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 list_del(&page->lru);
2593 } while (check_new_pcp(page));
2596 * We most definitely don't want callers attempting to
2597 * allocate greater than order-1 page units with __GFP_NOFAIL.
2599 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2600 spin_lock_irqsave(&zone->lock, flags);
2604 if (alloc_flags & ALLOC_HARDER) {
2605 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2607 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2610 page = __rmqueue(zone, order, migratetype);
2611 } while (page && check_new_pages(page, order));
2612 spin_unlock(&zone->lock);
2615 __mod_zone_freepage_state(zone, -(1 << order),
2616 get_pcppage_migratetype(page));
2619 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2620 zone_statistics(preferred_zone, zone, gfp_flags);
2621 local_irq_restore(flags);
2623 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2627 local_irq_restore(flags);
2631 #ifdef CONFIG_FAIL_PAGE_ALLOC
2634 struct fault_attr attr;
2636 bool ignore_gfp_highmem;
2637 bool ignore_gfp_reclaim;
2639 } fail_page_alloc = {
2640 .attr = FAULT_ATTR_INITIALIZER,
2641 .ignore_gfp_reclaim = true,
2642 .ignore_gfp_highmem = true,
2646 static int __init setup_fail_page_alloc(char *str)
2648 return setup_fault_attr(&fail_page_alloc.attr, str);
2650 __setup("fail_page_alloc=", setup_fail_page_alloc);
2652 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2654 if (order < fail_page_alloc.min_order)
2656 if (gfp_mask & __GFP_NOFAIL)
2658 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2660 if (fail_page_alloc.ignore_gfp_reclaim &&
2661 (gfp_mask & __GFP_DIRECT_RECLAIM))
2664 return should_fail(&fail_page_alloc.attr, 1 << order);
2667 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2669 static int __init fail_page_alloc_debugfs(void)
2671 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2674 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2675 &fail_page_alloc.attr);
2677 return PTR_ERR(dir);
2679 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2680 &fail_page_alloc.ignore_gfp_reclaim))
2682 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2683 &fail_page_alloc.ignore_gfp_highmem))
2685 if (!debugfs_create_u32("min-order", mode, dir,
2686 &fail_page_alloc.min_order))
2691 debugfs_remove_recursive(dir);
2696 late_initcall(fail_page_alloc_debugfs);
2698 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2700 #else /* CONFIG_FAIL_PAGE_ALLOC */
2702 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2707 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2710 * Return true if free base pages are above 'mark'. For high-order checks it
2711 * will return true of the order-0 watermark is reached and there is at least
2712 * one free page of a suitable size. Checking now avoids taking the zone lock
2713 * to check in the allocation paths if no pages are free.
2715 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2716 int classzone_idx, unsigned int alloc_flags,
2721 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2723 /* free_pages may go negative - that's OK */
2724 free_pages -= (1 << order) - 1;
2726 if (alloc_flags & ALLOC_HIGH)
2730 * If the caller does not have rights to ALLOC_HARDER then subtract
2731 * the high-atomic reserves. This will over-estimate the size of the
2732 * atomic reserve but it avoids a search.
2734 if (likely(!alloc_harder))
2735 free_pages -= z->nr_reserved_highatomic;
2740 /* If allocation can't use CMA areas don't use free CMA pages */
2741 if (!(alloc_flags & ALLOC_CMA))
2742 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2746 * Check watermarks for an order-0 allocation request. If these
2747 * are not met, then a high-order request also cannot go ahead
2748 * even if a suitable page happened to be free.
2750 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2753 /* If this is an order-0 request then the watermark is fine */
2757 /* For a high-order request, check at least one suitable page is free */
2758 for (o = order; o < MAX_ORDER; o++) {
2759 struct free_area *area = &z->free_area[o];
2768 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2769 if (!list_empty(&area->free_list[mt]))
2774 if ((alloc_flags & ALLOC_CMA) &&
2775 !list_empty(&area->free_list[MIGRATE_CMA])) {
2783 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2784 int classzone_idx, unsigned int alloc_flags)
2786 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2787 zone_page_state(z, NR_FREE_PAGES));
2790 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2791 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2793 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2797 /* If allocation can't use CMA areas don't use free CMA pages */
2798 if (!(alloc_flags & ALLOC_CMA))
2799 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2803 * Fast check for order-0 only. If this fails then the reserves
2804 * need to be calculated. There is a corner case where the check
2805 * passes but only the high-order atomic reserve are free. If
2806 * the caller is !atomic then it'll uselessly search the free
2807 * list. That corner case is then slower but it is harmless.
2809 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2812 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2816 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2817 unsigned long mark, int classzone_idx)
2819 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2821 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2822 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2824 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2829 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2831 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2834 #else /* CONFIG_NUMA */
2835 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2839 #endif /* CONFIG_NUMA */
2842 * get_page_from_freelist goes through the zonelist trying to allocate
2845 static struct page *
2846 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2847 const struct alloc_context *ac)
2849 struct zoneref *z = ac->preferred_zoneref;
2851 struct pglist_data *last_pgdat_dirty_limit = NULL;
2854 * Scan zonelist, looking for a zone with enough free.
2855 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2857 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2862 if (cpusets_enabled() &&
2863 (alloc_flags & ALLOC_CPUSET) &&
2864 !__cpuset_zone_allowed(zone, gfp_mask))
2867 * When allocating a page cache page for writing, we
2868 * want to get it from a node that is within its dirty
2869 * limit, such that no single node holds more than its
2870 * proportional share of globally allowed dirty pages.
2871 * The dirty limits take into account the node's
2872 * lowmem reserves and high watermark so that kswapd
2873 * should be able to balance it without having to
2874 * write pages from its LRU list.
2876 * XXX: For now, allow allocations to potentially
2877 * exceed the per-node dirty limit in the slowpath
2878 * (spread_dirty_pages unset) before going into reclaim,
2879 * which is important when on a NUMA setup the allowed
2880 * nodes are together not big enough to reach the
2881 * global limit. The proper fix for these situations
2882 * will require awareness of nodes in the
2883 * dirty-throttling and the flusher threads.
2885 if (ac->spread_dirty_pages) {
2886 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2889 if (!node_dirty_ok(zone->zone_pgdat)) {
2890 last_pgdat_dirty_limit = zone->zone_pgdat;
2895 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2896 if (!zone_watermark_fast(zone, order, mark,
2897 ac_classzone_idx(ac), alloc_flags)) {
2900 /* Checked here to keep the fast path fast */
2901 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2902 if (alloc_flags & ALLOC_NO_WATERMARKS)
2905 if (node_reclaim_mode == 0 ||
2906 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2909 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2911 case NODE_RECLAIM_NOSCAN:
2914 case NODE_RECLAIM_FULL:
2915 /* scanned but unreclaimable */
2918 /* did we reclaim enough */
2919 if (zone_watermark_ok(zone, order, mark,
2920 ac_classzone_idx(ac), alloc_flags))
2928 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2929 gfp_mask, alloc_flags, ac->migratetype);
2931 prep_new_page(page, order, gfp_mask, alloc_flags);
2934 * If this is a high-order atomic allocation then check
2935 * if the pageblock should be reserved for the future
2937 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2938 reserve_highatomic_pageblock(page, zone, order);
2948 * Large machines with many possible nodes should not always dump per-node
2949 * meminfo in irq context.
2951 static inline bool should_suppress_show_mem(void)
2956 ret = in_interrupt();
2961 static DEFINE_RATELIMIT_STATE(nopage_rs,
2962 DEFAULT_RATELIMIT_INTERVAL,
2963 DEFAULT_RATELIMIT_BURST);
2965 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2967 unsigned int filter = SHOW_MEM_FILTER_NODES;
2969 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2970 debug_guardpage_minorder() > 0)
2974 * This documents exceptions given to allocations in certain
2975 * contexts that are allowed to allocate outside current's set
2978 if (!(gfp_mask & __GFP_NOMEMALLOC))
2979 if (test_thread_flag(TIF_MEMDIE) ||
2980 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2981 filter &= ~SHOW_MEM_FILTER_NODES;
2982 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2983 filter &= ~SHOW_MEM_FILTER_NODES;
2986 struct va_format vaf;
2989 va_start(args, fmt);
2994 pr_warn("%pV", &vaf);
2999 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3000 current->comm, order, gfp_mask, &gfp_mask);
3002 if (!should_suppress_show_mem())
3006 static inline struct page *
3007 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3008 const struct alloc_context *ac, unsigned long *did_some_progress)
3010 struct oom_control oc = {
3011 .zonelist = ac->zonelist,
3012 .nodemask = ac->nodemask,
3014 .gfp_mask = gfp_mask,
3019 *did_some_progress = 0;
3022 * Acquire the oom lock. If that fails, somebody else is
3023 * making progress for us.
3025 if (!mutex_trylock(&oom_lock)) {
3026 *did_some_progress = 1;
3027 schedule_timeout_uninterruptible(1);
3032 * Go through the zonelist yet one more time, keep very high watermark
3033 * here, this is only to catch a parallel oom killing, we must fail if
3034 * we're still under heavy pressure.
3036 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3037 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3041 if (!(gfp_mask & __GFP_NOFAIL)) {
3042 /* Coredumps can quickly deplete all memory reserves */
3043 if (current->flags & PF_DUMPCORE)
3045 /* The OOM killer will not help higher order allocs */
3046 if (order > PAGE_ALLOC_COSTLY_ORDER)
3048 /* The OOM killer does not needlessly kill tasks for lowmem */
3049 if (ac->high_zoneidx < ZONE_NORMAL)
3051 if (pm_suspended_storage())
3054 * XXX: GFP_NOFS allocations should rather fail than rely on
3055 * other request to make a forward progress.
3056 * We are in an unfortunate situation where out_of_memory cannot
3057 * do much for this context but let's try it to at least get
3058 * access to memory reserved if the current task is killed (see
3059 * out_of_memory). Once filesystems are ready to handle allocation
3060 * failures more gracefully we should just bail out here.
3063 /* The OOM killer may not free memory on a specific node */
3064 if (gfp_mask & __GFP_THISNODE)
3067 /* Exhausted what can be done so it's blamo time */
3068 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3069 *did_some_progress = 1;
3071 if (gfp_mask & __GFP_NOFAIL) {
3072 page = get_page_from_freelist(gfp_mask, order,
3073 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3075 * fallback to ignore cpuset restriction if our nodes
3079 page = get_page_from_freelist(gfp_mask, order,
3080 ALLOC_NO_WATERMARKS, ac);
3084 mutex_unlock(&oom_lock);
3089 * Maximum number of compaction retries wit a progress before OOM
3090 * killer is consider as the only way to move forward.
3092 #define MAX_COMPACT_RETRIES 16
3094 #ifdef CONFIG_COMPACTION
3095 /* Try memory compaction for high-order allocations before reclaim */
3096 static struct page *
3097 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3098 unsigned int alloc_flags, const struct alloc_context *ac,
3099 enum migrate_mode mode, enum compact_result *compact_result)
3102 int contended_compaction;
3107 current->flags |= PF_MEMALLOC;
3108 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3109 mode, &contended_compaction);
3110 current->flags &= ~PF_MEMALLOC;
3112 if (*compact_result <= COMPACT_INACTIVE)
3116 * At least in one zone compaction wasn't deferred or skipped, so let's
3117 * count a compaction stall
3119 count_vm_event(COMPACTSTALL);
3121 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3124 struct zone *zone = page_zone(page);
3126 zone->compact_blockskip_flush = false;
3127 compaction_defer_reset(zone, order, true);
3128 count_vm_event(COMPACTSUCCESS);
3133 * It's bad if compaction run occurs and fails. The most likely reason
3134 * is that pages exist, but not enough to satisfy watermarks.
3136 count_vm_event(COMPACTFAIL);
3139 * In all zones where compaction was attempted (and not
3140 * deferred or skipped), lock contention has been detected.
3141 * For THP allocation we do not want to disrupt the others
3142 * so we fallback to base pages instead.
3144 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3145 *compact_result = COMPACT_CONTENDED;
3148 * If compaction was aborted due to need_resched(), we do not
3149 * want to further increase allocation latency, unless it is
3150 * khugepaged trying to collapse.
3152 if (contended_compaction == COMPACT_CONTENDED_SCHED
3153 && !(current->flags & PF_KTHREAD))
3154 *compact_result = COMPACT_CONTENDED;
3162 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3163 enum compact_result compact_result, enum migrate_mode *migrate_mode,
3164 int compaction_retries)
3166 int max_retries = MAX_COMPACT_RETRIES;
3172 * compaction considers all the zone as desperately out of memory
3173 * so it doesn't really make much sense to retry except when the
3174 * failure could be caused by weak migration mode.
3176 if (compaction_failed(compact_result)) {
3177 if (*migrate_mode == MIGRATE_ASYNC) {
3178 *migrate_mode = MIGRATE_SYNC_LIGHT;
3185 * make sure the compaction wasn't deferred or didn't bail out early
3186 * due to locks contention before we declare that we should give up.
3187 * But do not retry if the given zonelist is not suitable for
3190 if (compaction_withdrawn(compact_result))
3191 return compaction_zonelist_suitable(ac, order, alloc_flags);
3194 * !costly requests are much more important than __GFP_REPEAT
3195 * costly ones because they are de facto nofail and invoke OOM
3196 * killer to move on while costly can fail and users are ready
3197 * to cope with that. 1/4 retries is rather arbitrary but we
3198 * would need much more detailed feedback from compaction to
3199 * make a better decision.
3201 if (order > PAGE_ALLOC_COSTLY_ORDER)
3203 if (compaction_retries <= max_retries)
3209 static inline struct page *
3210 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3211 unsigned int alloc_flags, const struct alloc_context *ac,
3212 enum migrate_mode mode, enum compact_result *compact_result)
3214 *compact_result = COMPACT_SKIPPED;
3219 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3220 enum compact_result compact_result,
3221 enum migrate_mode *migrate_mode,
3222 int compaction_retries)
3227 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3231 * There are setups with compaction disabled which would prefer to loop
3232 * inside the allocator rather than hit the oom killer prematurely.
3233 * Let's give them a good hope and keep retrying while the order-0
3234 * watermarks are OK.
3236 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3238 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3239 ac_classzone_idx(ac), alloc_flags))
3244 #endif /* CONFIG_COMPACTION */
3246 /* Perform direct synchronous page reclaim */
3248 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3249 const struct alloc_context *ac)
3251 struct reclaim_state reclaim_state;
3256 /* We now go into synchronous reclaim */
3257 cpuset_memory_pressure_bump();
3258 current->flags |= PF_MEMALLOC;
3259 lockdep_set_current_reclaim_state(gfp_mask);
3260 reclaim_state.reclaimed_slab = 0;
3261 current->reclaim_state = &reclaim_state;
3263 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3266 current->reclaim_state = NULL;
3267 lockdep_clear_current_reclaim_state();
3268 current->flags &= ~PF_MEMALLOC;
3275 /* The really slow allocator path where we enter direct reclaim */
3276 static inline struct page *
3277 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3278 unsigned int alloc_flags, const struct alloc_context *ac,
3279 unsigned long *did_some_progress)
3281 struct page *page = NULL;
3282 bool drained = false;
3284 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3285 if (unlikely(!(*did_some_progress)))
3289 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3292 * If an allocation failed after direct reclaim, it could be because
3293 * pages are pinned on the per-cpu lists or in high alloc reserves.
3294 * Shrink them them and try again
3296 if (!page && !drained) {
3297 unreserve_highatomic_pageblock(ac);
3298 drain_all_pages(NULL);
3306 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3310 pg_data_t *last_pgdat = NULL;
3312 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3313 ac->high_zoneidx, ac->nodemask) {
3314 if (last_pgdat != zone->zone_pgdat)
3315 wakeup_kswapd(zone, order, ac->high_zoneidx);
3316 last_pgdat = zone->zone_pgdat;
3320 static inline unsigned int
3321 gfp_to_alloc_flags(gfp_t gfp_mask)
3323 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3325 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3326 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3329 * The caller may dip into page reserves a bit more if the caller
3330 * cannot run direct reclaim, or if the caller has realtime scheduling
3331 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3332 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3334 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3336 if (gfp_mask & __GFP_ATOMIC) {
3338 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3339 * if it can't schedule.
3341 if (!(gfp_mask & __GFP_NOMEMALLOC))
3342 alloc_flags |= ALLOC_HARDER;
3344 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3345 * comment for __cpuset_node_allowed().
3347 alloc_flags &= ~ALLOC_CPUSET;
3348 } else if (unlikely(rt_task(current)) && !in_interrupt())
3349 alloc_flags |= ALLOC_HARDER;
3352 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3353 alloc_flags |= ALLOC_CMA;
3358 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3360 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3363 if (gfp_mask & __GFP_MEMALLOC)
3365 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3367 if (!in_interrupt() &&
3368 ((current->flags & PF_MEMALLOC) ||
3369 unlikely(test_thread_flag(TIF_MEMDIE))))
3376 * Maximum number of reclaim retries without any progress before OOM killer
3377 * is consider as the only way to move forward.
3379 #define MAX_RECLAIM_RETRIES 16
3382 * Checks whether it makes sense to retry the reclaim to make a forward progress
3383 * for the given allocation request.
3384 * The reclaim feedback represented by did_some_progress (any progress during
3385 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3386 * any progress in a row) is considered as well as the reclaimable pages on the
3387 * applicable zone list (with a backoff mechanism which is a function of
3388 * no_progress_loops).
3390 * Returns true if a retry is viable or false to enter the oom path.
3393 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3394 struct alloc_context *ac, int alloc_flags,
3395 bool did_some_progress, int no_progress_loops)
3401 * Make sure we converge to OOM if we cannot make any progress
3402 * several times in the row.
3404 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3408 * Keep reclaiming pages while there is a chance this will lead
3409 * somewhere. If none of the target zones can satisfy our allocation
3410 * request even if all reclaimable pages are considered then we are
3411 * screwed and have to go OOM.
3413 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3415 unsigned long available;
3416 unsigned long reclaimable;
3418 available = reclaimable = zone_reclaimable_pages(zone);
3419 available -= DIV_ROUND_UP(no_progress_loops * available,
3420 MAX_RECLAIM_RETRIES);
3421 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3424 * Would the allocation succeed if we reclaimed the whole
3427 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3428 ac_classzone_idx(ac), alloc_flags, available)) {
3430 * If we didn't make any progress and have a lot of
3431 * dirty + writeback pages then we should wait for
3432 * an IO to complete to slow down the reclaim and
3433 * prevent from pre mature OOM
3435 if (!did_some_progress) {
3436 unsigned long write_pending;
3438 write_pending = zone_page_state_snapshot(zone,
3439 NR_ZONE_WRITE_PENDING);
3441 if (2 * write_pending > reclaimable) {
3442 congestion_wait(BLK_RW_ASYNC, HZ/10);
3448 * Memory allocation/reclaim might be called from a WQ
3449 * context and the current implementation of the WQ
3450 * concurrency control doesn't recognize that
3451 * a particular WQ is congested if the worker thread is
3452 * looping without ever sleeping. Therefore we have to
3453 * do a short sleep here rather than calling
3456 if (current->flags & PF_WQ_WORKER)
3457 schedule_timeout_uninterruptible(1);
3468 static inline struct page *
3469 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3470 struct alloc_context *ac)
3472 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3473 struct page *page = NULL;
3474 unsigned int alloc_flags;
3475 unsigned long did_some_progress;
3476 enum migrate_mode migration_mode = MIGRATE_SYNC_LIGHT;
3477 enum compact_result compact_result;
3478 int compaction_retries = 0;
3479 int no_progress_loops = 0;
3482 * In the slowpath, we sanity check order to avoid ever trying to
3483 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3484 * be using allocators in order of preference for an area that is
3487 if (order >= MAX_ORDER) {
3488 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3493 * We also sanity check to catch abuse of atomic reserves being used by
3494 * callers that are not in atomic context.
3496 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3497 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3498 gfp_mask &= ~__GFP_ATOMIC;
3501 * The fast path uses conservative alloc_flags to succeed only until
3502 * kswapd needs to be woken up, and to avoid the cost of setting up
3503 * alloc_flags precisely. So we do that now.
3505 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3507 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3508 wake_all_kswapds(order, ac);
3511 * The adjusted alloc_flags might result in immediate success, so try
3514 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3519 * For costly allocations, try direct compaction first, as it's likely
3520 * that we have enough base pages and don't need to reclaim. Don't try
3521 * that for allocations that are allowed to ignore watermarks, as the
3522 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3524 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3525 !gfp_pfmemalloc_allowed(gfp_mask)) {
3526 page = __alloc_pages_direct_compact(gfp_mask, order,
3534 * Checks for costly allocations with __GFP_NORETRY, which
3535 * includes THP page fault allocations
3537 if (gfp_mask & __GFP_NORETRY) {
3539 * If compaction is deferred for high-order allocations,
3540 * it is because sync compaction recently failed. If
3541 * this is the case and the caller requested a THP
3542 * allocation, we do not want to heavily disrupt the
3543 * system, so we fail the allocation instead of entering
3546 if (compact_result == COMPACT_DEFERRED)
3550 * Compaction is contended so rather back off than cause
3553 if (compact_result == COMPACT_CONTENDED)
3557 * Looks like reclaim/compaction is worth trying, but
3558 * sync compaction could be very expensive, so keep
3559 * using async compaction, unless it's khugepaged
3560 * trying to collapse.
3562 if (!(current->flags & PF_KTHREAD))
3563 migration_mode = MIGRATE_ASYNC;
3568 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3569 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3570 wake_all_kswapds(order, ac);
3572 if (gfp_pfmemalloc_allowed(gfp_mask))
3573 alloc_flags = ALLOC_NO_WATERMARKS;
3576 * Reset the zonelist iterators if memory policies can be ignored.
3577 * These allocations are high priority and system rather than user
3580 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3581 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3582 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3583 ac->high_zoneidx, ac->nodemask);
3586 /* Attempt with potentially adjusted zonelist and alloc_flags */
3587 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3591 /* Caller is not willing to reclaim, we can't balance anything */
3592 if (!can_direct_reclaim) {
3594 * All existing users of the __GFP_NOFAIL are blockable, so warn
3595 * of any new users that actually allow this type of allocation
3598 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3602 /* Avoid recursion of direct reclaim */
3603 if (current->flags & PF_MEMALLOC) {
3605 * __GFP_NOFAIL request from this context is rather bizarre
3606 * because we cannot reclaim anything and only can loop waiting
3607 * for somebody to do a work for us.
3609 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3616 /* Avoid allocations with no watermarks from looping endlessly */
3617 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3621 /* Try direct reclaim and then allocating */
3622 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3623 &did_some_progress);
3627 /* Try direct compaction and then allocating */
3628 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3634 if (order && compaction_made_progress(compact_result))
3635 compaction_retries++;
3637 /* Do not loop if specifically requested */
3638 if (gfp_mask & __GFP_NORETRY)
3642 * Do not retry costly high order allocations unless they are
3645 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3649 * Costly allocations might have made a progress but this doesn't mean
3650 * their order will become available due to high fragmentation so
3651 * always increment the no progress counter for them
3653 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3654 no_progress_loops = 0;
3656 no_progress_loops++;
3658 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3659 did_some_progress > 0, no_progress_loops))
3663 * It doesn't make any sense to retry for the compaction if the order-0
3664 * reclaim is not able to make any progress because the current
3665 * implementation of the compaction depends on the sufficient amount
3666 * of free memory (see __compaction_suitable)
3668 if (did_some_progress > 0 &&
3669 should_compact_retry(ac, order, alloc_flags,
3670 compact_result, &migration_mode,
3671 compaction_retries))
3674 /* Reclaim has failed us, start killing things */
3675 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3679 /* Retry as long as the OOM killer is making progress */
3680 if (did_some_progress) {
3681 no_progress_loops = 0;
3686 warn_alloc_failed(gfp_mask, order, NULL);
3692 * This is the 'heart' of the zoned buddy allocator.
3695 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3696 struct zonelist *zonelist, nodemask_t *nodemask)
3699 unsigned int cpuset_mems_cookie;
3700 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3701 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3702 struct alloc_context ac = {
3703 .high_zoneidx = gfp_zone(gfp_mask),
3704 .zonelist = zonelist,
3705 .nodemask = nodemask,
3706 .migratetype = gfpflags_to_migratetype(gfp_mask),
3709 if (cpusets_enabled()) {
3710 alloc_mask |= __GFP_HARDWALL;
3711 alloc_flags |= ALLOC_CPUSET;
3713 ac.nodemask = &cpuset_current_mems_allowed;
3716 gfp_mask &= gfp_allowed_mask;
3718 lockdep_trace_alloc(gfp_mask);
3720 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3722 if (should_fail_alloc_page(gfp_mask, order))
3726 * Check the zones suitable for the gfp_mask contain at least one
3727 * valid zone. It's possible to have an empty zonelist as a result
3728 * of __GFP_THISNODE and a memoryless node
3730 if (unlikely(!zonelist->_zonerefs->zone))
3733 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3734 alloc_flags |= ALLOC_CMA;
3737 cpuset_mems_cookie = read_mems_allowed_begin();
3739 /* Dirty zone balancing only done in the fast path */
3740 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3743 * The preferred zone is used for statistics but crucially it is
3744 * also used as the starting point for the zonelist iterator. It
3745 * may get reset for allocations that ignore memory policies.
3747 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3748 ac.high_zoneidx, ac.nodemask);
3749 if (!ac.preferred_zoneref) {
3754 /* First allocation attempt */
3755 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3760 * Runtime PM, block IO and its error handling path can deadlock
3761 * because I/O on the device might not complete.
3763 alloc_mask = memalloc_noio_flags(gfp_mask);
3764 ac.spread_dirty_pages = false;
3767 * Restore the original nodemask if it was potentially replaced with
3768 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3770 if (cpusets_enabled())
3771 ac.nodemask = nodemask;
3772 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3776 * When updating a task's mems_allowed, it is possible to race with
3777 * parallel threads in such a way that an allocation can fail while
3778 * the mask is being updated. If a page allocation is about to fail,
3779 * check if the cpuset changed during allocation and if so, retry.
3781 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3782 alloc_mask = gfp_mask;
3787 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page) {
3788 if (unlikely(memcg_kmem_charge(page, gfp_mask, order))) {
3789 __free_pages(page, order);
3792 __SetPageKmemcg(page);
3795 if (kmemcheck_enabled && page)
3796 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3798 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3802 EXPORT_SYMBOL(__alloc_pages_nodemask);
3805 * Common helper functions.
3807 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3812 * __get_free_pages() returns a 32-bit address, which cannot represent
3815 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3817 page = alloc_pages(gfp_mask, order);
3820 return (unsigned long) page_address(page);
3822 EXPORT_SYMBOL(__get_free_pages);
3824 unsigned long get_zeroed_page(gfp_t gfp_mask)
3826 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3828 EXPORT_SYMBOL(get_zeroed_page);
3830 void __free_pages(struct page *page, unsigned int order)
3832 if (put_page_testzero(page)) {
3834 free_hot_cold_page(page, false);
3836 __free_pages_ok(page, order);
3840 EXPORT_SYMBOL(__free_pages);
3842 void free_pages(unsigned long addr, unsigned int order)
3845 VM_BUG_ON(!virt_addr_valid((void *)addr));
3846 __free_pages(virt_to_page((void *)addr), order);
3850 EXPORT_SYMBOL(free_pages);
3854 * An arbitrary-length arbitrary-offset area of memory which resides
3855 * within a 0 or higher order page. Multiple fragments within that page
3856 * are individually refcounted, in the page's reference counter.
3858 * The page_frag functions below provide a simple allocation framework for
3859 * page fragments. This is used by the network stack and network device
3860 * drivers to provide a backing region of memory for use as either an
3861 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3863 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3866 struct page *page = NULL;
3867 gfp_t gfp = gfp_mask;
3869 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3870 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3872 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3873 PAGE_FRAG_CACHE_MAX_ORDER);
3874 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3876 if (unlikely(!page))
3877 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3879 nc->va = page ? page_address(page) : NULL;
3884 void *__alloc_page_frag(struct page_frag_cache *nc,
3885 unsigned int fragsz, gfp_t gfp_mask)
3887 unsigned int size = PAGE_SIZE;
3891 if (unlikely(!nc->va)) {
3893 page = __page_frag_refill(nc, gfp_mask);
3897 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3898 /* if size can vary use size else just use PAGE_SIZE */
3901 /* Even if we own the page, we do not use atomic_set().
3902 * This would break get_page_unless_zero() users.
3904 page_ref_add(page, size - 1);
3906 /* reset page count bias and offset to start of new frag */
3907 nc->pfmemalloc = page_is_pfmemalloc(page);
3908 nc->pagecnt_bias = size;
3912 offset = nc->offset - fragsz;
3913 if (unlikely(offset < 0)) {
3914 page = virt_to_page(nc->va);
3916 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3919 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3920 /* if size can vary use size else just use PAGE_SIZE */
3923 /* OK, page count is 0, we can safely set it */
3924 set_page_count(page, size);
3926 /* reset page count bias and offset to start of new frag */
3927 nc->pagecnt_bias = size;
3928 offset = size - fragsz;
3932 nc->offset = offset;
3934 return nc->va + offset;
3936 EXPORT_SYMBOL(__alloc_page_frag);
3939 * Frees a page fragment allocated out of either a compound or order 0 page.
3941 void __free_page_frag(void *addr)
3943 struct page *page = virt_to_head_page(addr);
3945 if (unlikely(put_page_testzero(page)))
3946 __free_pages_ok(page, compound_order(page));
3948 EXPORT_SYMBOL(__free_page_frag);
3950 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3954 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3955 unsigned long used = addr + PAGE_ALIGN(size);
3957 split_page(virt_to_page((void *)addr), order);
3958 while (used < alloc_end) {
3963 return (void *)addr;
3967 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3968 * @size: the number of bytes to allocate
3969 * @gfp_mask: GFP flags for the allocation
3971 * This function is similar to alloc_pages(), except that it allocates the
3972 * minimum number of pages to satisfy the request. alloc_pages() can only
3973 * allocate memory in power-of-two pages.
3975 * This function is also limited by MAX_ORDER.
3977 * Memory allocated by this function must be released by free_pages_exact().
3979 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3981 unsigned int order = get_order(size);
3984 addr = __get_free_pages(gfp_mask, order);
3985 return make_alloc_exact(addr, order, size);
3987 EXPORT_SYMBOL(alloc_pages_exact);
3990 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3992 * @nid: the preferred node ID where memory should be allocated
3993 * @size: the number of bytes to allocate
3994 * @gfp_mask: GFP flags for the allocation
3996 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3999 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4001 unsigned int order = get_order(size);
4002 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4005 return make_alloc_exact((unsigned long)page_address(p), order, size);
4009 * free_pages_exact - release memory allocated via alloc_pages_exact()
4010 * @virt: the value returned by alloc_pages_exact.
4011 * @size: size of allocation, same value as passed to alloc_pages_exact().
4013 * Release the memory allocated by a previous call to alloc_pages_exact.
4015 void free_pages_exact(void *virt, size_t size)
4017 unsigned long addr = (unsigned long)virt;
4018 unsigned long end = addr + PAGE_ALIGN(size);
4020 while (addr < end) {
4025 EXPORT_SYMBOL(free_pages_exact);
4028 * nr_free_zone_pages - count number of pages beyond high watermark
4029 * @offset: The zone index of the highest zone
4031 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4032 * high watermark within all zones at or below a given zone index. For each
4033 * zone, the number of pages is calculated as:
4034 * managed_pages - high_pages
4036 static unsigned long nr_free_zone_pages(int offset)
4041 /* Just pick one node, since fallback list is circular */
4042 unsigned long sum = 0;
4044 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4046 for_each_zone_zonelist(zone, z, zonelist, offset) {
4047 unsigned long size = zone->managed_pages;
4048 unsigned long high = high_wmark_pages(zone);
4057 * nr_free_buffer_pages - count number of pages beyond high watermark
4059 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4060 * watermark within ZONE_DMA and ZONE_NORMAL.
4062 unsigned long nr_free_buffer_pages(void)
4064 return nr_free_zone_pages(gfp_zone(GFP_USER));
4066 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4069 * nr_free_pagecache_pages - count number of pages beyond high watermark
4071 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4072 * high watermark within all zones.
4074 unsigned long nr_free_pagecache_pages(void)
4076 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4079 static inline void show_node(struct zone *zone)
4081 if (IS_ENABLED(CONFIG_NUMA))
4082 printk("Node %d ", zone_to_nid(zone));
4085 long si_mem_available(void)
4088 unsigned long pagecache;
4089 unsigned long wmark_low = 0;
4090 unsigned long pages[NR_LRU_LISTS];
4094 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4095 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4098 wmark_low += zone->watermark[WMARK_LOW];
4101 * Estimate the amount of memory available for userspace allocations,
4102 * without causing swapping.
4104 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4107 * Not all the page cache can be freed, otherwise the system will
4108 * start swapping. Assume at least half of the page cache, or the
4109 * low watermark worth of cache, needs to stay.
4111 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4112 pagecache -= min(pagecache / 2, wmark_low);
4113 available += pagecache;
4116 * Part of the reclaimable slab consists of items that are in use,
4117 * and cannot be freed. Cap this estimate at the low watermark.
4119 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4120 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4126 EXPORT_SYMBOL_GPL(si_mem_available);
4128 void si_meminfo(struct sysinfo *val)
4130 val->totalram = totalram_pages;
4131 val->sharedram = global_node_page_state(NR_SHMEM);
4132 val->freeram = global_page_state(NR_FREE_PAGES);
4133 val->bufferram = nr_blockdev_pages();
4134 val->totalhigh = totalhigh_pages;
4135 val->freehigh = nr_free_highpages();
4136 val->mem_unit = PAGE_SIZE;
4139 EXPORT_SYMBOL(si_meminfo);
4142 void si_meminfo_node(struct sysinfo *val, int nid)
4144 int zone_type; /* needs to be signed */
4145 unsigned long managed_pages = 0;
4146 unsigned long managed_highpages = 0;
4147 unsigned long free_highpages = 0;
4148 pg_data_t *pgdat = NODE_DATA(nid);
4150 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4151 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4152 val->totalram = managed_pages;
4153 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4154 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4155 #ifdef CONFIG_HIGHMEM
4156 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4157 struct zone *zone = &pgdat->node_zones[zone_type];
4159 if (is_highmem(zone)) {
4160 managed_highpages += zone->managed_pages;
4161 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4164 val->totalhigh = managed_highpages;
4165 val->freehigh = free_highpages;
4167 val->totalhigh = managed_highpages;
4168 val->freehigh = free_highpages;
4170 val->mem_unit = PAGE_SIZE;
4175 * Determine whether the node should be displayed or not, depending on whether
4176 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4178 bool skip_free_areas_node(unsigned int flags, int nid)
4181 unsigned int cpuset_mems_cookie;
4183 if (!(flags & SHOW_MEM_FILTER_NODES))
4187 cpuset_mems_cookie = read_mems_allowed_begin();
4188 ret = !node_isset(nid, cpuset_current_mems_allowed);
4189 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4194 #define K(x) ((x) << (PAGE_SHIFT-10))
4196 static void show_migration_types(unsigned char type)
4198 static const char types[MIGRATE_TYPES] = {
4199 [MIGRATE_UNMOVABLE] = 'U',
4200 [MIGRATE_MOVABLE] = 'M',
4201 [MIGRATE_RECLAIMABLE] = 'E',
4202 [MIGRATE_HIGHATOMIC] = 'H',
4204 [MIGRATE_CMA] = 'C',
4206 #ifdef CONFIG_MEMORY_ISOLATION
4207 [MIGRATE_ISOLATE] = 'I',
4210 char tmp[MIGRATE_TYPES + 1];
4214 for (i = 0; i < MIGRATE_TYPES; i++) {
4215 if (type & (1 << i))
4220 printk("(%s) ", tmp);
4224 * Show free area list (used inside shift_scroll-lock stuff)
4225 * We also calculate the percentage fragmentation. We do this by counting the
4226 * memory on each free list with the exception of the first item on the list.
4229 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4232 void show_free_areas(unsigned int filter)
4234 unsigned long free_pcp = 0;
4239 for_each_populated_zone(zone) {
4240 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4243 for_each_online_cpu(cpu)
4244 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4247 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4248 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4249 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4250 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4251 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4252 " free:%lu free_pcp:%lu free_cma:%lu\n",
4253 global_node_page_state(NR_ACTIVE_ANON),
4254 global_node_page_state(NR_INACTIVE_ANON),
4255 global_node_page_state(NR_ISOLATED_ANON),
4256 global_node_page_state(NR_ACTIVE_FILE),
4257 global_node_page_state(NR_INACTIVE_FILE),
4258 global_node_page_state(NR_ISOLATED_FILE),
4259 global_node_page_state(NR_UNEVICTABLE),
4260 global_node_page_state(NR_FILE_DIRTY),
4261 global_node_page_state(NR_WRITEBACK),
4262 global_node_page_state(NR_UNSTABLE_NFS),
4263 global_page_state(NR_SLAB_RECLAIMABLE),
4264 global_page_state(NR_SLAB_UNRECLAIMABLE),
4265 global_node_page_state(NR_FILE_MAPPED),
4266 global_node_page_state(NR_SHMEM),
4267 global_page_state(NR_PAGETABLE),
4268 global_page_state(NR_BOUNCE),
4269 global_page_state(NR_FREE_PAGES),
4271 global_page_state(NR_FREE_CMA_PAGES));
4273 for_each_online_pgdat(pgdat) {
4275 " active_anon:%lukB"
4276 " inactive_anon:%lukB"
4277 " active_file:%lukB"
4278 " inactive_file:%lukB"
4279 " unevictable:%lukB"
4280 " isolated(anon):%lukB"
4281 " isolated(file):%lukB"
4286 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4288 " shmem_pmdmapped: %lukB"
4291 " writeback_tmp:%lukB"
4293 " pages_scanned:%lu"
4294 " all_unreclaimable? %s"
4297 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4298 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4299 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4300 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4301 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4302 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4303 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4304 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4305 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4306 K(node_page_state(pgdat, NR_WRITEBACK)),
4307 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4308 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4309 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4311 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4313 K(node_page_state(pgdat, NR_SHMEM)),
4314 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4315 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4316 node_page_state(pgdat, NR_PAGES_SCANNED),
4317 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4320 for_each_populated_zone(zone) {
4323 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4327 for_each_online_cpu(cpu)
4328 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4336 " active_anon:%lukB"
4337 " inactive_anon:%lukB"
4338 " active_file:%lukB"
4339 " inactive_file:%lukB"
4340 " unevictable:%lukB"
4341 " writepending:%lukB"
4345 " slab_reclaimable:%lukB"
4346 " slab_unreclaimable:%lukB"
4347 " kernel_stack:%lukB"
4355 K(zone_page_state(zone, NR_FREE_PAGES)),
4356 K(min_wmark_pages(zone)),
4357 K(low_wmark_pages(zone)),
4358 K(high_wmark_pages(zone)),
4359 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4360 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4361 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4362 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4363 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4364 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4365 K(zone->present_pages),
4366 K(zone->managed_pages),
4367 K(zone_page_state(zone, NR_MLOCK)),
4368 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4369 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4370 zone_page_state(zone, NR_KERNEL_STACK_KB),
4371 K(zone_page_state(zone, NR_PAGETABLE)),
4372 K(zone_page_state(zone, NR_BOUNCE)),
4374 K(this_cpu_read(zone->pageset->pcp.count)),
4375 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4376 printk("lowmem_reserve[]:");
4377 for (i = 0; i < MAX_NR_ZONES; i++)
4378 printk(" %ld", zone->lowmem_reserve[i]);
4382 for_each_populated_zone(zone) {
4384 unsigned long nr[MAX_ORDER], flags, total = 0;
4385 unsigned char types[MAX_ORDER];
4387 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4390 printk("%s: ", zone->name);
4392 spin_lock_irqsave(&zone->lock, flags);
4393 for (order = 0; order < MAX_ORDER; order++) {
4394 struct free_area *area = &zone->free_area[order];
4397 nr[order] = area->nr_free;
4398 total += nr[order] << order;
4401 for (type = 0; type < MIGRATE_TYPES; type++) {
4402 if (!list_empty(&area->free_list[type]))
4403 types[order] |= 1 << type;
4406 spin_unlock_irqrestore(&zone->lock, flags);
4407 for (order = 0; order < MAX_ORDER; order++) {
4408 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4410 show_migration_types(types[order]);
4412 printk("= %lukB\n", K(total));
4415 hugetlb_show_meminfo();
4417 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4419 show_swap_cache_info();
4422 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4424 zoneref->zone = zone;
4425 zoneref->zone_idx = zone_idx(zone);
4429 * Builds allocation fallback zone lists.
4431 * Add all populated zones of a node to the zonelist.
4433 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4437 enum zone_type zone_type = MAX_NR_ZONES;
4441 zone = pgdat->node_zones + zone_type;
4442 if (populated_zone(zone)) {
4443 zoneref_set_zone(zone,
4444 &zonelist->_zonerefs[nr_zones++]);
4445 check_highest_zone(zone_type);
4447 } while (zone_type);
4455 * 0 = automatic detection of better ordering.
4456 * 1 = order by ([node] distance, -zonetype)
4457 * 2 = order by (-zonetype, [node] distance)
4459 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4460 * the same zonelist. So only NUMA can configure this param.
4462 #define ZONELIST_ORDER_DEFAULT 0
4463 #define ZONELIST_ORDER_NODE 1
4464 #define ZONELIST_ORDER_ZONE 2
4466 /* zonelist order in the kernel.
4467 * set_zonelist_order() will set this to NODE or ZONE.
4469 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4470 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4474 /* The value user specified ....changed by config */
4475 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4476 /* string for sysctl */
4477 #define NUMA_ZONELIST_ORDER_LEN 16
4478 char numa_zonelist_order[16] = "default";
4481 * interface for configure zonelist ordering.
4482 * command line option "numa_zonelist_order"
4483 * = "[dD]efault - default, automatic configuration.
4484 * = "[nN]ode - order by node locality, then by zone within node
4485 * = "[zZ]one - order by zone, then by locality within zone
4488 static int __parse_numa_zonelist_order(char *s)
4490 if (*s == 'd' || *s == 'D') {
4491 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4492 } else if (*s == 'n' || *s == 'N') {
4493 user_zonelist_order = ZONELIST_ORDER_NODE;
4494 } else if (*s == 'z' || *s == 'Z') {
4495 user_zonelist_order = ZONELIST_ORDER_ZONE;
4497 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4503 static __init int setup_numa_zonelist_order(char *s)
4510 ret = __parse_numa_zonelist_order(s);
4512 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4516 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4519 * sysctl handler for numa_zonelist_order
4521 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4522 void __user *buffer, size_t *length,
4525 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4527 static DEFINE_MUTEX(zl_order_mutex);
4529 mutex_lock(&zl_order_mutex);
4531 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4535 strcpy(saved_string, (char *)table->data);
4537 ret = proc_dostring(table, write, buffer, length, ppos);
4541 int oldval = user_zonelist_order;
4543 ret = __parse_numa_zonelist_order((char *)table->data);
4546 * bogus value. restore saved string
4548 strncpy((char *)table->data, saved_string,
4549 NUMA_ZONELIST_ORDER_LEN);
4550 user_zonelist_order = oldval;
4551 } else if (oldval != user_zonelist_order) {
4552 mutex_lock(&zonelists_mutex);
4553 build_all_zonelists(NULL, NULL);
4554 mutex_unlock(&zonelists_mutex);
4558 mutex_unlock(&zl_order_mutex);
4563 #define MAX_NODE_LOAD (nr_online_nodes)
4564 static int node_load[MAX_NUMNODES];
4567 * find_next_best_node - find the next node that should appear in a given node's fallback list
4568 * @node: node whose fallback list we're appending
4569 * @used_node_mask: nodemask_t of already used nodes
4571 * We use a number of factors to determine which is the next node that should
4572 * appear on a given node's fallback list. The node should not have appeared
4573 * already in @node's fallback list, and it should be the next closest node
4574 * according to the distance array (which contains arbitrary distance values
4575 * from each node to each node in the system), and should also prefer nodes
4576 * with no CPUs, since presumably they'll have very little allocation pressure
4577 * on them otherwise.
4578 * It returns -1 if no node is found.
4580 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4583 int min_val = INT_MAX;
4584 int best_node = NUMA_NO_NODE;
4585 const struct cpumask *tmp = cpumask_of_node(0);
4587 /* Use the local node if we haven't already */
4588 if (!node_isset(node, *used_node_mask)) {
4589 node_set(node, *used_node_mask);
4593 for_each_node_state(n, N_MEMORY) {
4595 /* Don't want a node to appear more than once */
4596 if (node_isset(n, *used_node_mask))
4599 /* Use the distance array to find the distance */
4600 val = node_distance(node, n);
4602 /* Penalize nodes under us ("prefer the next node") */
4605 /* Give preference to headless and unused nodes */
4606 tmp = cpumask_of_node(n);
4607 if (!cpumask_empty(tmp))
4608 val += PENALTY_FOR_NODE_WITH_CPUS;
4610 /* Slight preference for less loaded node */
4611 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4612 val += node_load[n];
4614 if (val < min_val) {
4621 node_set(best_node, *used_node_mask);
4628 * Build zonelists ordered by node and zones within node.
4629 * This results in maximum locality--normal zone overflows into local
4630 * DMA zone, if any--but risks exhausting DMA zone.
4632 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4635 struct zonelist *zonelist;
4637 zonelist = &pgdat->node_zonelists[0];
4638 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4640 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4641 zonelist->_zonerefs[j].zone = NULL;
4642 zonelist->_zonerefs[j].zone_idx = 0;
4646 * Build gfp_thisnode zonelists
4648 static void build_thisnode_zonelists(pg_data_t *pgdat)
4651 struct zonelist *zonelist;
4653 zonelist = &pgdat->node_zonelists[1];
4654 j = build_zonelists_node(pgdat, zonelist, 0);
4655 zonelist->_zonerefs[j].zone = NULL;
4656 zonelist->_zonerefs[j].zone_idx = 0;
4660 * Build zonelists ordered by zone and nodes within zones.
4661 * This results in conserving DMA zone[s] until all Normal memory is
4662 * exhausted, but results in overflowing to remote node while memory
4663 * may still exist in local DMA zone.
4665 static int node_order[MAX_NUMNODES];
4667 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4670 int zone_type; /* needs to be signed */
4672 struct zonelist *zonelist;
4674 zonelist = &pgdat->node_zonelists[0];
4676 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4677 for (j = 0; j < nr_nodes; j++) {
4678 node = node_order[j];
4679 z = &NODE_DATA(node)->node_zones[zone_type];
4680 if (populated_zone(z)) {
4682 &zonelist->_zonerefs[pos++]);
4683 check_highest_zone(zone_type);
4687 zonelist->_zonerefs[pos].zone = NULL;
4688 zonelist->_zonerefs[pos].zone_idx = 0;
4691 #if defined(CONFIG_64BIT)
4693 * Devices that require DMA32/DMA are relatively rare and do not justify a
4694 * penalty to every machine in case the specialised case applies. Default
4695 * to Node-ordering on 64-bit NUMA machines
4697 static int default_zonelist_order(void)
4699 return ZONELIST_ORDER_NODE;
4703 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4704 * by the kernel. If processes running on node 0 deplete the low memory zone
4705 * then reclaim will occur more frequency increasing stalls and potentially
4706 * be easier to OOM if a large percentage of the zone is under writeback or
4707 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4708 * Hence, default to zone ordering on 32-bit.
4710 static int default_zonelist_order(void)
4712 return ZONELIST_ORDER_ZONE;
4714 #endif /* CONFIG_64BIT */
4716 static void set_zonelist_order(void)
4718 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4719 current_zonelist_order = default_zonelist_order();
4721 current_zonelist_order = user_zonelist_order;
4724 static void build_zonelists(pg_data_t *pgdat)
4727 nodemask_t used_mask;
4728 int local_node, prev_node;
4729 struct zonelist *zonelist;
4730 unsigned int order = current_zonelist_order;
4732 /* initialize zonelists */
4733 for (i = 0; i < MAX_ZONELISTS; i++) {
4734 zonelist = pgdat->node_zonelists + i;
4735 zonelist->_zonerefs[0].zone = NULL;
4736 zonelist->_zonerefs[0].zone_idx = 0;
4739 /* NUMA-aware ordering of nodes */
4740 local_node = pgdat->node_id;
4741 load = nr_online_nodes;
4742 prev_node = local_node;
4743 nodes_clear(used_mask);
4745 memset(node_order, 0, sizeof(node_order));
4748 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4750 * We don't want to pressure a particular node.
4751 * So adding penalty to the first node in same
4752 * distance group to make it round-robin.
4754 if (node_distance(local_node, node) !=
4755 node_distance(local_node, prev_node))
4756 node_load[node] = load;
4760 if (order == ZONELIST_ORDER_NODE)
4761 build_zonelists_in_node_order(pgdat, node);
4763 node_order[i++] = node; /* remember order */
4766 if (order == ZONELIST_ORDER_ZONE) {
4767 /* calculate node order -- i.e., DMA last! */
4768 build_zonelists_in_zone_order(pgdat, i);
4771 build_thisnode_zonelists(pgdat);
4774 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4776 * Return node id of node used for "local" allocations.
4777 * I.e., first node id of first zone in arg node's generic zonelist.
4778 * Used for initializing percpu 'numa_mem', which is used primarily
4779 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4781 int local_memory_node(int node)
4785 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4786 gfp_zone(GFP_KERNEL),
4788 return z->zone->node;
4792 #else /* CONFIG_NUMA */
4794 static void set_zonelist_order(void)
4796 current_zonelist_order = ZONELIST_ORDER_ZONE;
4799 static void build_zonelists(pg_data_t *pgdat)
4801 int node, local_node;
4803 struct zonelist *zonelist;
4805 local_node = pgdat->node_id;
4807 zonelist = &pgdat->node_zonelists[0];
4808 j = build_zonelists_node(pgdat, zonelist, 0);
4811 * Now we build the zonelist so that it contains the zones
4812 * of all the other nodes.
4813 * We don't want to pressure a particular node, so when
4814 * building the zones for node N, we make sure that the
4815 * zones coming right after the local ones are those from
4816 * node N+1 (modulo N)
4818 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4819 if (!node_online(node))
4821 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4823 for (node = 0; node < local_node; node++) {
4824 if (!node_online(node))
4826 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4829 zonelist->_zonerefs[j].zone = NULL;
4830 zonelist->_zonerefs[j].zone_idx = 0;
4833 #endif /* CONFIG_NUMA */
4836 * Boot pageset table. One per cpu which is going to be used for all
4837 * zones and all nodes. The parameters will be set in such a way
4838 * that an item put on a list will immediately be handed over to
4839 * the buddy list. This is safe since pageset manipulation is done
4840 * with interrupts disabled.
4842 * The boot_pagesets must be kept even after bootup is complete for
4843 * unused processors and/or zones. They do play a role for bootstrapping
4844 * hotplugged processors.
4846 * zoneinfo_show() and maybe other functions do
4847 * not check if the processor is online before following the pageset pointer.
4848 * Other parts of the kernel may not check if the zone is available.
4850 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4851 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4852 static void setup_zone_pageset(struct zone *zone);
4855 * Global mutex to protect against size modification of zonelists
4856 * as well as to serialize pageset setup for the new populated zone.
4858 DEFINE_MUTEX(zonelists_mutex);
4860 /* return values int ....just for stop_machine() */
4861 static int __build_all_zonelists(void *data)
4865 pg_data_t *self = data;
4868 memset(node_load, 0, sizeof(node_load));
4871 if (self && !node_online(self->node_id)) {
4872 build_zonelists(self);
4875 for_each_online_node(nid) {
4876 pg_data_t *pgdat = NODE_DATA(nid);
4878 build_zonelists(pgdat);
4882 * Initialize the boot_pagesets that are going to be used
4883 * for bootstrapping processors. The real pagesets for
4884 * each zone will be allocated later when the per cpu
4885 * allocator is available.
4887 * boot_pagesets are used also for bootstrapping offline
4888 * cpus if the system is already booted because the pagesets
4889 * are needed to initialize allocators on a specific cpu too.
4890 * F.e. the percpu allocator needs the page allocator which
4891 * needs the percpu allocator in order to allocate its pagesets
4892 * (a chicken-egg dilemma).
4894 for_each_possible_cpu(cpu) {
4895 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4897 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4899 * We now know the "local memory node" for each node--
4900 * i.e., the node of the first zone in the generic zonelist.
4901 * Set up numa_mem percpu variable for on-line cpus. During
4902 * boot, only the boot cpu should be on-line; we'll init the
4903 * secondary cpus' numa_mem as they come on-line. During
4904 * node/memory hotplug, we'll fixup all on-line cpus.
4906 if (cpu_online(cpu))
4907 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4914 static noinline void __init
4915 build_all_zonelists_init(void)
4917 __build_all_zonelists(NULL);
4918 mminit_verify_zonelist();
4919 cpuset_init_current_mems_allowed();
4923 * Called with zonelists_mutex held always
4924 * unless system_state == SYSTEM_BOOTING.
4926 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4927 * [we're only called with non-NULL zone through __meminit paths] and
4928 * (2) call of __init annotated helper build_all_zonelists_init
4929 * [protected by SYSTEM_BOOTING].
4931 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4933 set_zonelist_order();
4935 if (system_state == SYSTEM_BOOTING) {
4936 build_all_zonelists_init();
4938 #ifdef CONFIG_MEMORY_HOTPLUG
4940 setup_zone_pageset(zone);
4942 /* we have to stop all cpus to guarantee there is no user
4944 stop_machine(__build_all_zonelists, pgdat, NULL);
4945 /* cpuset refresh routine should be here */
4947 vm_total_pages = nr_free_pagecache_pages();
4949 * Disable grouping by mobility if the number of pages in the
4950 * system is too low to allow the mechanism to work. It would be
4951 * more accurate, but expensive to check per-zone. This check is
4952 * made on memory-hotadd so a system can start with mobility
4953 * disabled and enable it later
4955 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4956 page_group_by_mobility_disabled = 1;
4958 page_group_by_mobility_disabled = 0;
4960 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4962 zonelist_order_name[current_zonelist_order],
4963 page_group_by_mobility_disabled ? "off" : "on",
4966 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4971 * Helper functions to size the waitqueue hash table.
4972 * Essentially these want to choose hash table sizes sufficiently
4973 * large so that collisions trying to wait on pages are rare.
4974 * But in fact, the number of active page waitqueues on typical
4975 * systems is ridiculously low, less than 200. So this is even
4976 * conservative, even though it seems large.
4978 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4979 * waitqueues, i.e. the size of the waitq table given the number of pages.
4981 #define PAGES_PER_WAITQUEUE 256
4983 #ifndef CONFIG_MEMORY_HOTPLUG
4984 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4986 unsigned long size = 1;
4988 pages /= PAGES_PER_WAITQUEUE;
4990 while (size < pages)
4994 * Once we have dozens or even hundreds of threads sleeping
4995 * on IO we've got bigger problems than wait queue collision.
4996 * Limit the size of the wait table to a reasonable size.
4998 size = min(size, 4096UL);
5000 return max(size, 4UL);
5004 * A zone's size might be changed by hot-add, so it is not possible to determine
5005 * a suitable size for its wait_table. So we use the maximum size now.
5007 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
5009 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
5010 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5011 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5013 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5014 * or more by the traditional way. (See above). It equals:
5016 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5017 * ia64(16K page size) : = ( 8G + 4M)byte.
5018 * powerpc (64K page size) : = (32G +16M)byte.
5020 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5027 * This is an integer logarithm so that shifts can be used later
5028 * to extract the more random high bits from the multiplicative
5029 * hash function before the remainder is taken.
5031 static inline unsigned long wait_table_bits(unsigned long size)
5037 * Initially all pages are reserved - free ones are freed
5038 * up by free_all_bootmem() once the early boot process is
5039 * done. Non-atomic initialization, single-pass.
5041 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5042 unsigned long start_pfn, enum memmap_context context)
5044 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5045 unsigned long end_pfn = start_pfn + size;
5046 pg_data_t *pgdat = NODE_DATA(nid);
5048 unsigned long nr_initialised = 0;
5049 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5050 struct memblock_region *r = NULL, *tmp;
5053 if (highest_memmap_pfn < end_pfn - 1)
5054 highest_memmap_pfn = end_pfn - 1;
5057 * Honor reservation requested by the driver for this ZONE_DEVICE
5060 if (altmap && start_pfn == altmap->base_pfn)
5061 start_pfn += altmap->reserve;
5063 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5065 * There can be holes in boot-time mem_map[]s handed to this
5066 * function. They do not exist on hotplugged memory.
5068 if (context != MEMMAP_EARLY)
5071 if (!early_pfn_valid(pfn))
5073 if (!early_pfn_in_nid(pfn, nid))
5075 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5078 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5080 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5081 * from zone_movable_pfn[nid] to end of each node should be
5082 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5084 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5085 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5089 * Check given memblock attribute by firmware which can affect
5090 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5091 * mirrored, it's an overlapped memmap init. skip it.
5093 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5094 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5095 for_each_memblock(memory, tmp)
5096 if (pfn < memblock_region_memory_end_pfn(tmp))
5100 if (pfn >= memblock_region_memory_base_pfn(r) &&
5101 memblock_is_mirror(r)) {
5102 /* already initialized as NORMAL */
5103 pfn = memblock_region_memory_end_pfn(r);
5111 * Mark the block movable so that blocks are reserved for
5112 * movable at startup. This will force kernel allocations
5113 * to reserve their blocks rather than leaking throughout
5114 * the address space during boot when many long-lived
5115 * kernel allocations are made.
5117 * bitmap is created for zone's valid pfn range. but memmap
5118 * can be created for invalid pages (for alignment)
5119 * check here not to call set_pageblock_migratetype() against
5122 if (!(pfn & (pageblock_nr_pages - 1))) {
5123 struct page *page = pfn_to_page(pfn);
5125 __init_single_page(page, pfn, zone, nid);
5126 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5128 __init_single_pfn(pfn, zone, nid);
5133 static void __meminit zone_init_free_lists(struct zone *zone)
5135 unsigned int order, t;
5136 for_each_migratetype_order(order, t) {
5137 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5138 zone->free_area[order].nr_free = 0;
5142 #ifndef __HAVE_ARCH_MEMMAP_INIT
5143 #define memmap_init(size, nid, zone, start_pfn) \
5144 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5147 static int zone_batchsize(struct zone *zone)
5153 * The per-cpu-pages pools are set to around 1000th of the
5154 * size of the zone. But no more than 1/2 of a meg.
5156 * OK, so we don't know how big the cache is. So guess.
5158 batch = zone->managed_pages / 1024;
5159 if (batch * PAGE_SIZE > 512 * 1024)
5160 batch = (512 * 1024) / PAGE_SIZE;
5161 batch /= 4; /* We effectively *= 4 below */
5166 * Clamp the batch to a 2^n - 1 value. Having a power
5167 * of 2 value was found to be more likely to have
5168 * suboptimal cache aliasing properties in some cases.
5170 * For example if 2 tasks are alternately allocating
5171 * batches of pages, one task can end up with a lot
5172 * of pages of one half of the possible page colors
5173 * and the other with pages of the other colors.
5175 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5180 /* The deferral and batching of frees should be suppressed under NOMMU
5183 * The problem is that NOMMU needs to be able to allocate large chunks
5184 * of contiguous memory as there's no hardware page translation to
5185 * assemble apparent contiguous memory from discontiguous pages.
5187 * Queueing large contiguous runs of pages for batching, however,
5188 * causes the pages to actually be freed in smaller chunks. As there
5189 * can be a significant delay between the individual batches being
5190 * recycled, this leads to the once large chunks of space being
5191 * fragmented and becoming unavailable for high-order allocations.
5198 * pcp->high and pcp->batch values are related and dependent on one another:
5199 * ->batch must never be higher then ->high.
5200 * The following function updates them in a safe manner without read side
5203 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5204 * those fields changing asynchronously (acording the the above rule).
5206 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5207 * outside of boot time (or some other assurance that no concurrent updaters
5210 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5211 unsigned long batch)
5213 /* start with a fail safe value for batch */
5217 /* Update high, then batch, in order */
5224 /* a companion to pageset_set_high() */
5225 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5227 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5230 static void pageset_init(struct per_cpu_pageset *p)
5232 struct per_cpu_pages *pcp;
5235 memset(p, 0, sizeof(*p));
5239 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5240 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5243 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5246 pageset_set_batch(p, batch);
5250 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5251 * to the value high for the pageset p.
5253 static void pageset_set_high(struct per_cpu_pageset *p,
5256 unsigned long batch = max(1UL, high / 4);
5257 if ((high / 4) > (PAGE_SHIFT * 8))
5258 batch = PAGE_SHIFT * 8;
5260 pageset_update(&p->pcp, high, batch);
5263 static void pageset_set_high_and_batch(struct zone *zone,
5264 struct per_cpu_pageset *pcp)
5266 if (percpu_pagelist_fraction)
5267 pageset_set_high(pcp,
5268 (zone->managed_pages /
5269 percpu_pagelist_fraction));
5271 pageset_set_batch(pcp, zone_batchsize(zone));
5274 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5276 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5279 pageset_set_high_and_batch(zone, pcp);
5282 static void __meminit setup_zone_pageset(struct zone *zone)
5285 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5286 for_each_possible_cpu(cpu)
5287 zone_pageset_init(zone, cpu);
5289 if (!zone->zone_pgdat->per_cpu_nodestats) {
5290 zone->zone_pgdat->per_cpu_nodestats =
5291 alloc_percpu(struct per_cpu_nodestat);
5296 * Allocate per cpu pagesets and initialize them.
5297 * Before this call only boot pagesets were available.
5299 void __init setup_per_cpu_pageset(void)
5303 for_each_populated_zone(zone)
5304 setup_zone_pageset(zone);
5307 static noinline __init_refok
5308 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5314 * The per-page waitqueue mechanism uses hashed waitqueues
5317 zone->wait_table_hash_nr_entries =
5318 wait_table_hash_nr_entries(zone_size_pages);
5319 zone->wait_table_bits =
5320 wait_table_bits(zone->wait_table_hash_nr_entries);
5321 alloc_size = zone->wait_table_hash_nr_entries
5322 * sizeof(wait_queue_head_t);
5324 if (!slab_is_available()) {
5325 zone->wait_table = (wait_queue_head_t *)
5326 memblock_virt_alloc_node_nopanic(
5327 alloc_size, zone->zone_pgdat->node_id);
5330 * This case means that a zone whose size was 0 gets new memory
5331 * via memory hot-add.
5332 * But it may be the case that a new node was hot-added. In
5333 * this case vmalloc() will not be able to use this new node's
5334 * memory - this wait_table must be initialized to use this new
5335 * node itself as well.
5336 * To use this new node's memory, further consideration will be
5339 zone->wait_table = vmalloc(alloc_size);
5341 if (!zone->wait_table)
5344 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5345 init_waitqueue_head(zone->wait_table + i);
5350 static __meminit void zone_pcp_init(struct zone *zone)
5353 * per cpu subsystem is not up at this point. The following code
5354 * relies on the ability of the linker to provide the
5355 * offset of a (static) per cpu variable into the per cpu area.
5357 zone->pageset = &boot_pageset;
5359 if (populated_zone(zone))
5360 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5361 zone->name, zone->present_pages,
5362 zone_batchsize(zone));
5365 int __meminit init_currently_empty_zone(struct zone *zone,
5366 unsigned long zone_start_pfn,
5369 struct pglist_data *pgdat = zone->zone_pgdat;
5371 ret = zone_wait_table_init(zone, size);
5374 pgdat->nr_zones = zone_idx(zone) + 1;
5376 zone->zone_start_pfn = zone_start_pfn;
5378 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5379 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5381 (unsigned long)zone_idx(zone),
5382 zone_start_pfn, (zone_start_pfn + size));
5384 zone_init_free_lists(zone);
5389 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5390 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5393 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5395 int __meminit __early_pfn_to_nid(unsigned long pfn,
5396 struct mminit_pfnnid_cache *state)
5398 unsigned long start_pfn, end_pfn;
5401 if (state->last_start <= pfn && pfn < state->last_end)
5402 return state->last_nid;
5404 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5406 state->last_start = start_pfn;
5407 state->last_end = end_pfn;
5408 state->last_nid = nid;
5413 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5416 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5417 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5418 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5420 * If an architecture guarantees that all ranges registered contain no holes
5421 * and may be freed, this this function may be used instead of calling
5422 * memblock_free_early_nid() manually.
5424 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5426 unsigned long start_pfn, end_pfn;
5429 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5430 start_pfn = min(start_pfn, max_low_pfn);
5431 end_pfn = min(end_pfn, max_low_pfn);
5433 if (start_pfn < end_pfn)
5434 memblock_free_early_nid(PFN_PHYS(start_pfn),
5435 (end_pfn - start_pfn) << PAGE_SHIFT,
5441 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5442 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5444 * If an architecture guarantees that all ranges registered contain no holes and may
5445 * be freed, this function may be used instead of calling memory_present() manually.
5447 void __init sparse_memory_present_with_active_regions(int nid)
5449 unsigned long start_pfn, end_pfn;
5452 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5453 memory_present(this_nid, start_pfn, end_pfn);
5457 * get_pfn_range_for_nid - Return the start and end page frames for a node
5458 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5459 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5460 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5462 * It returns the start and end page frame of a node based on information
5463 * provided by memblock_set_node(). If called for a node
5464 * with no available memory, a warning is printed and the start and end
5467 void __meminit get_pfn_range_for_nid(unsigned int nid,
5468 unsigned long *start_pfn, unsigned long *end_pfn)
5470 unsigned long this_start_pfn, this_end_pfn;
5476 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5477 *start_pfn = min(*start_pfn, this_start_pfn);
5478 *end_pfn = max(*end_pfn, this_end_pfn);
5481 if (*start_pfn == -1UL)
5486 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5487 * assumption is made that zones within a node are ordered in monotonic
5488 * increasing memory addresses so that the "highest" populated zone is used
5490 static void __init find_usable_zone_for_movable(void)
5493 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5494 if (zone_index == ZONE_MOVABLE)
5497 if (arch_zone_highest_possible_pfn[zone_index] >
5498 arch_zone_lowest_possible_pfn[zone_index])
5502 VM_BUG_ON(zone_index == -1);
5503 movable_zone = zone_index;
5507 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5508 * because it is sized independent of architecture. Unlike the other zones,
5509 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5510 * in each node depending on the size of each node and how evenly kernelcore
5511 * is distributed. This helper function adjusts the zone ranges
5512 * provided by the architecture for a given node by using the end of the
5513 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5514 * zones within a node are in order of monotonic increases memory addresses
5516 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5517 unsigned long zone_type,
5518 unsigned long node_start_pfn,
5519 unsigned long node_end_pfn,
5520 unsigned long *zone_start_pfn,
5521 unsigned long *zone_end_pfn)
5523 /* Only adjust if ZONE_MOVABLE is on this node */
5524 if (zone_movable_pfn[nid]) {
5525 /* Size ZONE_MOVABLE */
5526 if (zone_type == ZONE_MOVABLE) {
5527 *zone_start_pfn = zone_movable_pfn[nid];
5528 *zone_end_pfn = min(node_end_pfn,
5529 arch_zone_highest_possible_pfn[movable_zone]);
5531 /* Check if this whole range is within ZONE_MOVABLE */
5532 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5533 *zone_start_pfn = *zone_end_pfn;
5538 * Return the number of pages a zone spans in a node, including holes
5539 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5541 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5542 unsigned long zone_type,
5543 unsigned long node_start_pfn,
5544 unsigned long node_end_pfn,
5545 unsigned long *zone_start_pfn,
5546 unsigned long *zone_end_pfn,
5547 unsigned long *ignored)
5549 /* When hotadd a new node from cpu_up(), the node should be empty */
5550 if (!node_start_pfn && !node_end_pfn)
5553 /* Get the start and end of the zone */
5554 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5555 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5556 adjust_zone_range_for_zone_movable(nid, zone_type,
5557 node_start_pfn, node_end_pfn,
5558 zone_start_pfn, zone_end_pfn);
5560 /* Check that this node has pages within the zone's required range */
5561 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5564 /* Move the zone boundaries inside the node if necessary */
5565 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5566 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5568 /* Return the spanned pages */
5569 return *zone_end_pfn - *zone_start_pfn;
5573 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5574 * then all holes in the requested range will be accounted for.
5576 unsigned long __meminit __absent_pages_in_range(int nid,
5577 unsigned long range_start_pfn,
5578 unsigned long range_end_pfn)
5580 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5581 unsigned long start_pfn, end_pfn;
5584 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5585 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5586 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5587 nr_absent -= end_pfn - start_pfn;
5593 * absent_pages_in_range - Return number of page frames in holes within a range
5594 * @start_pfn: The start PFN to start searching for holes
5595 * @end_pfn: The end PFN to stop searching for holes
5597 * It returns the number of pages frames in memory holes within a range.
5599 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5600 unsigned long end_pfn)
5602 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5605 /* Return the number of page frames in holes in a zone on a node */
5606 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5607 unsigned long zone_type,
5608 unsigned long node_start_pfn,
5609 unsigned long node_end_pfn,
5610 unsigned long *ignored)
5612 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5613 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5614 unsigned long zone_start_pfn, zone_end_pfn;
5615 unsigned long nr_absent;
5617 /* When hotadd a new node from cpu_up(), the node should be empty */
5618 if (!node_start_pfn && !node_end_pfn)
5621 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5622 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5624 adjust_zone_range_for_zone_movable(nid, zone_type,
5625 node_start_pfn, node_end_pfn,
5626 &zone_start_pfn, &zone_end_pfn);
5627 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5630 * ZONE_MOVABLE handling.
5631 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5634 if (zone_movable_pfn[nid]) {
5635 if (mirrored_kernelcore) {
5636 unsigned long start_pfn, end_pfn;
5637 struct memblock_region *r;
5639 for_each_memblock(memory, r) {
5640 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5641 zone_start_pfn, zone_end_pfn);
5642 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5643 zone_start_pfn, zone_end_pfn);
5645 if (zone_type == ZONE_MOVABLE &&
5646 memblock_is_mirror(r))
5647 nr_absent += end_pfn - start_pfn;
5649 if (zone_type == ZONE_NORMAL &&
5650 !memblock_is_mirror(r))
5651 nr_absent += end_pfn - start_pfn;
5654 if (zone_type == ZONE_NORMAL)
5655 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5662 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5663 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5664 unsigned long zone_type,
5665 unsigned long node_start_pfn,
5666 unsigned long node_end_pfn,
5667 unsigned long *zone_start_pfn,
5668 unsigned long *zone_end_pfn,
5669 unsigned long *zones_size)
5673 *zone_start_pfn = node_start_pfn;
5674 for (zone = 0; zone < zone_type; zone++)
5675 *zone_start_pfn += zones_size[zone];
5677 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5679 return zones_size[zone_type];
5682 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5683 unsigned long zone_type,
5684 unsigned long node_start_pfn,
5685 unsigned long node_end_pfn,
5686 unsigned long *zholes_size)
5691 return zholes_size[zone_type];
5694 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5696 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5697 unsigned long node_start_pfn,
5698 unsigned long node_end_pfn,
5699 unsigned long *zones_size,
5700 unsigned long *zholes_size)
5702 unsigned long realtotalpages = 0, totalpages = 0;
5705 for (i = 0; i < MAX_NR_ZONES; i++) {
5706 struct zone *zone = pgdat->node_zones + i;
5707 unsigned long zone_start_pfn, zone_end_pfn;
5708 unsigned long size, real_size;
5710 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5716 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5717 node_start_pfn, node_end_pfn,
5720 zone->zone_start_pfn = zone_start_pfn;
5722 zone->zone_start_pfn = 0;
5723 zone->spanned_pages = size;
5724 zone->present_pages = real_size;
5727 realtotalpages += real_size;
5730 pgdat->node_spanned_pages = totalpages;
5731 pgdat->node_present_pages = realtotalpages;
5732 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5736 #ifndef CONFIG_SPARSEMEM
5738 * Calculate the size of the zone->blockflags rounded to an unsigned long
5739 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5740 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5741 * round what is now in bits to nearest long in bits, then return it in
5744 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5746 unsigned long usemapsize;
5748 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5749 usemapsize = roundup(zonesize, pageblock_nr_pages);
5750 usemapsize = usemapsize >> pageblock_order;
5751 usemapsize *= NR_PAGEBLOCK_BITS;
5752 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5754 return usemapsize / 8;
5757 static void __init setup_usemap(struct pglist_data *pgdat,
5759 unsigned long zone_start_pfn,
5760 unsigned long zonesize)
5762 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5763 zone->pageblock_flags = NULL;
5765 zone->pageblock_flags =
5766 memblock_virt_alloc_node_nopanic(usemapsize,
5770 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5771 unsigned long zone_start_pfn, unsigned long zonesize) {}
5772 #endif /* CONFIG_SPARSEMEM */
5774 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5776 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5777 void __paginginit set_pageblock_order(void)
5781 /* Check that pageblock_nr_pages has not already been setup */
5782 if (pageblock_order)
5785 if (HPAGE_SHIFT > PAGE_SHIFT)
5786 order = HUGETLB_PAGE_ORDER;
5788 order = MAX_ORDER - 1;
5791 * Assume the largest contiguous order of interest is a huge page.
5792 * This value may be variable depending on boot parameters on IA64 and
5795 pageblock_order = order;
5797 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5800 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5801 * is unused as pageblock_order is set at compile-time. See
5802 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5805 void __paginginit set_pageblock_order(void)
5809 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5811 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5812 unsigned long present_pages)
5814 unsigned long pages = spanned_pages;
5817 * Provide a more accurate estimation if there are holes within
5818 * the zone and SPARSEMEM is in use. If there are holes within the
5819 * zone, each populated memory region may cost us one or two extra
5820 * memmap pages due to alignment because memmap pages for each
5821 * populated regions may not naturally algined on page boundary.
5822 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5824 if (spanned_pages > present_pages + (present_pages >> 4) &&
5825 IS_ENABLED(CONFIG_SPARSEMEM))
5826 pages = present_pages;
5828 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5832 * Set up the zone data structures:
5833 * - mark all pages reserved
5834 * - mark all memory queues empty
5835 * - clear the memory bitmaps
5837 * NOTE: pgdat should get zeroed by caller.
5839 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5842 int nid = pgdat->node_id;
5845 pgdat_resize_init(pgdat);
5846 #ifdef CONFIG_NUMA_BALANCING
5847 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5848 pgdat->numabalancing_migrate_nr_pages = 0;
5849 pgdat->numabalancing_migrate_next_window = jiffies;
5851 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5852 spin_lock_init(&pgdat->split_queue_lock);
5853 INIT_LIST_HEAD(&pgdat->split_queue);
5854 pgdat->split_queue_len = 0;
5856 init_waitqueue_head(&pgdat->kswapd_wait);
5857 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5858 #ifdef CONFIG_COMPACTION
5859 init_waitqueue_head(&pgdat->kcompactd_wait);
5861 pgdat_page_ext_init(pgdat);
5862 spin_lock_init(&pgdat->lru_lock);
5863 lruvec_init(node_lruvec(pgdat));
5865 for (j = 0; j < MAX_NR_ZONES; j++) {
5866 struct zone *zone = pgdat->node_zones + j;
5867 unsigned long size, realsize, freesize, memmap_pages;
5868 unsigned long zone_start_pfn = zone->zone_start_pfn;
5870 size = zone->spanned_pages;
5871 realsize = freesize = zone->present_pages;
5874 * Adjust freesize so that it accounts for how much memory
5875 * is used by this zone for memmap. This affects the watermark
5876 * and per-cpu initialisations
5878 memmap_pages = calc_memmap_size(size, realsize);
5879 if (!is_highmem_idx(j)) {
5880 if (freesize >= memmap_pages) {
5881 freesize -= memmap_pages;
5884 " %s zone: %lu pages used for memmap\n",
5885 zone_names[j], memmap_pages);
5887 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5888 zone_names[j], memmap_pages, freesize);
5891 /* Account for reserved pages */
5892 if (j == 0 && freesize > dma_reserve) {
5893 freesize -= dma_reserve;
5894 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5895 zone_names[0], dma_reserve);
5898 if (!is_highmem_idx(j))
5899 nr_kernel_pages += freesize;
5900 /* Charge for highmem memmap if there are enough kernel pages */
5901 else if (nr_kernel_pages > memmap_pages * 2)
5902 nr_kernel_pages -= memmap_pages;
5903 nr_all_pages += freesize;
5906 * Set an approximate value for lowmem here, it will be adjusted
5907 * when the bootmem allocator frees pages into the buddy system.
5908 * And all highmem pages will be managed by the buddy system.
5910 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5913 pgdat->min_unmapped_pages += (freesize*sysctl_min_unmapped_ratio)
5915 pgdat->min_slab_pages += (freesize * sysctl_min_slab_ratio) / 100;
5917 zone->name = zone_names[j];
5918 zone->zone_pgdat = pgdat;
5919 spin_lock_init(&zone->lock);
5920 zone_seqlock_init(zone);
5921 zone_pcp_init(zone);
5926 set_pageblock_order();
5927 setup_usemap(pgdat, zone, zone_start_pfn, size);
5928 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5930 memmap_init(size, nid, j, zone_start_pfn);
5934 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5936 unsigned long __maybe_unused start = 0;
5937 unsigned long __maybe_unused offset = 0;
5939 /* Skip empty nodes */
5940 if (!pgdat->node_spanned_pages)
5943 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5944 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5945 offset = pgdat->node_start_pfn - start;
5946 /* ia64 gets its own node_mem_map, before this, without bootmem */
5947 if (!pgdat->node_mem_map) {
5948 unsigned long size, end;
5952 * The zone's endpoints aren't required to be MAX_ORDER
5953 * aligned but the node_mem_map endpoints must be in order
5954 * for the buddy allocator to function correctly.
5956 end = pgdat_end_pfn(pgdat);
5957 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5958 size = (end - start) * sizeof(struct page);
5959 map = alloc_remap(pgdat->node_id, size);
5961 map = memblock_virt_alloc_node_nopanic(size,
5963 pgdat->node_mem_map = map + offset;
5965 #ifndef CONFIG_NEED_MULTIPLE_NODES
5967 * With no DISCONTIG, the global mem_map is just set as node 0's
5969 if (pgdat == NODE_DATA(0)) {
5970 mem_map = NODE_DATA(0)->node_mem_map;
5971 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5972 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5974 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5977 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5980 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5981 unsigned long node_start_pfn, unsigned long *zholes_size)
5983 pg_data_t *pgdat = NODE_DATA(nid);
5984 unsigned long start_pfn = 0;
5985 unsigned long end_pfn = 0;
5987 /* pg_data_t should be reset to zero when it's allocated */
5988 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5990 reset_deferred_meminit(pgdat);
5991 pgdat->node_id = nid;
5992 pgdat->node_start_pfn = node_start_pfn;
5993 pgdat->per_cpu_nodestats = NULL;
5994 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5995 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5996 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5997 (u64)start_pfn << PAGE_SHIFT,
5998 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6000 start_pfn = node_start_pfn;
6002 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6003 zones_size, zholes_size);
6005 alloc_node_mem_map(pgdat);
6006 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6007 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6008 nid, (unsigned long)pgdat,
6009 (unsigned long)pgdat->node_mem_map);
6012 free_area_init_core(pgdat);
6015 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6017 #if MAX_NUMNODES > 1
6019 * Figure out the number of possible node ids.
6021 void __init setup_nr_node_ids(void)
6023 unsigned int highest;
6025 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6026 nr_node_ids = highest + 1;
6031 * node_map_pfn_alignment - determine the maximum internode alignment
6033 * This function should be called after node map is populated and sorted.
6034 * It calculates the maximum power of two alignment which can distinguish
6037 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6038 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6039 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6040 * shifted, 1GiB is enough and this function will indicate so.
6042 * This is used to test whether pfn -> nid mapping of the chosen memory
6043 * model has fine enough granularity to avoid incorrect mapping for the
6044 * populated node map.
6046 * Returns the determined alignment in pfn's. 0 if there is no alignment
6047 * requirement (single node).
6049 unsigned long __init node_map_pfn_alignment(void)
6051 unsigned long accl_mask = 0, last_end = 0;
6052 unsigned long start, end, mask;
6056 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6057 if (!start || last_nid < 0 || last_nid == nid) {
6064 * Start with a mask granular enough to pin-point to the
6065 * start pfn and tick off bits one-by-one until it becomes
6066 * too coarse to separate the current node from the last.
6068 mask = ~((1 << __ffs(start)) - 1);
6069 while (mask && last_end <= (start & (mask << 1)))
6072 /* accumulate all internode masks */
6076 /* convert mask to number of pages */
6077 return ~accl_mask + 1;
6080 /* Find the lowest pfn for a node */
6081 static unsigned long __init find_min_pfn_for_node(int nid)
6083 unsigned long min_pfn = ULONG_MAX;
6084 unsigned long start_pfn;
6087 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6088 min_pfn = min(min_pfn, start_pfn);
6090 if (min_pfn == ULONG_MAX) {
6091 pr_warn("Could not find start_pfn for node %d\n", nid);
6099 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6101 * It returns the minimum PFN based on information provided via
6102 * memblock_set_node().
6104 unsigned long __init find_min_pfn_with_active_regions(void)
6106 return find_min_pfn_for_node(MAX_NUMNODES);
6110 * early_calculate_totalpages()
6111 * Sum pages in active regions for movable zone.
6112 * Populate N_MEMORY for calculating usable_nodes.
6114 static unsigned long __init early_calculate_totalpages(void)
6116 unsigned long totalpages = 0;
6117 unsigned long start_pfn, end_pfn;
6120 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6121 unsigned long pages = end_pfn - start_pfn;
6123 totalpages += pages;
6125 node_set_state(nid, N_MEMORY);
6131 * Find the PFN the Movable zone begins in each node. Kernel memory
6132 * is spread evenly between nodes as long as the nodes have enough
6133 * memory. When they don't, some nodes will have more kernelcore than
6136 static void __init find_zone_movable_pfns_for_nodes(void)
6139 unsigned long usable_startpfn;
6140 unsigned long kernelcore_node, kernelcore_remaining;
6141 /* save the state before borrow the nodemask */
6142 nodemask_t saved_node_state = node_states[N_MEMORY];
6143 unsigned long totalpages = early_calculate_totalpages();
6144 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6145 struct memblock_region *r;
6147 /* Need to find movable_zone earlier when movable_node is specified. */
6148 find_usable_zone_for_movable();
6151 * If movable_node is specified, ignore kernelcore and movablecore
6154 if (movable_node_is_enabled()) {
6155 for_each_memblock(memory, r) {
6156 if (!memblock_is_hotpluggable(r))
6161 usable_startpfn = PFN_DOWN(r->base);
6162 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6163 min(usable_startpfn, zone_movable_pfn[nid]) :
6171 * If kernelcore=mirror is specified, ignore movablecore option
6173 if (mirrored_kernelcore) {
6174 bool mem_below_4gb_not_mirrored = false;
6176 for_each_memblock(memory, r) {
6177 if (memblock_is_mirror(r))
6182 usable_startpfn = memblock_region_memory_base_pfn(r);
6184 if (usable_startpfn < 0x100000) {
6185 mem_below_4gb_not_mirrored = true;
6189 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6190 min(usable_startpfn, zone_movable_pfn[nid]) :
6194 if (mem_below_4gb_not_mirrored)
6195 pr_warn("This configuration results in unmirrored kernel memory.");
6201 * If movablecore=nn[KMG] was specified, calculate what size of
6202 * kernelcore that corresponds so that memory usable for
6203 * any allocation type is evenly spread. If both kernelcore
6204 * and movablecore are specified, then the value of kernelcore
6205 * will be used for required_kernelcore if it's greater than
6206 * what movablecore would have allowed.
6208 if (required_movablecore) {
6209 unsigned long corepages;
6212 * Round-up so that ZONE_MOVABLE is at least as large as what
6213 * was requested by the user
6215 required_movablecore =
6216 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6217 required_movablecore = min(totalpages, required_movablecore);
6218 corepages = totalpages - required_movablecore;
6220 required_kernelcore = max(required_kernelcore, corepages);
6224 * If kernelcore was not specified or kernelcore size is larger
6225 * than totalpages, there is no ZONE_MOVABLE.
6227 if (!required_kernelcore || required_kernelcore >= totalpages)
6230 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6231 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6234 /* Spread kernelcore memory as evenly as possible throughout nodes */
6235 kernelcore_node = required_kernelcore / usable_nodes;
6236 for_each_node_state(nid, N_MEMORY) {
6237 unsigned long start_pfn, end_pfn;
6240 * Recalculate kernelcore_node if the division per node
6241 * now exceeds what is necessary to satisfy the requested
6242 * amount of memory for the kernel
6244 if (required_kernelcore < kernelcore_node)
6245 kernelcore_node = required_kernelcore / usable_nodes;
6248 * As the map is walked, we track how much memory is usable
6249 * by the kernel using kernelcore_remaining. When it is
6250 * 0, the rest of the node is usable by ZONE_MOVABLE
6252 kernelcore_remaining = kernelcore_node;
6254 /* Go through each range of PFNs within this node */
6255 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6256 unsigned long size_pages;
6258 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6259 if (start_pfn >= end_pfn)
6262 /* Account for what is only usable for kernelcore */
6263 if (start_pfn < usable_startpfn) {
6264 unsigned long kernel_pages;
6265 kernel_pages = min(end_pfn, usable_startpfn)
6268 kernelcore_remaining -= min(kernel_pages,
6269 kernelcore_remaining);
6270 required_kernelcore -= min(kernel_pages,
6271 required_kernelcore);
6273 /* Continue if range is now fully accounted */
6274 if (end_pfn <= usable_startpfn) {
6277 * Push zone_movable_pfn to the end so
6278 * that if we have to rebalance
6279 * kernelcore across nodes, we will
6280 * not double account here
6282 zone_movable_pfn[nid] = end_pfn;
6285 start_pfn = usable_startpfn;
6289 * The usable PFN range for ZONE_MOVABLE is from
6290 * start_pfn->end_pfn. Calculate size_pages as the
6291 * number of pages used as kernelcore
6293 size_pages = end_pfn - start_pfn;
6294 if (size_pages > kernelcore_remaining)
6295 size_pages = kernelcore_remaining;
6296 zone_movable_pfn[nid] = start_pfn + size_pages;
6299 * Some kernelcore has been met, update counts and
6300 * break if the kernelcore for this node has been
6303 required_kernelcore -= min(required_kernelcore,
6305 kernelcore_remaining -= size_pages;
6306 if (!kernelcore_remaining)
6312 * If there is still required_kernelcore, we do another pass with one
6313 * less node in the count. This will push zone_movable_pfn[nid] further
6314 * along on the nodes that still have memory until kernelcore is
6318 if (usable_nodes && required_kernelcore > usable_nodes)
6322 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6323 for (nid = 0; nid < MAX_NUMNODES; nid++)
6324 zone_movable_pfn[nid] =
6325 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6328 /* restore the node_state */
6329 node_states[N_MEMORY] = saved_node_state;
6332 /* Any regular or high memory on that node ? */
6333 static void check_for_memory(pg_data_t *pgdat, int nid)
6335 enum zone_type zone_type;
6337 if (N_MEMORY == N_NORMAL_MEMORY)
6340 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6341 struct zone *zone = &pgdat->node_zones[zone_type];
6342 if (populated_zone(zone)) {
6343 node_set_state(nid, N_HIGH_MEMORY);
6344 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6345 zone_type <= ZONE_NORMAL)
6346 node_set_state(nid, N_NORMAL_MEMORY);
6353 * free_area_init_nodes - Initialise all pg_data_t and zone data
6354 * @max_zone_pfn: an array of max PFNs for each zone
6356 * This will call free_area_init_node() for each active node in the system.
6357 * Using the page ranges provided by memblock_set_node(), the size of each
6358 * zone in each node and their holes is calculated. If the maximum PFN
6359 * between two adjacent zones match, it is assumed that the zone is empty.
6360 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6361 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6362 * starts where the previous one ended. For example, ZONE_DMA32 starts
6363 * at arch_max_dma_pfn.
6365 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6367 unsigned long start_pfn, end_pfn;
6370 /* Record where the zone boundaries are */
6371 memset(arch_zone_lowest_possible_pfn, 0,
6372 sizeof(arch_zone_lowest_possible_pfn));
6373 memset(arch_zone_highest_possible_pfn, 0,
6374 sizeof(arch_zone_highest_possible_pfn));
6376 start_pfn = find_min_pfn_with_active_regions();
6378 for (i = 0; i < MAX_NR_ZONES; i++) {
6379 if (i == ZONE_MOVABLE)
6382 end_pfn = max(max_zone_pfn[i], start_pfn);
6383 arch_zone_lowest_possible_pfn[i] = start_pfn;
6384 arch_zone_highest_possible_pfn[i] = end_pfn;
6386 start_pfn = end_pfn;
6388 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6389 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6391 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6392 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6393 find_zone_movable_pfns_for_nodes();
6395 /* Print out the zone ranges */
6396 pr_info("Zone ranges:\n");
6397 for (i = 0; i < MAX_NR_ZONES; i++) {
6398 if (i == ZONE_MOVABLE)
6400 pr_info(" %-8s ", zone_names[i]);
6401 if (arch_zone_lowest_possible_pfn[i] ==
6402 arch_zone_highest_possible_pfn[i])
6405 pr_cont("[mem %#018Lx-%#018Lx]\n",
6406 (u64)arch_zone_lowest_possible_pfn[i]
6408 ((u64)arch_zone_highest_possible_pfn[i]
6409 << PAGE_SHIFT) - 1);
6412 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6413 pr_info("Movable zone start for each node\n");
6414 for (i = 0; i < MAX_NUMNODES; i++) {
6415 if (zone_movable_pfn[i])
6416 pr_info(" Node %d: %#018Lx\n", i,
6417 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6420 /* Print out the early node map */
6421 pr_info("Early memory node ranges\n");
6422 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6423 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6424 (u64)start_pfn << PAGE_SHIFT,
6425 ((u64)end_pfn << PAGE_SHIFT) - 1);
6427 /* Initialise every node */
6428 mminit_verify_pageflags_layout();
6429 setup_nr_node_ids();
6430 for_each_online_node(nid) {
6431 pg_data_t *pgdat = NODE_DATA(nid);
6432 free_area_init_node(nid, NULL,
6433 find_min_pfn_for_node(nid), NULL);
6435 /* Any memory on that node */
6436 if (pgdat->node_present_pages)
6437 node_set_state(nid, N_MEMORY);
6438 check_for_memory(pgdat, nid);
6442 static int __init cmdline_parse_core(char *p, unsigned long *core)
6444 unsigned long long coremem;
6448 coremem = memparse(p, &p);
6449 *core = coremem >> PAGE_SHIFT;
6451 /* Paranoid check that UL is enough for the coremem value */
6452 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6458 * kernelcore=size sets the amount of memory for use for allocations that
6459 * cannot be reclaimed or migrated.
6461 static int __init cmdline_parse_kernelcore(char *p)
6463 /* parse kernelcore=mirror */
6464 if (parse_option_str(p, "mirror")) {
6465 mirrored_kernelcore = true;
6469 return cmdline_parse_core(p, &required_kernelcore);
6473 * movablecore=size sets the amount of memory for use for allocations that
6474 * can be reclaimed or migrated.
6476 static int __init cmdline_parse_movablecore(char *p)
6478 return cmdline_parse_core(p, &required_movablecore);
6481 early_param("kernelcore", cmdline_parse_kernelcore);
6482 early_param("movablecore", cmdline_parse_movablecore);
6484 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6486 void adjust_managed_page_count(struct page *page, long count)
6488 spin_lock(&managed_page_count_lock);
6489 page_zone(page)->managed_pages += count;
6490 totalram_pages += count;
6491 #ifdef CONFIG_HIGHMEM
6492 if (PageHighMem(page))
6493 totalhigh_pages += count;
6495 spin_unlock(&managed_page_count_lock);
6497 EXPORT_SYMBOL(adjust_managed_page_count);
6499 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6502 unsigned long pages = 0;
6504 start = (void *)PAGE_ALIGN((unsigned long)start);
6505 end = (void *)((unsigned long)end & PAGE_MASK);
6506 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6507 if ((unsigned int)poison <= 0xFF)
6508 memset(pos, poison, PAGE_SIZE);
6509 free_reserved_page(virt_to_page(pos));
6513 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6514 s, pages << (PAGE_SHIFT - 10), start, end);
6518 EXPORT_SYMBOL(free_reserved_area);
6520 #ifdef CONFIG_HIGHMEM
6521 void free_highmem_page(struct page *page)
6523 __free_reserved_page(page);
6525 page_zone(page)->managed_pages++;
6531 void __init mem_init_print_info(const char *str)
6533 unsigned long physpages, codesize, datasize, rosize, bss_size;
6534 unsigned long init_code_size, init_data_size;
6536 physpages = get_num_physpages();
6537 codesize = _etext - _stext;
6538 datasize = _edata - _sdata;
6539 rosize = __end_rodata - __start_rodata;
6540 bss_size = __bss_stop - __bss_start;
6541 init_data_size = __init_end - __init_begin;
6542 init_code_size = _einittext - _sinittext;
6545 * Detect special cases and adjust section sizes accordingly:
6546 * 1) .init.* may be embedded into .data sections
6547 * 2) .init.text.* may be out of [__init_begin, __init_end],
6548 * please refer to arch/tile/kernel/vmlinux.lds.S.
6549 * 3) .rodata.* may be embedded into .text or .data sections.
6551 #define adj_init_size(start, end, size, pos, adj) \
6553 if (start <= pos && pos < end && size > adj) \
6557 adj_init_size(__init_begin, __init_end, init_data_size,
6558 _sinittext, init_code_size);
6559 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6560 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6561 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6562 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6564 #undef adj_init_size
6566 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6567 #ifdef CONFIG_HIGHMEM
6571 nr_free_pages() << (PAGE_SHIFT - 10),
6572 physpages << (PAGE_SHIFT - 10),
6573 codesize >> 10, datasize >> 10, rosize >> 10,
6574 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6575 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6576 totalcma_pages << (PAGE_SHIFT - 10),
6577 #ifdef CONFIG_HIGHMEM
6578 totalhigh_pages << (PAGE_SHIFT - 10),
6580 str ? ", " : "", str ? str : "");
6584 * set_dma_reserve - set the specified number of pages reserved in the first zone
6585 * @new_dma_reserve: The number of pages to mark reserved
6587 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6588 * In the DMA zone, a significant percentage may be consumed by kernel image
6589 * and other unfreeable allocations which can skew the watermarks badly. This
6590 * function may optionally be used to account for unfreeable pages in the
6591 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6592 * smaller per-cpu batchsize.
6594 void __init set_dma_reserve(unsigned long new_dma_reserve)
6596 dma_reserve = new_dma_reserve;
6599 void __init free_area_init(unsigned long *zones_size)
6601 free_area_init_node(0, zones_size,
6602 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6605 static int page_alloc_cpu_notify(struct notifier_block *self,
6606 unsigned long action, void *hcpu)
6608 int cpu = (unsigned long)hcpu;
6610 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6611 lru_add_drain_cpu(cpu);
6615 * Spill the event counters of the dead processor
6616 * into the current processors event counters.
6617 * This artificially elevates the count of the current
6620 vm_events_fold_cpu(cpu);
6623 * Zero the differential counters of the dead processor
6624 * so that the vm statistics are consistent.
6626 * This is only okay since the processor is dead and cannot
6627 * race with what we are doing.
6629 cpu_vm_stats_fold(cpu);
6634 void __init page_alloc_init(void)
6636 hotcpu_notifier(page_alloc_cpu_notify, 0);
6640 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6641 * or min_free_kbytes changes.
6643 static void calculate_totalreserve_pages(void)
6645 struct pglist_data *pgdat;
6646 unsigned long reserve_pages = 0;
6647 enum zone_type i, j;
6649 for_each_online_pgdat(pgdat) {
6651 pgdat->totalreserve_pages = 0;
6653 for (i = 0; i < MAX_NR_ZONES; i++) {
6654 struct zone *zone = pgdat->node_zones + i;
6657 /* Find valid and maximum lowmem_reserve in the zone */
6658 for (j = i; j < MAX_NR_ZONES; j++) {
6659 if (zone->lowmem_reserve[j] > max)
6660 max = zone->lowmem_reserve[j];
6663 /* we treat the high watermark as reserved pages. */
6664 max += high_wmark_pages(zone);
6666 if (max > zone->managed_pages)
6667 max = zone->managed_pages;
6669 pgdat->totalreserve_pages += max;
6671 reserve_pages += max;
6674 totalreserve_pages = reserve_pages;
6678 * setup_per_zone_lowmem_reserve - called whenever
6679 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6680 * has a correct pages reserved value, so an adequate number of
6681 * pages are left in the zone after a successful __alloc_pages().
6683 static void setup_per_zone_lowmem_reserve(void)
6685 struct pglist_data *pgdat;
6686 enum zone_type j, idx;
6688 for_each_online_pgdat(pgdat) {
6689 for (j = 0; j < MAX_NR_ZONES; j++) {
6690 struct zone *zone = pgdat->node_zones + j;
6691 unsigned long managed_pages = zone->managed_pages;
6693 zone->lowmem_reserve[j] = 0;
6697 struct zone *lower_zone;
6701 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6702 sysctl_lowmem_reserve_ratio[idx] = 1;
6704 lower_zone = pgdat->node_zones + idx;
6705 lower_zone->lowmem_reserve[j] = managed_pages /
6706 sysctl_lowmem_reserve_ratio[idx];
6707 managed_pages += lower_zone->managed_pages;
6712 /* update totalreserve_pages */
6713 calculate_totalreserve_pages();
6716 static void __setup_per_zone_wmarks(void)
6718 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6719 unsigned long lowmem_pages = 0;
6721 unsigned long flags;
6723 /* Calculate total number of !ZONE_HIGHMEM pages */
6724 for_each_zone(zone) {
6725 if (!is_highmem(zone))
6726 lowmem_pages += zone->managed_pages;
6729 for_each_zone(zone) {
6732 spin_lock_irqsave(&zone->lock, flags);
6733 tmp = (u64)pages_min * zone->managed_pages;
6734 do_div(tmp, lowmem_pages);
6735 if (is_highmem(zone)) {
6737 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6738 * need highmem pages, so cap pages_min to a small
6741 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6742 * deltas control asynch page reclaim, and so should
6743 * not be capped for highmem.
6745 unsigned long min_pages;
6747 min_pages = zone->managed_pages / 1024;
6748 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6749 zone->watermark[WMARK_MIN] = min_pages;
6752 * If it's a lowmem zone, reserve a number of pages
6753 * proportionate to the zone's size.
6755 zone->watermark[WMARK_MIN] = tmp;
6759 * Set the kswapd watermarks distance according to the
6760 * scale factor in proportion to available memory, but
6761 * ensure a minimum size on small systems.
6763 tmp = max_t(u64, tmp >> 2,
6764 mult_frac(zone->managed_pages,
6765 watermark_scale_factor, 10000));
6767 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6768 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6770 spin_unlock_irqrestore(&zone->lock, flags);
6773 /* update totalreserve_pages */
6774 calculate_totalreserve_pages();
6778 * setup_per_zone_wmarks - called when min_free_kbytes changes
6779 * or when memory is hot-{added|removed}
6781 * Ensures that the watermark[min,low,high] values for each zone are set
6782 * correctly with respect to min_free_kbytes.
6784 void setup_per_zone_wmarks(void)
6786 mutex_lock(&zonelists_mutex);
6787 __setup_per_zone_wmarks();
6788 mutex_unlock(&zonelists_mutex);
6792 * Initialise min_free_kbytes.
6794 * For small machines we want it small (128k min). For large machines
6795 * we want it large (64MB max). But it is not linear, because network
6796 * bandwidth does not increase linearly with machine size. We use
6798 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6799 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6815 int __meminit init_per_zone_wmark_min(void)
6817 unsigned long lowmem_kbytes;
6818 int new_min_free_kbytes;
6820 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6821 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6823 if (new_min_free_kbytes > user_min_free_kbytes) {
6824 min_free_kbytes = new_min_free_kbytes;
6825 if (min_free_kbytes < 128)
6826 min_free_kbytes = 128;
6827 if (min_free_kbytes > 65536)
6828 min_free_kbytes = 65536;
6830 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6831 new_min_free_kbytes, user_min_free_kbytes);
6833 setup_per_zone_wmarks();
6834 refresh_zone_stat_thresholds();
6835 setup_per_zone_lowmem_reserve();
6838 core_initcall(init_per_zone_wmark_min)
6841 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6842 * that we can call two helper functions whenever min_free_kbytes
6845 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6846 void __user *buffer, size_t *length, loff_t *ppos)
6850 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6855 user_min_free_kbytes = min_free_kbytes;
6856 setup_per_zone_wmarks();
6861 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6862 void __user *buffer, size_t *length, loff_t *ppos)
6866 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6871 setup_per_zone_wmarks();
6877 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6878 void __user *buffer, size_t *length, loff_t *ppos)
6880 struct pglist_data *pgdat;
6884 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6888 for_each_online_pgdat(pgdat)
6889 pgdat->min_slab_pages = 0;
6892 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6893 sysctl_min_unmapped_ratio) / 100;
6897 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6898 void __user *buffer, size_t *length, loff_t *ppos)
6900 struct pglist_data *pgdat;
6904 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6908 for_each_online_pgdat(pgdat)
6909 pgdat->min_slab_pages = 0;
6912 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6913 sysctl_min_slab_ratio) / 100;
6919 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6920 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6921 * whenever sysctl_lowmem_reserve_ratio changes.
6923 * The reserve ratio obviously has absolutely no relation with the
6924 * minimum watermarks. The lowmem reserve ratio can only make sense
6925 * if in function of the boot time zone sizes.
6927 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6928 void __user *buffer, size_t *length, loff_t *ppos)
6930 proc_dointvec_minmax(table, write, buffer, length, ppos);
6931 setup_per_zone_lowmem_reserve();
6936 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6937 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6938 * pagelist can have before it gets flushed back to buddy allocator.
6940 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6941 void __user *buffer, size_t *length, loff_t *ppos)
6944 int old_percpu_pagelist_fraction;
6947 mutex_lock(&pcp_batch_high_lock);
6948 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6950 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6951 if (!write || ret < 0)
6954 /* Sanity checking to avoid pcp imbalance */
6955 if (percpu_pagelist_fraction &&
6956 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6957 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6963 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6966 for_each_populated_zone(zone) {
6969 for_each_possible_cpu(cpu)
6970 pageset_set_high_and_batch(zone,
6971 per_cpu_ptr(zone->pageset, cpu));
6974 mutex_unlock(&pcp_batch_high_lock);
6979 int hashdist = HASHDIST_DEFAULT;
6981 static int __init set_hashdist(char *str)
6985 hashdist = simple_strtoul(str, &str, 0);
6988 __setup("hashdist=", set_hashdist);
6992 * allocate a large system hash table from bootmem
6993 * - it is assumed that the hash table must contain an exact power-of-2
6994 * quantity of entries
6995 * - limit is the number of hash buckets, not the total allocation size
6997 void *__init alloc_large_system_hash(const char *tablename,
6998 unsigned long bucketsize,
6999 unsigned long numentries,
7002 unsigned int *_hash_shift,
7003 unsigned int *_hash_mask,
7004 unsigned long low_limit,
7005 unsigned long high_limit)
7007 unsigned long long max = high_limit;
7008 unsigned long log2qty, size;
7011 /* allow the kernel cmdline to have a say */
7013 /* round applicable memory size up to nearest megabyte */
7014 numentries = nr_kernel_pages;
7016 /* It isn't necessary when PAGE_SIZE >= 1MB */
7017 if (PAGE_SHIFT < 20)
7018 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7020 /* limit to 1 bucket per 2^scale bytes of low memory */
7021 if (scale > PAGE_SHIFT)
7022 numentries >>= (scale - PAGE_SHIFT);
7024 numentries <<= (PAGE_SHIFT - scale);
7026 /* Make sure we've got at least a 0-order allocation.. */
7027 if (unlikely(flags & HASH_SMALL)) {
7028 /* Makes no sense without HASH_EARLY */
7029 WARN_ON(!(flags & HASH_EARLY));
7030 if (!(numentries >> *_hash_shift)) {
7031 numentries = 1UL << *_hash_shift;
7032 BUG_ON(!numentries);
7034 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7035 numentries = PAGE_SIZE / bucketsize;
7037 numentries = roundup_pow_of_two(numentries);
7039 /* limit allocation size to 1/16 total memory by default */
7041 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7042 do_div(max, bucketsize);
7044 max = min(max, 0x80000000ULL);
7046 if (numentries < low_limit)
7047 numentries = low_limit;
7048 if (numentries > max)
7051 log2qty = ilog2(numentries);
7054 size = bucketsize << log2qty;
7055 if (flags & HASH_EARLY)
7056 table = memblock_virt_alloc_nopanic(size, 0);
7058 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7061 * If bucketsize is not a power-of-two, we may free
7062 * some pages at the end of hash table which
7063 * alloc_pages_exact() automatically does
7065 if (get_order(size) < MAX_ORDER) {
7066 table = alloc_pages_exact(size, GFP_ATOMIC);
7067 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7070 } while (!table && size > PAGE_SIZE && --log2qty);
7073 panic("Failed to allocate %s hash table\n", tablename);
7075 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7076 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7079 *_hash_shift = log2qty;
7081 *_hash_mask = (1 << log2qty) - 1;
7087 * This function checks whether pageblock includes unmovable pages or not.
7088 * If @count is not zero, it is okay to include less @count unmovable pages
7090 * PageLRU check without isolation or lru_lock could race so that
7091 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7092 * expect this function should be exact.
7094 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7095 bool skip_hwpoisoned_pages)
7097 unsigned long pfn, iter, found;
7101 * For avoiding noise data, lru_add_drain_all() should be called
7102 * If ZONE_MOVABLE, the zone never contains unmovable pages
7104 if (zone_idx(zone) == ZONE_MOVABLE)
7106 mt = get_pageblock_migratetype(page);
7107 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7110 pfn = page_to_pfn(page);
7111 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7112 unsigned long check = pfn + iter;
7114 if (!pfn_valid_within(check))
7117 page = pfn_to_page(check);
7120 * Hugepages are not in LRU lists, but they're movable.
7121 * We need not scan over tail pages bacause we don't
7122 * handle each tail page individually in migration.
7124 if (PageHuge(page)) {
7125 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7130 * We can't use page_count without pin a page
7131 * because another CPU can free compound page.
7132 * This check already skips compound tails of THP
7133 * because their page->_refcount is zero at all time.
7135 if (!page_ref_count(page)) {
7136 if (PageBuddy(page))
7137 iter += (1 << page_order(page)) - 1;
7142 * The HWPoisoned page may be not in buddy system, and
7143 * page_count() is not 0.
7145 if (skip_hwpoisoned_pages && PageHWPoison(page))
7151 * If there are RECLAIMABLE pages, we need to check
7152 * it. But now, memory offline itself doesn't call
7153 * shrink_node_slabs() and it still to be fixed.
7156 * If the page is not RAM, page_count()should be 0.
7157 * we don't need more check. This is an _used_ not-movable page.
7159 * The problematic thing here is PG_reserved pages. PG_reserved
7160 * is set to both of a memory hole page and a _used_ kernel
7169 bool is_pageblock_removable_nolock(struct page *page)
7175 * We have to be careful here because we are iterating over memory
7176 * sections which are not zone aware so we might end up outside of
7177 * the zone but still within the section.
7178 * We have to take care about the node as well. If the node is offline
7179 * its NODE_DATA will be NULL - see page_zone.
7181 if (!node_online(page_to_nid(page)))
7184 zone = page_zone(page);
7185 pfn = page_to_pfn(page);
7186 if (!zone_spans_pfn(zone, pfn))
7189 return !has_unmovable_pages(zone, page, 0, true);
7192 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7194 static unsigned long pfn_max_align_down(unsigned long pfn)
7196 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7197 pageblock_nr_pages) - 1);
7200 static unsigned long pfn_max_align_up(unsigned long pfn)
7202 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7203 pageblock_nr_pages));
7206 /* [start, end) must belong to a single zone. */
7207 static int __alloc_contig_migrate_range(struct compact_control *cc,
7208 unsigned long start, unsigned long end)
7210 /* This function is based on compact_zone() from compaction.c. */
7211 unsigned long nr_reclaimed;
7212 unsigned long pfn = start;
7213 unsigned int tries = 0;
7218 while (pfn < end || !list_empty(&cc->migratepages)) {
7219 if (fatal_signal_pending(current)) {
7224 if (list_empty(&cc->migratepages)) {
7225 cc->nr_migratepages = 0;
7226 pfn = isolate_migratepages_range(cc, pfn, end);
7232 } else if (++tries == 5) {
7233 ret = ret < 0 ? ret : -EBUSY;
7237 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7239 cc->nr_migratepages -= nr_reclaimed;
7241 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7242 NULL, 0, cc->mode, MR_CMA);
7245 putback_movable_pages(&cc->migratepages);
7252 * alloc_contig_range() -- tries to allocate given range of pages
7253 * @start: start PFN to allocate
7254 * @end: one-past-the-last PFN to allocate
7255 * @migratetype: migratetype of the underlaying pageblocks (either
7256 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7257 * in range must have the same migratetype and it must
7258 * be either of the two.
7260 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7261 * aligned, however it's the caller's responsibility to guarantee that
7262 * we are the only thread that changes migrate type of pageblocks the
7265 * The PFN range must belong to a single zone.
7267 * Returns zero on success or negative error code. On success all
7268 * pages which PFN is in [start, end) are allocated for the caller and
7269 * need to be freed with free_contig_range().
7271 int alloc_contig_range(unsigned long start, unsigned long end,
7272 unsigned migratetype)
7274 unsigned long outer_start, outer_end;
7278 struct compact_control cc = {
7279 .nr_migratepages = 0,
7281 .zone = page_zone(pfn_to_page(start)),
7282 .mode = MIGRATE_SYNC,
7283 .ignore_skip_hint = true,
7285 INIT_LIST_HEAD(&cc.migratepages);
7288 * What we do here is we mark all pageblocks in range as
7289 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7290 * have different sizes, and due to the way page allocator
7291 * work, we align the range to biggest of the two pages so
7292 * that page allocator won't try to merge buddies from
7293 * different pageblocks and change MIGRATE_ISOLATE to some
7294 * other migration type.
7296 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7297 * migrate the pages from an unaligned range (ie. pages that
7298 * we are interested in). This will put all the pages in
7299 * range back to page allocator as MIGRATE_ISOLATE.
7301 * When this is done, we take the pages in range from page
7302 * allocator removing them from the buddy system. This way
7303 * page allocator will never consider using them.
7305 * This lets us mark the pageblocks back as
7306 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7307 * aligned range but not in the unaligned, original range are
7308 * put back to page allocator so that buddy can use them.
7311 ret = start_isolate_page_range(pfn_max_align_down(start),
7312 pfn_max_align_up(end), migratetype,
7318 * In case of -EBUSY, we'd like to know which page causes problem.
7319 * So, just fall through. We will check it in test_pages_isolated().
7321 ret = __alloc_contig_migrate_range(&cc, start, end);
7322 if (ret && ret != -EBUSY)
7326 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7327 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7328 * more, all pages in [start, end) are free in page allocator.
7329 * What we are going to do is to allocate all pages from
7330 * [start, end) (that is remove them from page allocator).
7332 * The only problem is that pages at the beginning and at the
7333 * end of interesting range may be not aligned with pages that
7334 * page allocator holds, ie. they can be part of higher order
7335 * pages. Because of this, we reserve the bigger range and
7336 * once this is done free the pages we are not interested in.
7338 * We don't have to hold zone->lock here because the pages are
7339 * isolated thus they won't get removed from buddy.
7342 lru_add_drain_all();
7343 drain_all_pages(cc.zone);
7346 outer_start = start;
7347 while (!PageBuddy(pfn_to_page(outer_start))) {
7348 if (++order >= MAX_ORDER) {
7349 outer_start = start;
7352 outer_start &= ~0UL << order;
7355 if (outer_start != start) {
7356 order = page_order(pfn_to_page(outer_start));
7359 * outer_start page could be small order buddy page and
7360 * it doesn't include start page. Adjust outer_start
7361 * in this case to report failed page properly
7362 * on tracepoint in test_pages_isolated()
7364 if (outer_start + (1UL << order) <= start)
7365 outer_start = start;
7368 /* Make sure the range is really isolated. */
7369 if (test_pages_isolated(outer_start, end, false)) {
7370 pr_info("%s: [%lx, %lx) PFNs busy\n",
7371 __func__, outer_start, end);
7376 /* Grab isolated pages from freelists. */
7377 outer_end = isolate_freepages_range(&cc, outer_start, end);
7383 /* Free head and tail (if any) */
7384 if (start != outer_start)
7385 free_contig_range(outer_start, start - outer_start);
7386 if (end != outer_end)
7387 free_contig_range(end, outer_end - end);
7390 undo_isolate_page_range(pfn_max_align_down(start),
7391 pfn_max_align_up(end), migratetype);
7395 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7397 unsigned int count = 0;
7399 for (; nr_pages--; pfn++) {
7400 struct page *page = pfn_to_page(pfn);
7402 count += page_count(page) != 1;
7405 WARN(count != 0, "%d pages are still in use!\n", count);
7409 #ifdef CONFIG_MEMORY_HOTPLUG
7411 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7412 * page high values need to be recalulated.
7414 void __meminit zone_pcp_update(struct zone *zone)
7417 mutex_lock(&pcp_batch_high_lock);
7418 for_each_possible_cpu(cpu)
7419 pageset_set_high_and_batch(zone,
7420 per_cpu_ptr(zone->pageset, cpu));
7421 mutex_unlock(&pcp_batch_high_lock);
7425 void zone_pcp_reset(struct zone *zone)
7427 unsigned long flags;
7429 struct per_cpu_pageset *pset;
7431 /* avoid races with drain_pages() */
7432 local_irq_save(flags);
7433 if (zone->pageset != &boot_pageset) {
7434 for_each_online_cpu(cpu) {
7435 pset = per_cpu_ptr(zone->pageset, cpu);
7436 drain_zonestat(zone, pset);
7438 free_percpu(zone->pageset);
7439 zone->pageset = &boot_pageset;
7441 local_irq_restore(flags);
7444 #ifdef CONFIG_MEMORY_HOTREMOVE
7446 * All pages in the range must be in a single zone and isolated
7447 * before calling this.
7450 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7454 unsigned int order, i;
7456 unsigned long flags;
7457 /* find the first valid pfn */
7458 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7463 zone = page_zone(pfn_to_page(pfn));
7464 spin_lock_irqsave(&zone->lock, flags);
7466 while (pfn < end_pfn) {
7467 if (!pfn_valid(pfn)) {
7471 page = pfn_to_page(pfn);
7473 * The HWPoisoned page may be not in buddy system, and
7474 * page_count() is not 0.
7476 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7478 SetPageReserved(page);
7482 BUG_ON(page_count(page));
7483 BUG_ON(!PageBuddy(page));
7484 order = page_order(page);
7485 #ifdef CONFIG_DEBUG_VM
7486 pr_info("remove from free list %lx %d %lx\n",
7487 pfn, 1 << order, end_pfn);
7489 list_del(&page->lru);
7490 rmv_page_order(page);
7491 zone->free_area[order].nr_free--;
7492 for (i = 0; i < (1 << order); i++)
7493 SetPageReserved((page+i));
7494 pfn += (1 << order);
7496 spin_unlock_irqrestore(&zone->lock, flags);
7500 bool is_free_buddy_page(struct page *page)
7502 struct zone *zone = page_zone(page);
7503 unsigned long pfn = page_to_pfn(page);
7504 unsigned long flags;
7507 spin_lock_irqsave(&zone->lock, flags);
7508 for (order = 0; order < MAX_ORDER; order++) {
7509 struct page *page_head = page - (pfn & ((1 << order) - 1));
7511 if (PageBuddy(page_head) && page_order(page_head) >= order)
7514 spin_unlock_irqrestore(&zone->lock, flags);
7516 return order < MAX_ORDER;