pages from freelists. */
7414 outer_end = isolate_freepages_range(&cc, outer_start, end);
7420 /* Free head and tail (if any) */
7421 if (start != outer_start)
7422 free_contig_range(outer_start, start - outer_start);
7423 if (end != outer_end)
7424 free_contig_range(end, outer_end - end);
7427 undo_isolate_page_range(pfn_max_align_down(start),
7428 pfn_max_align_up(end), migratetype);
7432 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7434 unsigned int count = 0;
7436 for (; nr_pages--; pfn++) {
7437 struct page *page = pfn_to_page(pfn);
7439 count += page_count(page) != 1;
7442 WARN(count != 0, "%d pages are still in use!\n", count);
7446 #ifdef CONFIG_MEMORY_HOTPLUG
7448 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7449 * page high values need to be recalulated.
7451 void __meminit zone_pcp_update(struct zone *zone)
7454 mutex_lock(&pcp_batch_high_lock);
7455 for_each_possible_cpu(cpu)
7456 pageset_set_high_and_batch(zone,
7457 per_cpu_ptr(zone->pageset, cpu));
7458 mutex_unlock(&pcp_batch_high_lock);
7462 void zone_pcp_reset(struct zone *zone)
7464 unsigned long flags;
7466 struct per_cpu_pageset *pset;
7468 /* avoid races with drain_pages() */
7469 local_irq_save(flags);
7470 if (zone->pageset != &boot_pageset) {
7471 for_each_online_cpu(cpu) {
7472 pset = per_cpu_ptr(zone->pageset, cpu);
7473 drain_zonestat(zone, pset);
7475 free_percpu(zone->pageset);
7476 zone->pageset = &boot_pageset;
7478 local_irq_restore(flags);
7481 #ifdef CONFIG_MEMORY_HOTREMOVE
7483 * All pages in the range must be in a single zone and isolated
7484 * before calling this.
7487 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7491 unsigned int order, i;
7493 unsigned long flags;
7494 /* find the first valid pfn */
7495 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7500 zone = page_zone(pfn_to_page(pfn));
7501 spin_lock_irqsave(&zone->lock, flags);
7503 while (pfn < end_pfn) {
7504 if (!pfn_valid(pfn)) {
7508 page = pfn_to_page(pfn);
7510 * The HWPoisoned page may be not in buddy system, and
7511 * page_count() is not 0.
7513 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7515 SetPageReserved(page);
7519 BUG_ON(page_count(page));
7520 BUG_ON(!PageBuddy(page));
7521 order = page_order(page);
7522 #ifdef CONFIG_DEBUG_VM
7523 pr_info("remove from free list %lx %d %lx\n",
7524 pfn, 1 << order, end_pfn);
7526 list_del(&page->lru);
7527 rmv_page_order(page);
7528 zone->free_area[order].nr_free--;
7529 for (i = 0; i < (1 << order); i++)
7530 SetPageReserved((page+i));
7531 pfn += (1 << order);
7533 spin_unlock_irqrestore(&zone->lock, flags);
7537 bool is_free_buddy_page(struct page *page)
7539 struct zone *zone = page_zone(page);
7540 unsigned long pfn = page_to_pfn(page);
7541 unsigned long flags;
7544 spin_lock_irqsave(&zone->lock, flags);
7545 for (order = 0; order < MAX_ORDER; order++) {
7546 struct page *page_head = page - (pfn & ((1 << order) - 1));
7548 if (PageBuddy(page_head) && page_order(page_head) >= order)
7551 spin_unlock_irqrestore(&zone->lock, flags);
7553 return order < MAX_ORDER;