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())
610 if (!debug_guardpage_minorder())
616 static void init_debug_guardpage(void)
618 if (!debug_pagealloc_enabled())
621 if (!debug_guardpage_minorder())
624 _debug_guardpage_enabled = true;
627 struct page_ext_operations debug_guardpage_ops = {
628 .need = need_debug_guardpage,
629 .init = init_debug_guardpage,
632 static int __init debug_guardpage_minorder_setup(char *buf)
636 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
637 pr_err("Bad debug_guardpage_minorder value\n");
640 _debug_guardpage_minorder = res;
641 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
644 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
646 static inline bool set_page_guard(struct zone *zone, struct page *page,
647 unsigned int order, int migratetype)
649 struct page_ext *page_ext;
651 if (!debug_guardpage_enabled())
654 if (order >= debug_guardpage_minorder())
657 page_ext = lookup_page_ext(page);
658 if (unlikely(!page_ext))
661 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
663 INIT_LIST_HEAD(&page->lru);
664 set_page_private(page, order);
665 /* Guard pages are not available for any usage */
666 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
671 static inline void clear_page_guard(struct zone *zone, struct page *page,
672 unsigned int order, int migratetype)
674 struct page_ext *page_ext;
676 if (!debug_guardpage_enabled())
679 page_ext = lookup_page_ext(page);
680 if (unlikely(!page_ext))
683 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
685 set_page_private(page, 0);
686 if (!is_migrate_isolate(migratetype))
687 __mod_zone_freepage_state(zone, (1 << order), migratetype);
690 struct page_ext_operations debug_guardpage_ops;
691 static inline bool set_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype) { return false; }
693 static inline void clear_page_guard(struct zone *zone, struct page *page,
694 unsigned int order, int migratetype) {}
697 static inline void set_page_order(struct page *page, unsigned int order)
699 set_page_private(page, order);
700 __SetPageBuddy(page);
703 static inline void rmv_page_order(struct page *page)
705 __ClearPageBuddy(page);
706 set_page_private(page, 0);
710 * This function checks whether a page is free && is the buddy
711 * we can do coalesce a page and its buddy if
712 * (a) the buddy is not in a hole &&
713 * (b) the buddy is in the buddy system &&
714 * (c) a page and its buddy have the same order &&
715 * (d) a page and its buddy are in the same zone.
717 * For recording whether a page is in the buddy system, we set ->_mapcount
718 * PAGE_BUDDY_MAPCOUNT_VALUE.
719 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
720 * serialized by zone->lock.
722 * For recording page's order, we use page_private(page).
724 static inline int page_is_buddy(struct page *page, struct page *buddy,
727 if (!pfn_valid_within(page_to_pfn(buddy)))
730 if (page_is_guard(buddy) && page_order(buddy) == order) {
731 if (page_zone_id(page) != page_zone_id(buddy))
734 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
739 if (PageBuddy(buddy) && page_order(buddy) == order) {
741 * zone check is done late to avoid uselessly
742 * calculating zone/node ids for pages that could
745 if (page_zone_id(page) != page_zone_id(buddy))
748 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
756 * Freeing function for a buddy system allocator.
758 * The concept of a buddy system is to maintain direct-mapped table
759 * (containing bit values) for memory blocks of various "orders".
760 * The bottom level table contains the map for the smallest allocatable
761 * units of memory (here, pages), and each level above it describes
762 * pairs of units from the levels below, hence, "buddies".
763 * At a high level, all that happens here is marking the table entry
764 * at the bottom level available, and propagating the changes upward
765 * as necessary, plus some accounting needed to play nicely with other
766 * parts of the VM system.
767 * At each level, we keep a list of pages, which are heads of continuous
768 * free pages of length of (1 << order) and marked with _mapcount
769 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
771 * So when we are allocating or freeing one, we can derive the state of the
772 * other. That is, if we allocate a small block, and both were
773 * free, the remainder of the region must be split into blocks.
774 * If a block is freed, and its buddy is also free, then this
775 * triggers coalescing into a block of larger size.
780 static inline void __free_one_page(struct page *page,
782 struct zone *zone, unsigned int order,
785 unsigned long page_idx;
786 unsigned long combined_idx;
787 unsigned long uninitialized_var(buddy_idx);
789 unsigned int max_order;
791 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
793 VM_BUG_ON(!zone_is_initialized(zone));
794 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
796 VM_BUG_ON(migratetype == -1);
797 if (likely(!is_migrate_isolate(migratetype)))
798 __mod_zone_freepage_state(zone, 1 << order, migratetype);
800 page_idx = pfn & ((1 << MAX_ORDER) - 1);
802 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
803 VM_BUG_ON_PAGE(bad_range(zone, page), page);
806 while (order < max_order - 1) {
807 buddy_idx = __find_buddy_index(page_idx, order);
808 buddy = page + (buddy_idx - page_idx);
809 if (!page_is_buddy(page, buddy, order))
812 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
813 * merge with it and move up one order.
815 if (page_is_guard(buddy)) {
816 clear_page_guard(zone, buddy, order, migratetype);
818 list_del(&buddy->lru);
819 zone->free_area[order].nr_free--;
820 rmv_page_order(buddy);
822 combined_idx = buddy_idx & page_idx;
823 page = page + (combined_idx - page_idx);
824 page_idx = combined_idx;
827 if (max_order < MAX_ORDER) {
828 /* If we are here, it means order is >= pageblock_order.
829 * We want to prevent merge between freepages on isolate
830 * pageblock and normal pageblock. Without this, pageblock
831 * isolation could cause incorrect freepage or CMA accounting.
833 * We don't want to hit this code for the more frequent
836 if (unlikely(has_isolate_pageblock(zone))) {
839 buddy_idx = __find_buddy_index(page_idx, order);
840 buddy = page + (buddy_idx - page_idx);
841 buddy_mt = get_pageblock_migratetype(buddy);
843 if (migratetype != buddy_mt
844 && (is_migrate_isolate(migratetype) ||
845 is_migrate_isolate(buddy_mt)))
849 goto continue_merging;
853 set_page_order(page, order);
856 * If this is not the largest possible page, check if the buddy
857 * of the next-highest order is free. If it is, it's possible
858 * that pages are being freed that will coalesce soon. In case,
859 * that is happening, add the free page to the tail of the list
860 * so it's less likely to be used soon and more likely to be merged
861 * as a higher order page
863 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
864 struct page *higher_page, *higher_buddy;
865 combined_idx = buddy_idx & page_idx;
866 higher_page = page + (combined_idx - page_idx);
867 buddy_idx = __find_buddy_index(combined_idx, order + 1);
868 higher_buddy = higher_page + (buddy_idx - combined_idx);
869 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
870 list_add_tail(&page->lru,
871 &zone->free_area[order].free_list[migratetype]);
876 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
878 zone->free_area[order].nr_free++;
882 * A bad page could be due to a number of fields. Instead of multiple branches,
883 * try and check multiple fields with one check. The caller must do a detailed
884 * check if necessary.
886 static inline bool page_expected_state(struct page *page,
887 unsigned long check_flags)
889 if (unlikely(atomic_read(&page->_mapcount) != -1))
892 if (unlikely((unsigned long)page->mapping |
893 page_ref_count(page) |
895 (unsigned long)page->mem_cgroup |
897 (page->flags & check_flags)))
903 static void free_pages_check_bad(struct page *page)
905 const char *bad_reason;
906 unsigned long bad_flags;
911 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 bad_reason = "nonzero mapcount";
913 if (unlikely(page->mapping != NULL))
914 bad_reason = "non-NULL mapping";
915 if (unlikely(page_ref_count(page) != 0))
916 bad_reason = "nonzero _refcount";
917 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
918 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
919 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
922 if (unlikely(page->mem_cgroup))
923 bad_reason = "page still charged to cgroup";
925 bad_page(page, bad_reason, bad_flags);
928 static inline int free_pages_check(struct page *page)
930 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
933 /* Something has gone sideways, find it */
934 free_pages_check_bad(page);
938 static int free_tail_pages_check(struct page *head_page, struct page *page)
943 * We rely page->lru.next never has bit 0 set, unless the page
944 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
946 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
948 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
952 switch (page - head_page) {
954 /* the first tail page: ->mapping is compound_mapcount() */
955 if (unlikely(compound_mapcount(page))) {
956 bad_page(page, "nonzero compound_mapcount", 0);
962 * the second tail page: ->mapping is
963 * page_deferred_list().next -- ignore value.
967 if (page->mapping != TAIL_MAPPING) {
968 bad_page(page, "corrupted mapping in tail page", 0);
973 if (unlikely(!PageTail(page))) {
974 bad_page(page, "PageTail not set", 0);
977 if (unlikely(compound_head(page) != head_page)) {
978 bad_page(page, "compound_head not consistent", 0);
983 page->mapping = NULL;
984 clear_compound_head(page);
988 static __always_inline bool free_pages_prepare(struct page *page,
989 unsigned int order, bool check_free)
993 VM_BUG_ON_PAGE(PageTail(page), page);
995 trace_mm_page_free(page, order);
996 kmemcheck_free_shadow(page, order);
999 * Check tail pages before head page information is cleared to
1000 * avoid checking PageCompound for order-0 pages.
1002 if (unlikely(order)) {
1003 bool compound = PageCompound(page);
1006 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1009 ClearPageDoubleMap(page);
1010 for (i = 1; i < (1 << order); i++) {
1012 bad += free_tail_pages_check(page, page + i);
1013 if (unlikely(free_pages_check(page + i))) {
1017 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1020 if (PageMappingFlags(page))
1021 page->mapping = NULL;
1022 if (memcg_kmem_enabled() && PageKmemcg(page))
1023 memcg_kmem_uncharge(page, order);
1025 bad += free_pages_check(page);
1029 page_cpupid_reset_last(page);
1030 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1031 reset_page_owner(page, order);
1033 if (!PageHighMem(page)) {
1034 debug_check_no_locks_freed(page_address(page),
1035 PAGE_SIZE << order);
1036 debug_check_no_obj_freed(page_address(page),
1037 PAGE_SIZE << order);
1039 arch_free_page(page, order);
1040 kernel_poison_pages(page, 1 << order, 0);
1041 kernel_map_pages(page, 1 << order, 0);
1042 kasan_free_pages(page, order);
1047 #ifdef CONFIG_DEBUG_VM
1048 static inline bool free_pcp_prepare(struct page *page)
1050 return free_pages_prepare(page, 0, true);
1053 static inline bool bulkfree_pcp_prepare(struct page *page)
1058 static bool free_pcp_prepare(struct page *page)
1060 return free_pages_prepare(page, 0, false);
1063 static bool bulkfree_pcp_prepare(struct page *page)
1065 return free_pages_check(page);
1067 #endif /* CONFIG_DEBUG_VM */
1070 * Frees a number of pages from the PCP lists
1071 * Assumes all pages on list are in same zone, and of same order.
1072 * count is the number of pages to free.
1074 * If the zone was previously in an "all pages pinned" state then look to
1075 * see if this freeing clears that state.
1077 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1078 * pinned" detection logic.
1080 static void free_pcppages_bulk(struct zone *zone, int count,
1081 struct per_cpu_pages *pcp)
1083 int migratetype = 0;
1085 unsigned long nr_scanned;
1086 bool isolated_pageblocks;
1088 spin_lock(&zone->lock);
1089 isolated_pageblocks = has_isolate_pageblock(zone);
1090 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1092 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1096 struct list_head *list;
1099 * Remove pages from lists in a round-robin fashion. A
1100 * batch_free count is maintained that is incremented when an
1101 * empty list is encountered. This is so more pages are freed
1102 * off fuller lists instead of spinning excessively around empty
1107 if (++migratetype == MIGRATE_PCPTYPES)
1109 list = &pcp->lists[migratetype];
1110 } while (list_empty(list));
1112 /* This is the only non-empty list. Free them all. */
1113 if (batch_free == MIGRATE_PCPTYPES)
1117 int mt; /* migratetype of the to-be-freed page */
1119 page = list_last_entry(list, struct page, lru);
1120 /* must delete as __free_one_page list manipulates */
1121 list_del(&page->lru);
1123 mt = get_pcppage_migratetype(page);
1124 /* MIGRATE_ISOLATE page should not go to pcplists */
1125 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1126 /* Pageblock could have been isolated meanwhile */
1127 if (unlikely(isolated_pageblocks))
1128 mt = get_pageblock_migratetype(page);
1130 if (bulkfree_pcp_prepare(page))
1133 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1134 trace_mm_page_pcpu_drain(page, 0, mt);
1135 } while (--count && --batch_free && !list_empty(list));
1137 spin_unlock(&zone->lock);
1140 static void free_one_page(struct zone *zone,
1141 struct page *page, unsigned long pfn,
1145 unsigned long nr_scanned;
1146 spin_lock(&zone->lock);
1147 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1149 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1151 if (unlikely(has_isolate_pageblock(zone) ||
1152 is_migrate_isolate(migratetype))) {
1153 migratetype = get_pfnblock_migratetype(page, pfn);
1155 __free_one_page(page, pfn, zone, order, migratetype);
1156 spin_unlock(&zone->lock);
1159 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1160 unsigned long zone, int nid)
1162 set_page_links(page, zone, nid, pfn);
1163 init_page_count(page);
1164 page_mapcount_reset(page);
1165 page_cpupid_reset_last(page);
1167 INIT_LIST_HEAD(&page->lru);
1168 #ifdef WANT_PAGE_VIRTUAL
1169 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1170 if (!is_highmem_idx(zone))
1171 set_page_address(page, __va(pfn << PAGE_SHIFT));
1175 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1178 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1181 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1182 static void init_reserved_page(unsigned long pfn)
1187 if (!early_page_uninitialised(pfn))
1190 nid = early_pfn_to_nid(pfn);
1191 pgdat = NODE_DATA(nid);
1193 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1194 struct zone *zone = &pgdat->node_zones[zid];
1196 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1199 __init_single_pfn(pfn, zid, nid);
1202 static inline void init_reserved_page(unsigned long pfn)
1205 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1208 * Initialised pages do not have PageReserved set. This function is
1209 * called for each range allocated by the bootmem allocator and
1210 * marks the pages PageReserved. The remaining valid pages are later
1211 * sent to the buddy page allocator.
1213 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1215 unsigned long start_pfn = PFN_DOWN(start);
1216 unsigned long end_pfn = PFN_UP(end);
1218 for (; start_pfn < end_pfn; start_pfn++) {
1219 if (pfn_valid(start_pfn)) {
1220 struct page *page = pfn_to_page(start_pfn);
1222 init_reserved_page(start_pfn);
1224 /* Avoid false-positive PageTail() */
1225 INIT_LIST_HEAD(&page->lru);
1227 SetPageReserved(page);
1232 static void __free_pages_ok(struct page *page, unsigned int order)
1234 unsigned long flags;
1236 unsigned long pfn = page_to_pfn(page);
1238 if (!free_pages_prepare(page, order, true))
1241 migratetype = get_pfnblock_migratetype(page, pfn);
1242 local_irq_save(flags);
1243 __count_vm_events(PGFREE, 1 << order);
1244 free_one_page(page_zone(page), page, pfn, order, migratetype);
1245 local_irq_restore(flags);
1248 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1250 unsigned int nr_pages = 1 << order;
1251 struct page *p = page;
1255 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1257 __ClearPageReserved(p);
1258 set_page_count(p, 0);
1260 __ClearPageReserved(p);
1261 set_page_count(p, 0);
1263 page_zone(page)->managed_pages += nr_pages;
1264 set_page_refcounted(page);
1265 __free_pages(page, order);
1268 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1269 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1271 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1273 int __meminit early_pfn_to_nid(unsigned long pfn)
1275 static DEFINE_SPINLOCK(early_pfn_lock);
1278 spin_lock(&early_pfn_lock);
1279 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1281 nid = first_online_node;
1282 spin_unlock(&early_pfn_lock);
1288 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1289 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1290 struct mminit_pfnnid_cache *state)
1294 nid = __early_pfn_to_nid(pfn, state);
1295 if (nid >= 0 && nid != node)
1300 /* Only safe to use early in boot when initialisation is single-threaded */
1301 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1303 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1308 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1312 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1313 struct mminit_pfnnid_cache *state)
1320 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1323 if (early_page_uninitialised(pfn))
1325 return __free_pages_boot_core(page, order);
1329 * Check that the whole (or subset of) a pageblock given by the interval of
1330 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1331 * with the migration of free compaction scanner. The scanners then need to
1332 * use only pfn_valid_within() check for arches that allow holes within
1335 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1337 * It's possible on some configurations to have a setup like node0 node1 node0
1338 * i.e. it's possible that all pages within a zones range of pages do not
1339 * belong to a single zone. We assume that a border between node0 and node1
1340 * can occur within a single pageblock, but not a node0 node1 node0
1341 * interleaving within a single pageblock. It is therefore sufficient to check
1342 * the first and last page of a pageblock and avoid checking each individual
1343 * page in a pageblock.
1345 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1346 unsigned long end_pfn, struct zone *zone)
1348 struct page *start_page;
1349 struct page *end_page;
1351 /* end_pfn is one past the range we are checking */
1354 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1357 start_page = pfn_to_page(start_pfn);
1359 if (page_zone(start_page) != zone)
1362 end_page = pfn_to_page(end_pfn);
1364 /* This gives a shorter code than deriving page_zone(end_page) */
1365 if (page_zone_id(start_page) != page_zone_id(end_page))
1371 void set_zone_contiguous(struct zone *zone)
1373 unsigned long block_start_pfn = zone->zone_start_pfn;
1374 unsigned long block_end_pfn;
1376 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1377 for (; block_start_pfn < zone_end_pfn(zone);
1378 block_start_pfn = block_end_pfn,
1379 block_end_pfn += pageblock_nr_pages) {
1381 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1383 if (!__pageblock_pfn_to_page(block_start_pfn,
1384 block_end_pfn, zone))
1388 /* We confirm that there is no hole */
1389 zone->contiguous = true;
1392 void clear_zone_contiguous(struct zone *zone)
1394 zone->contiguous = false;
1397 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1398 static void __init deferred_free_range(struct page *page,
1399 unsigned long pfn, int nr_pages)
1406 /* Free a large naturally-aligned chunk if possible */
1407 if (nr_pages == pageblock_nr_pages &&
1408 (pfn & (pageblock_nr_pages - 1)) == 0) {
1409 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1410 __free_pages_boot_core(page, pageblock_order);
1414 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1415 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1416 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1417 __free_pages_boot_core(page, 0);
1421 /* Completion tracking for deferred_init_memmap() threads */
1422 static atomic_t pgdat_init_n_undone __initdata;
1423 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1425 static inline void __init pgdat_init_report_one_done(void)
1427 if (atomic_dec_and_test(&pgdat_init_n_undone))
1428 complete(&pgdat_init_all_done_comp);
1431 /* Initialise remaining memory on a node */
1432 static int __init deferred_init_memmap(void *data)
1434 pg_data_t *pgdat = data;
1435 int nid = pgdat->node_id;
1436 struct mminit_pfnnid_cache nid_init_state = { };
1437 unsigned long start = jiffies;
1438 unsigned long nr_pages = 0;
1439 unsigned long walk_start, walk_end;
1442 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1443 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1445 if (first_init_pfn == ULONG_MAX) {
1446 pgdat_init_report_one_done();
1450 /* Bind memory initialisation thread to a local node if possible */
1451 if (!cpumask_empty(cpumask))
1452 set_cpus_allowed_ptr(current, cpumask);
1454 /* Sanity check boundaries */
1455 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1456 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1457 pgdat->first_deferred_pfn = ULONG_MAX;
1459 /* Only the highest zone is deferred so find it */
1460 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1461 zone = pgdat->node_zones + zid;
1462 if (first_init_pfn < zone_end_pfn(zone))
1466 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1467 unsigned long pfn, end_pfn;
1468 struct page *page = NULL;
1469 struct page *free_base_page = NULL;
1470 unsigned long free_base_pfn = 0;
1473 end_pfn = min(walk_end, zone_end_pfn(zone));
1474 pfn = first_init_pfn;
1475 if (pfn < walk_start)
1477 if (pfn < zone->zone_start_pfn)
1478 pfn = zone->zone_start_pfn;
1480 for (; pfn < end_pfn; pfn++) {
1481 if (!pfn_valid_within(pfn))
1485 * Ensure pfn_valid is checked every
1486 * pageblock_nr_pages for memory holes
1488 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1489 if (!pfn_valid(pfn)) {
1495 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1500 /* Minimise pfn page lookups and scheduler checks */
1501 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1504 nr_pages += nr_to_free;
1505 deferred_free_range(free_base_page,
1506 free_base_pfn, nr_to_free);
1507 free_base_page = NULL;
1508 free_base_pfn = nr_to_free = 0;
1510 page = pfn_to_page(pfn);
1515 VM_BUG_ON(page_zone(page) != zone);
1519 __init_single_page(page, pfn, zid, nid);
1520 if (!free_base_page) {
1521 free_base_page = page;
1522 free_base_pfn = pfn;
1527 /* Where possible, batch up pages for a single free */
1530 /* Free the current block of pages to allocator */
1531 nr_pages += nr_to_free;
1532 deferred_free_range(free_base_page, free_base_pfn,
1534 free_base_page = NULL;
1535 free_base_pfn = nr_to_free = 0;
1537 /* Free the last block of pages to allocator */
1538 nr_pages += nr_to_free;
1539 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1541 first_init_pfn = max(end_pfn, first_init_pfn);
1544 /* Sanity check that the next zone really is unpopulated */
1545 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1547 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1548 jiffies_to_msecs(jiffies - start));
1550 pgdat_init_report_one_done();
1553 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1555 void __init page_alloc_init_late(void)
1559 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1562 /* There will be num_node_state(N_MEMORY) threads */
1563 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1564 for_each_node_state(nid, N_MEMORY) {
1565 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1568 /* Block until all are initialised */
1569 wait_for_completion(&pgdat_init_all_done_comp);
1571 /* Reinit limits that are based on free pages after the kernel is up */
1572 files_maxfiles_init();
1575 for_each_populated_zone(zone)
1576 set_zone_contiguous(zone);
1580 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1581 void __init init_cma_reserved_pageblock(struct page *page)
1583 unsigned i = pageblock_nr_pages;
1584 struct page *p = page;
1587 __ClearPageReserved(p);
1588 set_page_count(p, 0);
1591 set_pageblock_migratetype(page, MIGRATE_CMA);
1593 if (pageblock_order >= MAX_ORDER) {
1594 i = pageblock_nr_pages;
1597 set_page_refcounted(p);
1598 __free_pages(p, MAX_ORDER - 1);
1599 p += MAX_ORDER_NR_PAGES;
1600 } while (i -= MAX_ORDER_NR_PAGES);
1602 set_page_refcounted(page);
1603 __free_pages(page, pageblock_order);
1606 adjust_managed_page_count(page, pageblock_nr_pages);
1611 * The order of subdivision here is critical for the IO subsystem.
1612 * Please do not alter this order without good reasons and regression
1613 * testing. Specifically, as large blocks of memory are subdivided,
1614 * the order in which smaller blocks are delivered depends on the order
1615 * they're subdivided in this function. This is the primary factor
1616 * influencing the order in which pages are delivered to the IO
1617 * subsystem according to empirical testing, and this is also justified
1618 * by considering the behavior of a buddy system containing a single
1619 * large block of memory acted on by a series of small allocations.
1620 * This behavior is a critical factor in sglist merging's success.
1624 static inline void expand(struct zone *zone, struct page *page,
1625 int low, int high, struct free_area *area,
1628 unsigned long size = 1 << high;
1630 while (high > low) {
1634 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1637 * Mark as guard pages (or page), that will allow to
1638 * merge back to allocator when buddy will be freed.
1639 * Corresponding page table entries will not be touched,
1640 * pages will stay not present in virtual address space
1642 if (set_page_guard(zone, &page[size], high, migratetype))
1645 list_add(&page[size].lru, &area->free_list[migratetype]);
1647 set_page_order(&page[size], high);
1651 static void check_new_page_bad(struct page *page)
1653 const char *bad_reason = NULL;
1654 unsigned long bad_flags = 0;
1656 if (unlikely(atomic_read(&page->_mapcount) != -1))
1657 bad_reason = "nonzero mapcount";
1658 if (unlikely(page->mapping != NULL))
1659 bad_reason = "non-NULL mapping";
1660 if (unlikely(page_ref_count(page) != 0))
1661 bad_reason = "nonzero _count";
1662 if (unlikely(page->flags & __PG_HWPOISON)) {
1663 bad_reason = "HWPoisoned (hardware-corrupted)";
1664 bad_flags = __PG_HWPOISON;
1665 /* Don't complain about hwpoisoned pages */
1666 page_mapcount_reset(page); /* remove PageBuddy */
1669 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1670 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1671 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1674 if (unlikely(page->mem_cgroup))
1675 bad_reason = "page still charged to cgroup";
1677 bad_page(page, bad_reason, bad_flags);
1681 * This page is about to be returned from the page allocator
1683 static inline int check_new_page(struct page *page)
1685 if (likely(page_expected_state(page,
1686 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1689 check_new_page_bad(page);
1693 static inline bool free_pages_prezeroed(bool poisoned)
1695 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1696 page_poisoning_enabled() && poisoned;
1699 #ifdef CONFIG_DEBUG_VM
1700 static bool check_pcp_refill(struct page *page)
1705 static bool check_new_pcp(struct page *page)
1707 return check_new_page(page);
1710 static bool check_pcp_refill(struct page *page)
1712 return check_new_page(page);
1714 static bool check_new_pcp(struct page *page)
1718 #endif /* CONFIG_DEBUG_VM */
1720 static bool check_new_pages(struct page *page, unsigned int order)
1723 for (i = 0; i < (1 << order); i++) {
1724 struct page *p = page + i;
1726 if (unlikely(check_new_page(p)))
1733 inline void post_alloc_hook(struct page *page, unsigned int order,
1736 set_page_private(page, 0);
1737 set_page_refcounted(page);
1739 arch_alloc_page(page, order);
1740 kernel_map_pages(page, 1 << order, 1);
1741 kernel_poison_pages(page, 1 << order, 1);
1742 kasan_alloc_pages(page, order);
1743 set_page_owner(page, order, gfp_flags);
1746 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1747 unsigned int alloc_flags)
1750 bool poisoned = true;
1752 for (i = 0; i < (1 << order); i++) {
1753 struct page *p = page + i;
1755 poisoned &= page_is_poisoned(p);
1758 post_alloc_hook(page, order, gfp_flags);
1760 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1761 for (i = 0; i < (1 << order); i++)
1762 clear_highpage(page + i);
1764 if (order && (gfp_flags & __GFP_COMP))
1765 prep_compound_page(page, order);
1768 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1769 * allocate the page. The expectation is that the caller is taking
1770 * steps that will free more memory. The caller should avoid the page
1771 * being used for !PFMEMALLOC purposes.
1773 if (alloc_flags & ALLOC_NO_WATERMARKS)
1774 set_page_pfmemalloc(page);
1776 clear_page_pfmemalloc(page);
1780 * Go through the free lists for the given migratetype and remove
1781 * the smallest available page from the freelists
1784 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1787 unsigned int current_order;
1788 struct free_area *area;
1791 /* Find a page of the appropriate size in the preferred list */
1792 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1793 area = &(zone->free_area[current_order]);
1794 page = list_first_entry_or_null(&area->free_list[migratetype],
1798 list_del(&page->lru);
1799 rmv_page_order(page);
1801 expand(zone, page, order, current_order, area, migratetype);
1802 set_pcppage_migratetype(page, migratetype);
1811 * This array describes the order lists are fallen back to when
1812 * the free lists for the desirable migrate type are depleted
1814 static int fallbacks[MIGRATE_TYPES][4] = {
1815 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1816 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1817 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1819 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1821 #ifdef CONFIG_MEMORY_ISOLATION
1822 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1827 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1830 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1833 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1834 unsigned int order) { return NULL; }
1838 * Move the free pages in a range to the free lists of the requested type.
1839 * Note that start_page and end_pages are not aligned on a pageblock
1840 * boundary. If alignment is required, use move_freepages_block()
1842 int move_freepages(struct zone *zone,
1843 struct page *start_page, struct page *end_page,
1848 int pages_moved = 0;
1850 #ifndef CONFIG_HOLES_IN_ZONE
1852 * page_zone is not safe to call in this context when
1853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1854 * anyway as we check zone boundaries in move_freepages_block().
1855 * Remove at a later date when no bug reports exist related to
1856 * grouping pages by mobility
1858 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1861 for (page = start_page; page <= end_page;) {
1862 /* Make sure we are not inadvertently changing nodes */
1863 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1865 if (!pfn_valid_within(page_to_pfn(page))) {
1870 if (!PageBuddy(page)) {
1875 order = page_order(page);
1876 list_move(&page->lru,
1877 &zone->free_area[order].free_list[migratetype]);
1879 pages_moved += 1 << order;
1885 int move_freepages_block(struct zone *zone, struct page *page,
1888 unsigned long start_pfn, end_pfn;
1889 struct page *start_page, *end_page;
1891 start_pfn = page_to_pfn(page);
1892 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1893 start_page = pfn_to_page(start_pfn);
1894 end_page = start_page + pageblock_nr_pages - 1;
1895 end_pfn = start_pfn + pageblock_nr_pages - 1;
1897 /* Do not cross zone boundaries */
1898 if (!zone_spans_pfn(zone, start_pfn))
1900 if (!zone_spans_pfn(zone, end_pfn))
1903 return move_freepages(zone, start_page, end_page, migratetype);
1906 static void change_pageblock_range(struct page *pageblock_page,
1907 int start_order, int migratetype)
1909 int nr_pageblocks = 1 << (start_order - pageblock_order);
1911 while (nr_pageblocks--) {
1912 set_pageblock_migratetype(pageblock_page, migratetype);
1913 pageblock_page += pageblock_nr_pages;
1918 * When we are falling back to another migratetype during allocation, try to
1919 * steal extra free pages from the same pageblocks to satisfy further
1920 * allocations, instead of polluting multiple pageblocks.
1922 * If we are stealing a relatively large buddy page, it is likely there will
1923 * be more free pages in the pageblock, so try to steal them all. For
1924 * reclaimable and unmovable allocations, we steal regardless of page size,
1925 * as fragmentation caused by those allocations polluting movable pageblocks
1926 * is worse than movable allocations stealing from unmovable and reclaimable
1929 static bool can_steal_fallback(unsigned int order, int start_mt)
1932 * Leaving this order check is intended, although there is
1933 * relaxed order check in next check. The reason is that
1934 * we can actually steal whole pageblock if this condition met,
1935 * but, below check doesn't guarantee it and that is just heuristic
1936 * so could be changed anytime.
1938 if (order >= pageblock_order)
1941 if (order >= pageblock_order / 2 ||
1942 start_mt == MIGRATE_RECLAIMABLE ||
1943 start_mt == MIGRATE_UNMOVABLE ||
1944 page_group_by_mobility_disabled)
1951 * This function implements actual steal behaviour. If order is large enough,
1952 * we can steal whole pageblock. If not, we first move freepages in this
1953 * pageblock and check whether half of pages are moved or not. If half of
1954 * pages are moved, we can change migratetype of pageblock and permanently
1955 * use it's pages as requested migratetype in the future.
1957 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1960 unsigned int current_order = page_order(page);
1963 /* Take ownership for orders >= pageblock_order */
1964 if (current_order >= pageblock_order) {
1965 change_pageblock_range(page, current_order, start_type);
1969 pages = move_freepages_block(zone, page, start_type);
1971 /* Claim the whole block if over half of it is free */
1972 if (pages >= (1 << (pageblock_order-1)) ||
1973 page_group_by_mobility_disabled)
1974 set_pageblock_migratetype(page, start_type);
1978 * Check whether there is a suitable fallback freepage with requested order.
1979 * If only_stealable is true, this function returns fallback_mt only if
1980 * we can steal other freepages all together. This would help to reduce
1981 * fragmentation due to mixed migratetype pages in one pageblock.
1983 int find_suitable_fallback(struct free_area *area, unsigned int order,
1984 int migratetype, bool only_stealable, bool *can_steal)
1989 if (area->nr_free == 0)
1994 fallback_mt = fallbacks[migratetype][i];
1995 if (fallback_mt == MIGRATE_TYPES)
1998 if (list_empty(&area->free_list[fallback_mt]))
2001 if (can_steal_fallback(order, migratetype))
2004 if (!only_stealable)
2015 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2016 * there are no empty page blocks that contain a page with a suitable order
2018 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2019 unsigned int alloc_order)
2022 unsigned long max_managed, flags;
2025 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2026 * Check is race-prone but harmless.
2028 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2029 if (zone->nr_reserved_highatomic >= max_managed)
2032 spin_lock_irqsave(&zone->lock, flags);
2034 /* Recheck the nr_reserved_highatomic limit under the lock */
2035 if (zone->nr_reserved_highatomic >= max_managed)
2039 mt = get_pageblock_migratetype(page);
2040 if (mt != MIGRATE_HIGHATOMIC &&
2041 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2042 zone->nr_reserved_highatomic += pageblock_nr_pages;
2043 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2044 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2048 spin_unlock_irqrestore(&zone->lock, flags);
2052 * Used when an allocation is about to fail under memory pressure. This
2053 * potentially hurts the reliability of high-order allocations when under
2054 * intense memory pressure but failed atomic allocations should be easier
2055 * to recover from than an OOM.
2057 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2059 struct zonelist *zonelist = ac->zonelist;
2060 unsigned long flags;
2066 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2068 /* Preserve at least one pageblock */
2069 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2072 spin_lock_irqsave(&zone->lock, flags);
2073 for (order = 0; order < MAX_ORDER; order++) {
2074 struct free_area *area = &(zone->free_area[order]);
2076 page = list_first_entry_or_null(
2077 &area->free_list[MIGRATE_HIGHATOMIC],
2083 * It should never happen but changes to locking could
2084 * inadvertently allow a per-cpu drain to add pages
2085 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2086 * and watch for underflows.
2088 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2089 zone->nr_reserved_highatomic);
2092 * Convert to ac->migratetype and avoid the normal
2093 * pageblock stealing heuristics. Minimally, the caller
2094 * is doing the work and needs the pages. More
2095 * importantly, if the block was always converted to
2096 * MIGRATE_UNMOVABLE or another type then the number
2097 * of pageblocks that cannot be completely freed
2100 set_pageblock_migratetype(page, ac->migratetype);
2101 move_freepages_block(zone, page, ac->migratetype);
2102 spin_unlock_irqrestore(&zone->lock, flags);
2105 spin_unlock_irqrestore(&zone->lock, flags);
2109 /* Remove an element from the buddy allocator from the fallback list */
2110 static inline struct page *
2111 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2113 struct free_area *area;
2114 unsigned int current_order;
2119 /* Find the largest possible block of pages in the other list */
2120 for (current_order = MAX_ORDER-1;
2121 current_order >= order && current_order <= MAX_ORDER-1;
2123 area = &(zone->free_area[current_order]);
2124 fallback_mt = find_suitable_fallback(area, current_order,
2125 start_migratetype, false, &can_steal);
2126 if (fallback_mt == -1)
2129 page = list_first_entry(&area->free_list[fallback_mt],
2132 steal_suitable_fallback(zone, page, start_migratetype);
2134 /* Remove the page from the freelists */
2136 list_del(&page->lru);
2137 rmv_page_order(page);
2139 expand(zone, page, order, current_order, area,
2142 * The pcppage_migratetype may differ from pageblock's
2143 * migratetype depending on the decisions in
2144 * find_suitable_fallback(). This is OK as long as it does not
2145 * differ for MIGRATE_CMA pageblocks. Those can be used as
2146 * fallback only via special __rmqueue_cma_fallback() function
2148 set_pcppage_migratetype(page, start_migratetype);
2150 trace_mm_page_alloc_extfrag(page, order, current_order,
2151 start_migratetype, fallback_mt);
2160 * Do the hard work of removing an element from the buddy allocator.
2161 * Call me with the zone->lock already held.
2163 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2168 page = __rmqueue_smallest(zone, order, migratetype);
2169 if (unlikely(!page)) {
2170 if (migratetype == MIGRATE_MOVABLE)
2171 page = __rmqueue_cma_fallback(zone, order);
2174 page = __rmqueue_fallback(zone, order, migratetype);
2177 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2182 * Obtain a specified number of elements from the buddy allocator, all under
2183 * a single hold of the lock, for efficiency. Add them to the supplied list.
2184 * Returns the number of new pages which were placed at *list.
2186 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2187 unsigned long count, struct list_head *list,
2188 int migratetype, bool cold)
2192 spin_lock(&zone->lock);
2193 for (i = 0; i < count; ++i) {
2194 struct page *page = __rmqueue(zone, order, migratetype);
2195 if (unlikely(page == NULL))
2198 if (unlikely(check_pcp_refill(page)))
2202 * Split buddy pages returned by expand() are received here
2203 * in physical page order. The page is added to the callers and
2204 * list and the list head then moves forward. From the callers
2205 * perspective, the linked list is ordered by page number in
2206 * some conditions. This is useful for IO devices that can
2207 * merge IO requests if the physical pages are ordered
2211 list_add(&page->lru, list);
2213 list_add_tail(&page->lru, list);
2215 if (is_migrate_cma(get_pcppage_migratetype(page)))
2216 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2219 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2220 spin_unlock(&zone->lock);
2226 * Called from the vmstat counter updater to drain pagesets of this
2227 * currently executing processor on remote nodes after they have
2230 * Note that this function must be called with the thread pinned to
2231 * a single processor.
2233 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2235 unsigned long flags;
2236 int to_drain, batch;
2238 local_irq_save(flags);
2239 batch = READ_ONCE(pcp->batch);
2240 to_drain = min(pcp->count, batch);
2242 free_pcppages_bulk(zone, to_drain, pcp);
2243 pcp->count -= to_drain;
2245 local_irq_restore(flags);
2250 * Drain pcplists of the indicated processor and zone.
2252 * The processor must either be the current processor and the
2253 * thread pinned to the current processor or a processor that
2256 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2258 unsigned long flags;
2259 struct per_cpu_pageset *pset;
2260 struct per_cpu_pages *pcp;
2262 local_irq_save(flags);
2263 pset = per_cpu_ptr(zone->pageset, cpu);
2267 free_pcppages_bulk(zone, pcp->count, pcp);
2270 local_irq_restore(flags);
2274 * Drain pcplists of all zones on the indicated processor.
2276 * The processor must either be the current processor and the
2277 * thread pinned to the current processor or a processor that
2280 static void drain_pages(unsigned int cpu)
2284 for_each_populated_zone(zone) {
2285 drain_pages_zone(cpu, zone);
2290 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2292 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2293 * the single zone's pages.
2295 void drain_local_pages(struct zone *zone)
2297 int cpu = smp_processor_id();
2300 drain_pages_zone(cpu, zone);
2306 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2308 * When zone parameter is non-NULL, spill just the single zone's pages.
2310 * Note that this code is protected against sending an IPI to an offline
2311 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2312 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2313 * nothing keeps CPUs from showing up after we populated the cpumask and
2314 * before the call to on_each_cpu_mask().
2316 void drain_all_pages(struct zone *zone)
2321 * Allocate in the BSS so we wont require allocation in
2322 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2324 static cpumask_t cpus_with_pcps;
2327 * We don't care about racing with CPU hotplug event
2328 * as offline notification will cause the notified
2329 * cpu to drain that CPU pcps and on_each_cpu_mask
2330 * disables preemption as part of its processing
2332 for_each_online_cpu(cpu) {
2333 struct per_cpu_pageset *pcp;
2335 bool has_pcps = false;
2338 pcp = per_cpu_ptr(zone->pageset, cpu);
2342 for_each_populated_zone(z) {
2343 pcp = per_cpu_ptr(z->pageset, cpu);
2344 if (pcp->pcp.count) {
2352 cpumask_set_cpu(cpu, &cpus_with_pcps);
2354 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2356 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2360 #ifdef CONFIG_HIBERNATION
2362 void mark_free_pages(struct zone *zone)
2364 unsigned long pfn, max_zone_pfn;
2365 unsigned long flags;
2366 unsigned int order, t;
2369 if (zone_is_empty(zone))
2372 spin_lock_irqsave(&zone->lock, flags);
2374 max_zone_pfn = zone_end_pfn(zone);
2375 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2376 if (pfn_valid(pfn)) {
2377 page = pfn_to_page(pfn);
2379 if (page_zone(page) != zone)
2382 if (!swsusp_page_is_forbidden(page))
2383 swsusp_unset_page_free(page);
2386 for_each_migratetype_order(order, t) {
2387 list_for_each_entry(page,
2388 &zone->free_area[order].free_list[t], lru) {
2391 pfn = page_to_pfn(page);
2392 for (i = 0; i < (1UL << order); i++)
2393 swsusp_set_page_free(pfn_to_page(pfn + i));
2396 spin_unlock_irqrestore(&zone->lock, flags);
2398 #endif /* CONFIG_PM */
2401 * Free a 0-order page
2402 * cold == true ? free a cold page : free a hot page
2404 void free_hot_cold_page(struct page *page, bool cold)
2406 struct zone *zone = page_zone(page);
2407 struct per_cpu_pages *pcp;
2408 unsigned long flags;
2409 unsigned long pfn = page_to_pfn(page);
2412 if (!free_pcp_prepare(page))
2415 migratetype = get_pfnblock_migratetype(page, pfn);
2416 set_pcppage_migratetype(page, migratetype);
2417 local_irq_save(flags);
2418 __count_vm_event(PGFREE);
2421 * We only track unmovable, reclaimable and movable on pcp lists.
2422 * Free ISOLATE pages back to the allocator because they are being
2423 * offlined but treat RESERVE as movable pages so we can get those
2424 * areas back if necessary. Otherwise, we may have to free
2425 * excessively into the page allocator
2427 if (migratetype >= MIGRATE_PCPTYPES) {
2428 if (unlikely(is_migrate_isolate(migratetype))) {
2429 free_one_page(zone, page, pfn, 0, migratetype);
2432 migratetype = MIGRATE_MOVABLE;
2435 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2437 list_add(&page->lru, &pcp->lists[migratetype]);
2439 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2441 if (pcp->count >= pcp->high) {
2442 unsigned long batch = READ_ONCE(pcp->batch);
2443 free_pcppages_bulk(zone, batch, pcp);
2444 pcp->count -= batch;
2448 local_irq_restore(flags);
2452 * Free a list of 0-order pages
2454 void free_hot_cold_page_list(struct list_head *list, bool cold)
2456 struct page *page, *next;
2458 list_for_each_entry_safe(page, next, list, lru) {
2459 trace_mm_page_free_batched(page, cold);
2460 free_hot_cold_page(page, cold);
2465 * split_page takes a non-compound higher-order page, and splits it into
2466 * n (1<<order) sub-pages: page[0..n]
2467 * Each sub-page must be freed individually.
2469 * Note: this is probably too low level an operation for use in drivers.
2470 * Please consult with lkml before using this in your driver.
2472 void split_page(struct page *page, unsigned int order)
2476 VM_BUG_ON_PAGE(PageCompound(page), page);
2477 VM_BUG_ON_PAGE(!page_count(page), page);
2479 #ifdef CONFIG_KMEMCHECK
2481 * Split shadow pages too, because free(page[0]) would
2482 * otherwise free the whole shadow.
2484 if (kmemcheck_page_is_tracked(page))
2485 split_page(virt_to_page(page[0].shadow), order);
2488 for (i = 1; i < (1 << order); i++)
2489 set_page_refcounted(page + i);
2490 split_page_owner(page, order);
2492 EXPORT_SYMBOL_GPL(split_page);
2494 int __isolate_free_page(struct page *page, unsigned int order)
2496 unsigned long watermark;
2500 BUG_ON(!PageBuddy(page));
2502 zone = page_zone(page);
2503 mt = get_pageblock_migratetype(page);
2505 if (!is_migrate_isolate(mt)) {
2507 * Obey watermarks as if the page was being allocated. We can
2508 * emulate a high-order watermark check with a raised order-0
2509 * watermark, because we already know our high-order page
2512 watermark = min_wmark_pages(zone) + (1UL << order);
2513 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2516 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2519 /* Remove page from free list */
2520 list_del(&page->lru);
2521 zone->free_area[order].nr_free--;
2522 rmv_page_order(page);
2525 * Set the pageblock if the isolated page is at least half of a
2528 if (order >= pageblock_order - 1) {
2529 struct page *endpage = page + (1 << order) - 1;
2530 for (; page < endpage; page += pageblock_nr_pages) {
2531 int mt = get_pageblock_migratetype(page);
2532 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2533 set_pageblock_migratetype(page,
2539 return 1UL << order;
2543 * Update NUMA hit/miss statistics
2545 * Must be called with interrupts disabled.
2547 * When __GFP_OTHER_NODE is set assume the node of the preferred
2548 * zone is the local node. This is useful for daemons who allocate
2549 * memory on behalf of other processes.
2551 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2555 int local_nid = numa_node_id();
2556 enum zone_stat_item local_stat = NUMA_LOCAL;
2558 if (unlikely(flags & __GFP_OTHER_NODE)) {
2559 local_stat = NUMA_OTHER;
2560 local_nid = preferred_zone->node;
2563 if (z->node == local_nid) {
2564 __inc_zone_state(z, NUMA_HIT);
2565 __inc_zone_state(z, local_stat);
2567 __inc_zone_state(z, NUMA_MISS);
2568 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2574 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2577 struct page *buffered_rmqueue(struct zone *preferred_zone,
2578 struct zone *zone, unsigned int order,
2579 gfp_t gfp_flags, unsigned int alloc_flags,
2582 unsigned long flags;
2584 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2586 if (likely(order == 0)) {
2587 struct per_cpu_pages *pcp;
2588 struct list_head *list;
2590 local_irq_save(flags);
2592 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2593 list = &pcp->lists[migratetype];
2594 if (list_empty(list)) {
2595 pcp->count += rmqueue_bulk(zone, 0,
2598 if (unlikely(list_empty(list)))
2603 page = list_last_entry(list, struct page, lru);
2605 page = list_first_entry(list, struct page, lru);
2607 list_del(&page->lru);
2610 } while (check_new_pcp(page));
2613 * We most definitely don't want callers attempting to
2614 * allocate greater than order-1 page units with __GFP_NOFAIL.
2616 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2617 spin_lock_irqsave(&zone->lock, flags);
2621 if (alloc_flags & ALLOC_HARDER) {
2622 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2624 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2627 page = __rmqueue(zone, order, migratetype);
2628 } while (page && check_new_pages(page, order));
2629 spin_unlock(&zone->lock);
2632 __mod_zone_freepage_state(zone, -(1 << order),
2633 get_pcppage_migratetype(page));
2636 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2637 zone_statistics(preferred_zone, zone, gfp_flags);
2638 local_irq_restore(flags);
2640 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2644 local_irq_restore(flags);
2648 #ifdef CONFIG_FAIL_PAGE_ALLOC
2651 struct fault_attr attr;
2653 bool ignore_gfp_highmem;
2654 bool ignore_gfp_reclaim;
2656 } fail_page_alloc = {
2657 .attr = FAULT_ATTR_INITIALIZER,
2658 .ignore_gfp_reclaim = true,
2659 .ignore_gfp_highmem = true,
2663 static int __init setup_fail_page_alloc(char *str)
2665 return setup_fault_attr(&fail_page_alloc.attr, str);
2667 __setup("fail_page_alloc=", setup_fail_page_alloc);
2669 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2671 if (order < fail_page_alloc.min_order)
2673 if (gfp_mask & __GFP_NOFAIL)
2675 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2677 if (fail_page_alloc.ignore_gfp_reclaim &&
2678 (gfp_mask & __GFP_DIRECT_RECLAIM))
2681 return should_fail(&fail_page_alloc.attr, 1 << order);
2684 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2686 static int __init fail_page_alloc_debugfs(void)
2688 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2691 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2692 &fail_page_alloc.attr);
2694 return PTR_ERR(dir);
2696 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2697 &fail_page_alloc.ignore_gfp_reclaim))
2699 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2700 &fail_page_alloc.ignore_gfp_highmem))
2702 if (!debugfs_create_u32("min-order", mode, dir,
2703 &fail_page_alloc.min_order))
2708 debugfs_remove_recursive(dir);
2713 late_initcall(fail_page_alloc_debugfs);
2715 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2717 #else /* CONFIG_FAIL_PAGE_ALLOC */
2719 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2724 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2727 * Return true if free base pages are above 'mark'. For high-order checks it
2728 * will return true of the order-0 watermark is reached and there is at least
2729 * one free page of a suitable size. Checking now avoids taking the zone lock
2730 * to check in the allocation paths if no pages are free.
2732 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2733 int classzone_idx, unsigned int alloc_flags,
2738 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2740 /* free_pages may go negative - that's OK */
2741 free_pages -= (1 << order) - 1;
2743 if (alloc_flags & ALLOC_HIGH)
2747 * If the caller does not have rights to ALLOC_HARDER then subtract
2748 * the high-atomic reserves. This will over-estimate the size of the
2749 * atomic reserve but it avoids a search.
2751 if (likely(!alloc_harder))
2752 free_pages -= z->nr_reserved_highatomic;
2757 /* If allocation can't use CMA areas don't use free CMA pages */
2758 if (!(alloc_flags & ALLOC_CMA))
2759 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2763 * Check watermarks for an order-0 allocation request. If these
2764 * are not met, then a high-order request also cannot go ahead
2765 * even if a suitable page happened to be free.
2767 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2770 /* If this is an order-0 request then the watermark is fine */
2774 /* For a high-order request, check at least one suitable page is free */
2775 for (o = order; o < MAX_ORDER; o++) {
2776 struct free_area *area = &z->free_area[o];
2785 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2786 if (!list_empty(&area->free_list[mt]))
2791 if ((alloc_flags & ALLOC_CMA) &&
2792 !list_empty(&area->free_list[MIGRATE_CMA])) {
2800 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2801 int classzone_idx, unsigned int alloc_flags)
2803 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2804 zone_page_state(z, NR_FREE_PAGES));
2807 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2808 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2810 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2814 /* If allocation can't use CMA areas don't use free CMA pages */
2815 if (!(alloc_flags & ALLOC_CMA))
2816 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2820 * Fast check for order-0 only. If this fails then the reserves
2821 * need to be calculated. There is a corner case where the check
2822 * passes but only the high-order atomic reserve are free. If
2823 * the caller is !atomic then it'll uselessly search the free
2824 * list. That corner case is then slower but it is harmless.
2826 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2829 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2833 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2834 unsigned long mark, int classzone_idx)
2836 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2838 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2839 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2841 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2846 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2848 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2851 #else /* CONFIG_NUMA */
2852 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2856 #endif /* CONFIG_NUMA */
2859 * get_page_from_freelist goes through the zonelist trying to allocate
2862 static struct page *
2863 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2864 const struct alloc_context *ac)
2866 struct zoneref *z = ac->preferred_zoneref;
2868 struct pglist_data *last_pgdat_dirty_limit = NULL;
2871 * Scan zonelist, looking for a zone with enough free.
2872 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2874 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2879 if (cpusets_enabled() &&
2880 (alloc_flags & ALLOC_CPUSET) &&
2881 !__cpuset_zone_allowed(zone, gfp_mask))
2884 * When allocating a page cache page for writing, we
2885 * want to get it from a node that is within its dirty
2886 * limit, such that no single node holds more than its
2887 * proportional share of globally allowed dirty pages.
2888 * The dirty limits take into account the node's
2889 * lowmem reserves and high watermark so that kswapd
2890 * should be able to balance it without having to
2891 * write pages from its LRU list.
2893 * XXX: For now, allow allocations to potentially
2894 * exceed the per-node dirty limit in the slowpath
2895 * (spread_dirty_pages unset) before going into reclaim,
2896 * which is important when on a NUMA setup the allowed
2897 * nodes are together not big enough to reach the
2898 * global limit. The proper fix for these situations
2899 * will require awareness of nodes in the
2900 * dirty-throttling and the flusher threads.
2902 if (ac->spread_dirty_pages) {
2903 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2906 if (!node_dirty_ok(zone->zone_pgdat)) {
2907 last_pgdat_dirty_limit = zone->zone_pgdat;
2912 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2913 if (!zone_watermark_fast(zone, order, mark,
2914 ac_classzone_idx(ac), alloc_flags)) {
2917 /* Checked here to keep the fast path fast */
2918 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2919 if (alloc_flags & ALLOC_NO_WATERMARKS)
2922 if (node_reclaim_mode == 0 ||
2923 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2926 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2928 case NODE_RECLAIM_NOSCAN:
2931 case NODE_RECLAIM_FULL:
2932 /* scanned but unreclaimable */
2935 /* did we reclaim enough */
2936 if (zone_watermark_ok(zone, order, mark,
2937 ac_classzone_idx(ac), alloc_flags))
2945 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2946 gfp_mask, alloc_flags, ac->migratetype);
2948 prep_new_page(page, order, gfp_mask, alloc_flags);
2951 * If this is a high-order atomic allocation then check
2952 * if the pageblock should be reserved for the future
2954 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2955 reserve_highatomic_pageblock(page, zone, order);
2965 * Large machines with many possible nodes should not always dump per-node
2966 * meminfo in irq context.
2968 static inline bool should_suppress_show_mem(void)
2973 ret = in_interrupt();
2978 static DEFINE_RATELIMIT_STATE(nopage_rs,
2979 DEFAULT_RATELIMIT_INTERVAL,
2980 DEFAULT_RATELIMIT_BURST);
2982 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2984 unsigned int filter = SHOW_MEM_FILTER_NODES;
2986 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2987 debug_guardpage_minorder() > 0)
2991 * This documents exceptions given to allocations in certain
2992 * contexts that are allowed to allocate outside current's set
2995 if (!(gfp_mask & __GFP_NOMEMALLOC))
2996 if (test_thread_flag(TIF_MEMDIE) ||
2997 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2998 filter &= ~SHOW_MEM_FILTER_NODES;
2999 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3000 filter &= ~SHOW_MEM_FILTER_NODES;
3003 struct va_format vaf;
3006 va_start(args, fmt);
3011 pr_warn("%pV", &vaf);
3016 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3017 current->comm, order, gfp_mask, &gfp_mask);
3019 if (!should_suppress_show_mem())
3023 static inline struct page *
3024 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3025 const struct alloc_context *ac, unsigned long *did_some_progress)
3027 struct oom_control oc = {
3028 .zonelist = ac->zonelist,
3029 .nodemask = ac->nodemask,
3031 .gfp_mask = gfp_mask,
3036 *did_some_progress = 0;
3039 * Acquire the oom lock. If that fails, somebody else is
3040 * making progress for us.
3042 if (!mutex_trylock(&oom_lock)) {
3043 *did_some_progress = 1;
3044 schedule_timeout_uninterruptible(1);
3049 * Go through the zonelist yet one more time, keep very high watermark
3050 * here, this is only to catch a parallel oom killing, we must fail if
3051 * we're still under heavy pressure.
3053 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3054 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3058 if (!(gfp_mask & __GFP_NOFAIL)) {
3059 /* Coredumps can quickly deplete all memory reserves */
3060 if (current->flags & PF_DUMPCORE)
3062 /* The OOM killer will not help higher order allocs */
3063 if (order > PAGE_ALLOC_COSTLY_ORDER)
3065 /* The OOM killer does not needlessly kill tasks for lowmem */
3066 if (ac->high_zoneidx < ZONE_NORMAL)
3068 if (pm_suspended_storage())
3071 * XXX: GFP_NOFS allocations should rather fail than rely on
3072 * other request to make a forward progress.
3073 * We are in an unfortunate situation where out_of_memory cannot
3074 * do much for this context but let's try it to at least get
3075 * access to memory reserved if the current task is killed (see
3076 * out_of_memory). Once filesystems are ready to handle allocation
3077 * failures more gracefully we should just bail out here.
3080 /* The OOM killer may not free memory on a specific node */
3081 if (gfp_mask & __GFP_THISNODE)
3084 /* Exhausted what can be done so it's blamo time */
3085 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3086 *did_some_progress = 1;
3088 if (gfp_mask & __GFP_NOFAIL) {
3089 page = get_page_from_freelist(gfp_mask, order,
3090 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3092 * fallback to ignore cpuset restriction if our nodes
3096 page = get_page_from_freelist(gfp_mask, order,
3097 ALLOC_NO_WATERMARKS, ac);
3101 mutex_unlock(&oom_lock);
3106 * Maximum number of compaction retries wit a progress before OOM
3107 * killer is consider as the only way to move forward.
3109 #define MAX_COMPACT_RETRIES 16
3111 #ifdef CONFIG_COMPACTION
3112 /* Try memory compaction for high-order allocations before reclaim */
3113 static struct page *
3114 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3115 unsigned int alloc_flags, const struct alloc_context *ac,
3116 enum compact_priority prio, enum compact_result *compact_result)
3123 current->flags |= PF_MEMALLOC;
3124 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3126 current->flags &= ~PF_MEMALLOC;
3128 if (*compact_result <= COMPACT_INACTIVE)
3132 * At least in one zone compaction wasn't deferred or skipped, so let's
3133 * count a compaction stall
3135 count_vm_event(COMPACTSTALL);
3137 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3140 struct zone *zone = page_zone(page);
3142 zone->compact_blockskip_flush = false;
3143 compaction_defer_reset(zone, order, true);
3144 count_vm_event(COMPACTSUCCESS);
3149 * It's bad if compaction run occurs and fails. The most likely reason
3150 * is that pages exist, but not enough to satisfy watermarks.
3152 count_vm_event(COMPACTFAIL);
3160 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3161 enum compact_result compact_result,
3162 enum compact_priority *compact_priority,
3163 int compaction_retries)
3165 int max_retries = MAX_COMPACT_RETRIES;
3171 * compaction considers all the zone as desperately out of memory
3172 * so it doesn't really make much sense to retry except when the
3173 * failure could be caused by insufficient priority
3175 if (compaction_failed(compact_result)) {
3176 if (*compact_priority > MIN_COMPACT_PRIORITY) {
3177 (*compact_priority)--;
3184 * make sure the compaction wasn't deferred or didn't bail out early
3185 * due to locks contention before we declare that we should give up.
3186 * But do not retry if the given zonelist is not suitable for
3189 if (compaction_withdrawn(compact_result))
3190 return compaction_zonelist_suitable(ac, order, alloc_flags);
3193 * !costly requests are much more important than __GFP_REPEAT
3194 * costly ones because they are de facto nofail and invoke OOM
3195 * killer to move on while costly can fail and users are ready
3196 * to cope with that. 1/4 retries is rather arbitrary but we
3197 * would need much more detailed feedback from compaction to
3198 * make a better decision.
3200 if (order > PAGE_ALLOC_COSTLY_ORDER)
3202 if (compaction_retries <= max_retries)
3208 static inline struct page *
3209 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3210 unsigned int alloc_flags, const struct alloc_context *ac,
3211 enum compact_priority prio, enum compact_result *compact_result)
3213 *compact_result = COMPACT_SKIPPED;
3218 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3219 enum compact_result compact_result,
3220 enum compact_priority *compact_priority,
3221 int compaction_retries)
3226 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3230 * There are setups with compaction disabled which would prefer to loop
3231 * inside the allocator rather than hit the oom killer prematurely.
3232 * Let's give them a good hope and keep retrying while the order-0
3233 * watermarks are OK.
3235 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3237 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3238 ac_classzone_idx(ac), alloc_flags))
3243 #endif /* CONFIG_COMPACTION */
3245 /* Perform direct synchronous page reclaim */
3247 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3248 const struct alloc_context *ac)
3250 struct reclaim_state reclaim_state;
3255 /* We now go into synchronous reclaim */
3256 cpuset_memory_pressure_bump();
3257 current->flags |= PF_MEMALLOC;
3258 lockdep_set_current_reclaim_state(gfp_mask);
3259 reclaim_state.reclaimed_slab = 0;
3260 current->reclaim_state = &reclaim_state;
3262 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3265 current->reclaim_state = NULL;
3266 lockdep_clear_current_reclaim_state();
3267 current->flags &= ~PF_MEMALLOC;
3274 /* The really slow allocator path where we enter direct reclaim */
3275 static inline struct page *
3276 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3277 unsigned int alloc_flags, const struct alloc_context *ac,
3278 unsigned long *did_some_progress)
3280 struct page *page = NULL;
3281 bool drained = false;
3283 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3284 if (unlikely(!(*did_some_progress)))
3288 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3291 * If an allocation failed after direct reclaim, it could be because
3292 * pages are pinned on the per-cpu lists or in high alloc reserves.
3293 * Shrink them them and try again
3295 if (!page && !drained) {
3296 unreserve_highatomic_pageblock(ac);
3297 drain_all_pages(NULL);
3305 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3309 pg_data_t *last_pgdat = NULL;
3311 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3312 ac->high_zoneidx, ac->nodemask) {
3313 if (last_pgdat != zone->zone_pgdat)
3314 wakeup_kswapd(zone, order, ac->high_zoneidx);
3315 last_pgdat = zone->zone_pgdat;
3319 static inline unsigned int
3320 gfp_to_alloc_flags(gfp_t gfp_mask)
3322 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3324 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3325 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3328 * The caller may dip into page reserves a bit more if the caller
3329 * cannot run direct reclaim, or if the caller has realtime scheduling
3330 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3331 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3333 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3335 if (gfp_mask & __GFP_ATOMIC) {
3337 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3338 * if it can't schedule.
3340 if (!(gfp_mask & __GFP_NOMEMALLOC))
3341 alloc_flags |= ALLOC_HARDER;
3343 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3344 * comment for __cpuset_node_allowed().
3346 alloc_flags &= ~ALLOC_CPUSET;
3347 } else if (unlikely(rt_task(current)) && !in_interrupt())
3348 alloc_flags |= ALLOC_HARDER;
3351 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3352 alloc_flags |= ALLOC_CMA;
3357 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3359 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3362 if (gfp_mask & __GFP_MEMALLOC)
3364 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3366 if (!in_interrupt() &&
3367 ((current->flags & PF_MEMALLOC) ||
3368 unlikely(test_thread_flag(TIF_MEMDIE))))
3375 * Maximum number of reclaim retries without any progress before OOM killer
3376 * is consider as the only way to move forward.
3378 #define MAX_RECLAIM_RETRIES 16
3381 * Checks whether it makes sense to retry the reclaim to make a forward progress
3382 * for the given allocation request.
3383 * The reclaim feedback represented by did_some_progress (any progress during
3384 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3385 * any progress in a row) is considered as well as the reclaimable pages on the
3386 * applicable zone list (with a backoff mechanism which is a function of
3387 * no_progress_loops).
3389 * Returns true if a retry is viable or false to enter the oom path.
3392 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3393 struct alloc_context *ac, int alloc_flags,
3394 bool did_some_progress, int no_progress_loops)
3400 * Make sure we converge to OOM if we cannot make any progress
3401 * several times in the row.
3403 if (no_progress_loops > MAX_RECLAIM_RETRIES)
3407 * Keep reclaiming pages while there is a chance this will lead
3408 * somewhere. If none of the target zones can satisfy our allocation
3409 * request even if all reclaimable pages are considered then we are
3410 * screwed and have to go OOM.
3412 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3414 unsigned long available;
3415 unsigned long reclaimable;
3417 available = reclaimable = zone_reclaimable_pages(zone);
3418 available -= DIV_ROUND_UP(no_progress_loops * available,
3419 MAX_RECLAIM_RETRIES);
3420 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3423 * Would the allocation succeed if we reclaimed the whole
3426 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3427 ac_classzone_idx(ac), alloc_flags, available)) {
3429 * If we didn't make any progress and have a lot of
3430 * dirty + writeback pages then we should wait for
3431 * an IO to complete to slow down the reclaim and
3432 * prevent from pre mature OOM
3434 if (!did_some_progress) {
3435 unsigned long write_pending;
3437 write_pending = zone_page_state_snapshot(zone,
3438 NR_ZONE_WRITE_PENDING);
3440 if (2 * write_pending > reclaimable) {
3441 congestion_wait(BLK_RW_ASYNC, HZ/10);
3447 * Memory allocation/reclaim might be called from a WQ
3448 * context and the current implementation of the WQ
3449 * concurrency control doesn't recognize that
3450 * a particular WQ is congested if the worker thread is
3451 * looping without ever sleeping. Therefore we have to
3452 * do a short sleep here rather than calling
3455 if (current->flags & PF_WQ_WORKER)
3456 schedule_timeout_uninterruptible(1);
3467 static inline struct page *
3468 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3469 struct alloc_context *ac)
3471 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3472 struct page *page = NULL;
3473 unsigned int alloc_flags;
3474 unsigned long did_some_progress;
3475 enum compact_priority compact_priority = DEF_COMPACT_PRIORITY;
3476 enum compact_result compact_result;
3477 int compaction_retries = 0;
3478 int no_progress_loops = 0;
3481 * In the slowpath, we sanity check order to avoid ever trying to
3482 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3483 * be using allocators in order of preference for an area that is
3486 if (order >= MAX_ORDER) {
3487 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3492 * We also sanity check to catch abuse of atomic reserves being used by
3493 * callers that are not in atomic context.
3495 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3496 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3497 gfp_mask &= ~__GFP_ATOMIC;
3500 * The fast path uses conservative alloc_flags to succeed only until
3501 * kswapd needs to be woken up, and to avoid the cost of setting up
3502 * alloc_flags precisely. So we do that now.
3504 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3506 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3507 wake_all_kswapds(order, ac);
3510 * The adjusted alloc_flags might result in immediate success, so try
3513 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3518 * For costly allocations, try direct compaction first, as it's likely
3519 * that we have enough base pages and don't need to reclaim. Don't try
3520 * that for allocations that are allowed to ignore watermarks, as the
3521 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3523 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3524 !gfp_pfmemalloc_allowed(gfp_mask)) {
3525 page = __alloc_pages_direct_compact(gfp_mask, order,
3527 INIT_COMPACT_PRIORITY,
3533 * Checks for costly allocations with __GFP_NORETRY, which
3534 * includes THP page fault allocations
3536 if (gfp_mask & __GFP_NORETRY) {
3538 * If compaction is deferred for high-order allocations,
3539 * it is because sync compaction recently failed. If
3540 * this is the case and the caller requested a THP
3541 * allocation, we do not want to heavily disrupt the
3542 * system, so we fail the allocation instead of entering
3545 if (compact_result == COMPACT_DEFERRED)
3549 * Looks like reclaim/compaction is worth trying, but
3550 * sync compaction could be very expensive, so keep
3551 * using async compaction.
3553 compact_priority = INIT_COMPACT_PRIORITY;
3558 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3559 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3560 wake_all_kswapds(order, ac);
3562 if (gfp_pfmemalloc_allowed(gfp_mask))
3563 alloc_flags = ALLOC_NO_WATERMARKS;
3566 * Reset the zonelist iterators if memory policies can be ignored.
3567 * These allocations are high priority and system rather than user
3570 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3571 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3572 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3573 ac->high_zoneidx, ac->nodemask);
3576 /* Attempt with potentially adjusted zonelist and alloc_flags */
3577 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3581 /* Caller is not willing to reclaim, we can't balance anything */
3582 if (!can_direct_reclaim) {
3584 * All existing users of the __GFP_NOFAIL are blockable, so warn
3585 * of any new users that actually allow this type of allocation
3588 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3592 /* Avoid recursion of direct reclaim */
3593 if (current->flags & PF_MEMALLOC) {
3595 * __GFP_NOFAIL request from this context is rather bizarre
3596 * because we cannot reclaim anything and only can loop waiting
3597 * for somebody to do a work for us.
3599 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3606 /* Avoid allocations with no watermarks from looping endlessly */
3607 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3611 /* Try direct reclaim and then allocating */
3612 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3613 &did_some_progress);
3617 /* Try direct compaction and then allocating */
3618 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3619 compact_priority, &compact_result);
3623 if (order && compaction_made_progress(compact_result))
3624 compaction_retries++;
3626 /* Do not loop if specifically requested */
3627 if (gfp_mask & __GFP_NORETRY)
3631 * Do not retry costly high order allocations unless they are
3634 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3638 * Costly allocations might have made a progress but this doesn't mean
3639 * their order will become available due to high fragmentation so
3640 * always increment the no progress counter for them
3642 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3643 no_progress_loops = 0;
3645 no_progress_loops++;
3647 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3648 did_some_progress > 0, no_progress_loops))
3652 * It doesn't make any sense to retry for the compaction if the order-0
3653 * reclaim is not able to make any progress because the current
3654 * implementation of the compaction depends on the sufficient amount
3655 * of free memory (see __compaction_suitable)
3657 if (did_some_progress > 0 &&
3658 should_compact_retry(ac, order, alloc_flags,
3659 compact_result, &compact_priority,
3660 compaction_retries))
3663 /* Reclaim has failed us, start killing things */
3664 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3668 /* Retry as long as the OOM killer is making progress */
3669 if (did_some_progress) {
3670 no_progress_loops = 0;
3675 warn_alloc_failed(gfp_mask, order, NULL);
3681 * This is the 'heart' of the zoned buddy allocator.
3684 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3685 struct zonelist *zonelist, nodemask_t *nodemask)
3688 unsigned int cpuset_mems_cookie;
3689 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3690 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3691 struct alloc_context ac = {
3692 .high_zoneidx = gfp_zone(gfp_mask),
3693 .zonelist = zonelist,
3694 .nodemask = nodemask,
3695 .migratetype = gfpflags_to_migratetype(gfp_mask),
3698 if (cpusets_enabled()) {
3699 alloc_mask |= __GFP_HARDWALL;
3700 alloc_flags |= ALLOC_CPUSET;
3702 ac.nodemask = &cpuset_current_mems_allowed;
3705 gfp_mask &= gfp_allowed_mask;
3707 lockdep_trace_alloc(gfp_mask);
3709 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3711 if (should_fail_alloc_page(gfp_mask, order))
3715 * Check the zones suitable for the gfp_mask contain at least one
3716 * valid zone. It's possible to have an empty zonelist as a result
3717 * of __GFP_THISNODE and a memoryless node
3719 if (unlikely(!zonelist->_zonerefs->zone))
3722 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3723 alloc_flags |= ALLOC_CMA;
3726 cpuset_mems_cookie = read_mems_allowed_begin();
3728 /* Dirty zone balancing only done in the fast path */
3729 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3732 * The preferred zone is used for statistics but crucially it is
3733 * also used as the starting point for the zonelist iterator. It
3734 * may get reset for allocations that ignore memory policies.
3736 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3737 ac.high_zoneidx, ac.nodemask);
3738 if (!ac.preferred_zoneref) {
3743 /* First allocation attempt */
3744 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3749 * Runtime PM, block IO and its error handling path can deadlock
3750 * because I/O on the device might not complete.
3752 alloc_mask = memalloc_noio_flags(gfp_mask);
3753 ac.spread_dirty_pages = false;
3756 * Restore the original nodemask if it was potentially replaced with
3757 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3759 if (cpusets_enabled())
3760 ac.nodemask = nodemask;
3761 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3765 * When updating a task's mems_allowed, it is possible to race with
3766 * parallel threads in such a way that an allocation can fail while
3767 * the mask is being updated. If a page allocation is about to fail,
3768 * check if the cpuset changed during allocation and if so, retry.
3770 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3771 alloc_mask = gfp_mask;
3776 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3777 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3778 __free_pages(page, order);
3782 if (kmemcheck_enabled && page)
3783 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3785 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3789 EXPORT_SYMBOL(__alloc_pages_nodemask);
3792 * Common helper functions.
3794 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3799 * __get_free_pages() returns a 32-bit address, which cannot represent
3802 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3804 page = alloc_pages(gfp_mask, order);
3807 return (unsigned long) page_address(page);
3809 EXPORT_SYMBOL(__get_free_pages);
3811 unsigned long get_zeroed_page(gfp_t gfp_mask)
3813 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3815 EXPORT_SYMBOL(get_zeroed_page);
3817 void __free_pages(struct page *page, unsigned int order)
3819 if (put_page_testzero(page)) {
3821 free_hot_cold_page(page, false);
3823 __free_pages_ok(page, order);
3827 EXPORT_SYMBOL(__free_pages);
3829 void free_pages(unsigned long addr, unsigned int order)
3832 VM_BUG_ON(!virt_addr_valid((void *)addr));
3833 __free_pages(virt_to_page((void *)addr), order);
3837 EXPORT_SYMBOL(free_pages);
3841 * An arbitrary-length arbitrary-offset area of memory which resides
3842 * within a 0 or higher order page. Multiple fragments within that page
3843 * are individually refcounted, in the page's reference counter.
3845 * The page_frag functions below provide a simple allocation framework for
3846 * page fragments. This is used by the network stack and network device
3847 * drivers to provide a backing region of memory for use as either an
3848 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3850 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3853 struct page *page = NULL;
3854 gfp_t gfp = gfp_mask;
3856 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3857 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3859 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3860 PAGE_FRAG_CACHE_MAX_ORDER);
3861 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3863 if (unlikely(!page))
3864 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3866 nc->va = page ? page_address(page) : NULL;
3871 void *__alloc_page_frag(struct page_frag_cache *nc,
3872 unsigned int fragsz, gfp_t gfp_mask)
3874 unsigned int size = PAGE_SIZE;
3878 if (unlikely(!nc->va)) {
3880 page = __page_frag_refill(nc, gfp_mask);
3884 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3885 /* if size can vary use size else just use PAGE_SIZE */
3888 /* Even if we own the page, we do not use atomic_set().
3889 * This would break get_page_unless_zero() users.
3891 page_ref_add(page, size - 1);
3893 /* reset page count bias and offset to start of new frag */
3894 nc->pfmemalloc = page_is_pfmemalloc(page);
3895 nc->pagecnt_bias = size;
3899 offset = nc->offset - fragsz;
3900 if (unlikely(offset < 0)) {
3901 page = virt_to_page(nc->va);
3903 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3906 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3907 /* if size can vary use size else just use PAGE_SIZE */
3910 /* OK, page count is 0, we can safely set it */
3911 set_page_count(page, size);
3913 /* reset page count bias and offset to start of new frag */
3914 nc->pagecnt_bias = size;
3915 offset = size - fragsz;
3919 nc->offset = offset;
3921 return nc->va + offset;
3923 EXPORT_SYMBOL(__alloc_page_frag);
3926 * Frees a page fragment allocated out of either a compound or order 0 page.
3928 void __free_page_frag(void *addr)
3930 struct page *page = virt_to_head_page(addr);
3932 if (unlikely(put_page_testzero(page)))
3933 __free_pages_ok(page, compound_order(page));
3935 EXPORT_SYMBOL(__free_page_frag);
3937 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3941 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3942 unsigned long used = addr + PAGE_ALIGN(size);
3944 split_page(virt_to_page((void *)addr), order);
3945 while (used < alloc_end) {
3950 return (void *)addr;
3954 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3955 * @size: the number of bytes to allocate
3956 * @gfp_mask: GFP flags for the allocation
3958 * This function is similar to alloc_pages(), except that it allocates the
3959 * minimum number of pages to satisfy the request. alloc_pages() can only
3960 * allocate memory in power-of-two pages.
3962 * This function is also limited by MAX_ORDER.
3964 * Memory allocated by this function must be released by free_pages_exact().
3966 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3968 unsigned int order = get_order(size);
3971 addr = __get_free_pages(gfp_mask, order);
3972 return make_alloc_exact(addr, order, size);
3974 EXPORT_SYMBOL(alloc_pages_exact);
3977 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3979 * @nid: the preferred node ID where memory should be allocated
3980 * @size: the number of bytes to allocate
3981 * @gfp_mask: GFP flags for the allocation
3983 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3986 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3988 unsigned int order = get_order(size);
3989 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3992 return make_alloc_exact((unsigned long)page_address(p), order, size);
3996 * free_pages_exact - release memory allocated via alloc_pages_exact()
3997 * @virt: the value returned by alloc_pages_exact.
3998 * @size: size of allocation, same value as passed to alloc_pages_exact().
4000 * Release the memory allocated by a previous call to alloc_pages_exact.
4002 void free_pages_exact(void *virt, size_t size)
4004 unsigned long addr = (unsigned long)virt;
4005 unsigned long end = addr + PAGE_ALIGN(size);
4007 while (addr < end) {
4012 EXPORT_SYMBOL(free_pages_exact);
4015 * nr_free_zone_pages - count number of pages beyond high watermark
4016 * @offset: The zone index of the highest zone
4018 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4019 * high watermark within all zones at or below a given zone index. For each
4020 * zone, the number of pages is calculated as:
4021 * managed_pages - high_pages
4023 static unsigned long nr_free_zone_pages(int offset)
4028 /* Just pick one node, since fallback list is circular */
4029 unsigned long sum = 0;
4031 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4033 for_each_zone_zonelist(zone, z, zonelist, offset) {
4034 unsigned long size = zone->managed_pages;
4035 unsigned long high = high_wmark_pages(zone);
4044 * nr_free_buffer_pages - count number of pages beyond high watermark
4046 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4047 * watermark within ZONE_DMA and ZONE_NORMAL.
4049 unsigned long nr_free_buffer_pages(void)
4051 return nr_free_zone_pages(gfp_zone(GFP_USER));
4053 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4056 * nr_free_pagecache_pages - count number of pages beyond high watermark
4058 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4059 * high watermark within all zones.
4061 unsigned long nr_free_pagecache_pages(void)
4063 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4066 static inline void show_node(struct zone *zone)
4068 if (IS_ENABLED(CONFIG_NUMA))
4069 printk("Node %d ", zone_to_nid(zone));
4072 long si_mem_available(void)
4075 unsigned long pagecache;
4076 unsigned long wmark_low = 0;
4077 unsigned long pages[NR_LRU_LISTS];
4081 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4082 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4085 wmark_low += zone->watermark[WMARK_LOW];
4088 * Estimate the amount of memory available for userspace allocations,
4089 * without causing swapping.
4091 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4094 * Not all the page cache can be freed, otherwise the system will
4095 * start swapping. Assume at least half of the page cache, or the
4096 * low watermark worth of cache, needs to stay.
4098 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4099 pagecache -= min(pagecache / 2, wmark_low);
4100 available += pagecache;
4103 * Part of the reclaimable slab consists of items that are in use,
4104 * and cannot be freed. Cap this estimate at the low watermark.
4106 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4107 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4113 EXPORT_SYMBOL_GPL(si_mem_available);
4115 void si_meminfo(struct sysinfo *val)
4117 val->totalram = totalram_pages;
4118 val->sharedram = global_node_page_state(NR_SHMEM);
4119 val->freeram = global_page_state(NR_FREE_PAGES);
4120 val->bufferram = nr_blockdev_pages();
4121 val->totalhigh = totalhigh_pages;
4122 val->freehigh = nr_free_highpages();
4123 val->mem_unit = PAGE_SIZE;
4126 EXPORT_SYMBOL(si_meminfo);
4129 void si_meminfo_node(struct sysinfo *val, int nid)
4131 int zone_type; /* needs to be signed */
4132 unsigned long managed_pages = 0;
4133 unsigned long managed_highpages = 0;
4134 unsigned long free_highpages = 0;
4135 pg_data_t *pgdat = NODE_DATA(nid);
4137 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4138 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4139 val->totalram = managed_pages;
4140 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4141 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4142 #ifdef CONFIG_HIGHMEM
4143 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4144 struct zone *zone = &pgdat->node_zones[zone_type];
4146 if (is_highmem(zone)) {
4147 managed_highpages += zone->managed_pages;
4148 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4151 val->totalhigh = managed_highpages;
4152 val->freehigh = free_highpages;
4154 val->totalhigh = managed_highpages;
4155 val->freehigh = free_highpages;
4157 val->mem_unit = PAGE_SIZE;
4162 * Determine whether the node should be displayed or not, depending on whether
4163 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4165 bool skip_free_areas_node(unsigned int flags, int nid)
4168 unsigned int cpuset_mems_cookie;
4170 if (!(flags & SHOW_MEM_FILTER_NODES))
4174 cpuset_mems_cookie = read_mems_allowed_begin();
4175 ret = !node_isset(nid, cpuset_current_mems_allowed);
4176 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4181 #define K(x) ((x) << (PAGE_SHIFT-10))
4183 static void show_migration_types(unsigned char type)
4185 static const char types[MIGRATE_TYPES] = {
4186 [MIGRATE_UNMOVABLE] = 'U',
4187 [MIGRATE_MOVABLE] = 'M',
4188 [MIGRATE_RECLAIMABLE] = 'E',
4189 [MIGRATE_HIGHATOMIC] = 'H',
4191 [MIGRATE_CMA] = 'C',
4193 #ifdef CONFIG_MEMORY_ISOLATION
4194 [MIGRATE_ISOLATE] = 'I',
4197 char tmp[MIGRATE_TYPES + 1];
4201 for (i = 0; i < MIGRATE_TYPES; i++) {
4202 if (type & (1 << i))
4207 printk("(%s) ", tmp);
4211 * Show free area list (used inside shift_scroll-lock stuff)
4212 * We also calculate the percentage fragmentation. We do this by counting the
4213 * memory on each free list with the exception of the first item on the list.
4216 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4219 void show_free_areas(unsigned int filter)
4221 unsigned long free_pcp = 0;
4226 for_each_populated_zone(zone) {
4227 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4230 for_each_online_cpu(cpu)
4231 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4234 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4235 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4236 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4237 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4238 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4239 " free:%lu free_pcp:%lu free_cma:%lu\n",
4240 global_node_page_state(NR_ACTIVE_ANON),
4241 global_node_page_state(NR_INACTIVE_ANON),
4242 global_node_page_state(NR_ISOLATED_ANON),
4243 global_node_page_state(NR_ACTIVE_FILE),
4244 global_node_page_state(NR_INACTIVE_FILE),
4245 global_node_page_state(NR_ISOLATED_FILE),
4246 global_node_page_state(NR_UNEVICTABLE),
4247 global_node_page_state(NR_FILE_DIRTY),
4248 global_node_page_state(NR_WRITEBACK),
4249 global_node_page_state(NR_UNSTABLE_NFS),
4250 global_page_state(NR_SLAB_RECLAIMABLE),
4251 global_page_state(NR_SLAB_UNRECLAIMABLE),
4252 global_node_page_state(NR_FILE_MAPPED),
4253 global_node_page_state(NR_SHMEM),
4254 global_page_state(NR_PAGETABLE),
4255 global_page_state(NR_BOUNCE),
4256 global_page_state(NR_FREE_PAGES),
4258 global_page_state(NR_FREE_CMA_PAGES));
4260 for_each_online_pgdat(pgdat) {
4262 " active_anon:%lukB"
4263 " inactive_anon:%lukB"
4264 " active_file:%lukB"
4265 " inactive_file:%lukB"
4266 " unevictable:%lukB"
4267 " isolated(anon):%lukB"
4268 " isolated(file):%lukB"
4273 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4275 " shmem_pmdmapped: %lukB"
4278 " writeback_tmp:%lukB"
4280 " pages_scanned:%lu"
4281 " all_unreclaimable? %s"
4284 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4285 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4286 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4287 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4288 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4289 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4290 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4291 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4292 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4293 K(node_page_state(pgdat, NR_WRITEBACK)),
4294 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4295 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4296 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4298 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4300 K(node_page_state(pgdat, NR_SHMEM)),
4301 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4302 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4303 node_page_state(pgdat, NR_PAGES_SCANNED),
4304 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4307 for_each_populated_zone(zone) {
4310 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4314 for_each_online_cpu(cpu)
4315 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4323 " active_anon:%lukB"
4324 " inactive_anon:%lukB"
4325 " active_file:%lukB"
4326 " inactive_file:%lukB"
4327 " unevictable:%lukB"
4328 " writepending:%lukB"
4332 " slab_reclaimable:%lukB"
4333 " slab_unreclaimable:%lukB"
4334 " kernel_stack:%lukB"
4342 K(zone_page_state(zone, NR_FREE_PAGES)),
4343 K(min_wmark_pages(zone)),
4344 K(low_wmark_pages(zone)),
4345 K(high_wmark_pages(zone)),
4346 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4347 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4348 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4349 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4350 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4351 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4352 K(zone->present_pages),
4353 K(zone->managed_pages),
4354 K(zone_page_state(zone, NR_MLOCK)),
4355 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4356 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4357 zone_page_state(zone, NR_KERNEL_STACK_KB),
4358 K(zone_page_state(zone, NR_PAGETABLE)),
4359 K(zone_page_state(zone, NR_BOUNCE)),
4361 K(this_cpu_read(zone->pageset->pcp.count)),
4362 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4363 printk("lowmem_reserve[]:");
4364 for (i = 0; i < MAX_NR_ZONES; i++)
4365 printk(" %ld", zone->lowmem_reserve[i]);
4369 for_each_populated_zone(zone) {
4371 unsigned long nr[MAX_ORDER], flags, total = 0;
4372 unsigned char types[MAX_ORDER];
4374 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4377 printk("%s: ", zone->name);
4379 spin_lock_irqsave(&zone->lock, flags);
4380 for (order = 0; order < MAX_ORDER; order++) {
4381 struct free_area *area = &zone->free_area[order];
4384 nr[order] = area->nr_free;
4385 total += nr[order] << order;
4388 for (type = 0; type < MIGRATE_TYPES; type++) {
4389 if (!list_empty(&area->free_list[type]))
4390 types[order] |= 1 << type;
4393 spin_unlock_irqrestore(&zone->lock, flags);
4394 for (order = 0; order < MAX_ORDER; order++) {
4395 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4397 show_migration_types(types[order]);
4399 printk("= %lukB\n", K(total));
4402 hugetlb_show_meminfo();
4404 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4406 show_swap_cache_info();
4409 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4411 zoneref->zone = zone;
4412 zoneref->zone_idx = zone_idx(zone);
4416 * Builds allocation fallback zone lists.
4418 * Add all populated zones of a node to the zonelist.
4420 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4424 enum zone_type zone_type = MAX_NR_ZONES;
4428 zone = pgdat->node_zones + zone_type;
4429 if (managed_zone(zone)) {
4430 zoneref_set_zone(zone,
4431 &zonelist->_zonerefs[nr_zones++]);
4432 check_highest_zone(zone_type);
4434 } while (zone_type);
4442 * 0 = automatic detection of better ordering.
4443 * 1 = order by ([node] distance, -zonetype)
4444 * 2 = order by (-zonetype, [node] distance)
4446 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4447 * the same zonelist. So only NUMA can configure this param.
4449 #define ZONELIST_ORDER_DEFAULT 0
4450 #define ZONELIST_ORDER_NODE 1
4451 #define ZONELIST_ORDER_ZONE 2
4453 /* zonelist order in the kernel.
4454 * set_zonelist_order() will set this to NODE or ZONE.
4456 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4457 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4461 /* The value user specified ....changed by config */
4462 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4463 /* string for sysctl */
4464 #define NUMA_ZONELIST_ORDER_LEN 16
4465 char numa_zonelist_order[16] = "default";
4468 * interface for configure zonelist ordering.
4469 * command line option "numa_zonelist_order"
4470 * = "[dD]efault - default, automatic configuration.
4471 * = "[nN]ode - order by node locality, then by zone within node
4472 * = "[zZ]one - order by zone, then by locality within zone
4475 static int __parse_numa_zonelist_order(char *s)
4477 if (*s == 'd' || *s == 'D') {
4478 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4479 } else if (*s == 'n' || *s == 'N') {
4480 user_zonelist_order = ZONELIST_ORDER_NODE;
4481 } else if (*s == 'z' || *s == 'Z') {
4482 user_zonelist_order = ZONELIST_ORDER_ZONE;
4484 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4490 static __init int setup_numa_zonelist_order(char *s)
4497 ret = __parse_numa_zonelist_order(s);
4499 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4503 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4506 * sysctl handler for numa_zonelist_order
4508 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4509 void __user *buffer, size_t *length,
4512 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4514 static DEFINE_MUTEX(zl_order_mutex);
4516 mutex_lock(&zl_order_mutex);
4518 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4522 strcpy(saved_string, (char *)table->data);
4524 ret = proc_dostring(table, write, buffer, length, ppos);
4528 int oldval = user_zonelist_order;
4530 ret = __parse_numa_zonelist_order((char *)table->data);
4533 * bogus value. restore saved string
4535 strncpy((char *)table->data, saved_string,
4536 NUMA_ZONELIST_ORDER_LEN);
4537 user_zonelist_order = oldval;
4538 } else if (oldval != user_zonelist_order) {
4539 mutex_lock(&zonelists_mutex);
4540 build_all_zonelists(NULL, NULL);
4541 mutex_unlock(&zonelists_mutex);
4545 mutex_unlock(&zl_order_mutex);
4550 #define MAX_NODE_LOAD (nr_online_nodes)
4551 static int node_load[MAX_NUMNODES];
4554 * find_next_best_node - find the next node that should appear in a given node's fallback list
4555 * @node: node whose fallback list we're appending
4556 * @used_node_mask: nodemask_t of already used nodes
4558 * We use a number of factors to determine which is the next node that should
4559 * appear on a given node's fallback list. The node should not have appeared
4560 * already in @node's fallback list, and it should be the next closest node
4561 * according to the distance array (which contains arbitrary distance values
4562 * from each node to each node in the system), and should also prefer nodes
4563 * with no CPUs, since presumably they'll have very little allocation pressure
4564 * on them otherwise.
4565 * It returns -1 if no node is found.
4567 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4570 int min_val = INT_MAX;
4571 int best_node = NUMA_NO_NODE;
4572 const struct cpumask *tmp = cpumask_of_node(0);
4574 /* Use the local node if we haven't already */
4575 if (!node_isset(node, *used_node_mask)) {
4576 node_set(node, *used_node_mask);
4580 for_each_node_state(n, N_MEMORY) {
4582 /* Don't want a node to appear more than once */
4583 if (node_isset(n, *used_node_mask))
4586 /* Use the distance array to find the distance */
4587 val = node_distance(node, n);
4589 /* Penalize nodes under us ("prefer the next node") */
4592 /* Give preference to headless and unused nodes */
4593 tmp = cpumask_of_node(n);
4594 if (!cpumask_empty(tmp))
4595 val += PENALTY_FOR_NODE_WITH_CPUS;
4597 /* Slight preference for less loaded node */
4598 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4599 val += node_load[n];
4601 if (val < min_val) {
4608 node_set(best_node, *used_node_mask);
4615 * Build zonelists ordered by node and zones within node.
4616 * This results in maximum locality--normal zone overflows into local
4617 * DMA zone, if any--but risks exhausting DMA zone.
4619 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4622 struct zonelist *zonelist;
4624 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4625 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4627 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4628 zonelist->_zonerefs[j].zone = NULL;
4629 zonelist->_zonerefs[j].zone_idx = 0;
4633 * Build gfp_thisnode zonelists
4635 static void build_thisnode_zonelists(pg_data_t *pgdat)
4638 struct zonelist *zonelist;
4640 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4641 j = build_zonelists_node(pgdat, zonelist, 0);
4642 zonelist->_zonerefs[j].zone = NULL;
4643 zonelist->_zonerefs[j].zone_idx = 0;
4647 * Build zonelists ordered by zone and nodes within zones.
4648 * This results in conserving DMA zone[s] until all Normal memory is
4649 * exhausted, but results in overflowing to remote node while memory
4650 * may still exist in local DMA zone.
4652 static int node_order[MAX_NUMNODES];
4654 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4657 int zone_type; /* needs to be signed */
4659 struct zonelist *zonelist;
4661 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4663 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4664 for (j = 0; j < nr_nodes; j++) {
4665 node = node_order[j];
4666 z = &NODE_DATA(node)->node_zones[zone_type];
4667 if (managed_zone(z)) {
4669 &zonelist->_zonerefs[pos++]);
4670 check_highest_zone(zone_type);
4674 zonelist->_zonerefs[pos].zone = NULL;
4675 zonelist->_zonerefs[pos].zone_idx = 0;
4678 #if defined(CONFIG_64BIT)
4680 * Devices that require DMA32/DMA are relatively rare and do not justify a
4681 * penalty to every machine in case the specialised case applies. Default
4682 * to Node-ordering on 64-bit NUMA machines
4684 static int default_zonelist_order(void)
4686 return ZONELIST_ORDER_NODE;
4690 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4691 * by the kernel. If processes running on node 0 deplete the low memory zone
4692 * then reclaim will occur more frequency increasing stalls and potentially
4693 * be easier to OOM if a large percentage of the zone is under writeback or
4694 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4695 * Hence, default to zone ordering on 32-bit.
4697 static int default_zonelist_order(void)
4699 return ZONELIST_ORDER_ZONE;
4701 #endif /* CONFIG_64BIT */
4703 static void set_zonelist_order(void)
4705 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4706 current_zonelist_order = default_zonelist_order();
4708 current_zonelist_order = user_zonelist_order;
4711 static void build_zonelists(pg_data_t *pgdat)
4714 nodemask_t used_mask;
4715 int local_node, prev_node;
4716 struct zonelist *zonelist;
4717 unsigned int order = current_zonelist_order;
4719 /* initialize zonelists */
4720 for (i = 0; i < MAX_ZONELISTS; i++) {
4721 zonelist = pgdat->node_zonelists + i;
4722 zonelist->_zonerefs[0].zone = NULL;
4723 zonelist->_zonerefs[0].zone_idx = 0;
4726 /* NUMA-aware ordering of nodes */
4727 local_node = pgdat->node_id;
4728 load = nr_online_nodes;
4729 prev_node = local_node;
4730 nodes_clear(used_mask);
4732 memset(node_order, 0, sizeof(node_order));
4735 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4737 * We don't want to pressure a particular node.
4738 * So adding penalty to the first node in same
4739 * distance group to make it round-robin.
4741 if (node_distance(local_node, node) !=
4742 node_distance(local_node, prev_node))
4743 node_load[node] = load;
4747 if (order == ZONELIST_ORDER_NODE)
4748 build_zonelists_in_node_order(pgdat, node);
4750 node_order[i++] = node; /* remember order */
4753 if (order == ZONELIST_ORDER_ZONE) {
4754 /* calculate node order -- i.e., DMA last! */
4755 build_zonelists_in_zone_order(pgdat, i);
4758 build_thisnode_zonelists(pgdat);
4761 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4763 * Return node id of node used for "local" allocations.
4764 * I.e., first node id of first zone in arg node's generic zonelist.
4765 * Used for initializing percpu 'numa_mem', which is used primarily
4766 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4768 int local_memory_node(int node)
4772 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4773 gfp_zone(GFP_KERNEL),
4775 return z->zone->node;
4779 static void setup_min_unmapped_ratio(void);
4780 static void setup_min_slab_ratio(void);
4781 #else /* CONFIG_NUMA */
4783 static void set_zonelist_order(void)
4785 current_zonelist_order = ZONELIST_ORDER_ZONE;
4788 static void build_zonelists(pg_data_t *pgdat)
4790 int node, local_node;
4792 struct zonelist *zonelist;
4794 local_node = pgdat->node_id;
4796 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4797 j = build_zonelists_node(pgdat, zonelist, 0);
4800 * Now we build the zonelist so that it contains the zones
4801 * of all the other nodes.
4802 * We don't want to pressure a particular node, so when
4803 * building the zones for node N, we make sure that the
4804 * zones coming right after the local ones are those from
4805 * node N+1 (modulo N)
4807 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4808 if (!node_online(node))
4810 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4812 for (node = 0; node < local_node; node++) {
4813 if (!node_online(node))
4815 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4818 zonelist->_zonerefs[j].zone = NULL;
4819 zonelist->_zonerefs[j].zone_idx = 0;
4822 #endif /* CONFIG_NUMA */
4825 * Boot pageset table. One per cpu which is going to be used for all
4826 * zones and all nodes. The parameters will be set in such a way
4827 * that an item put on a list will immediately be handed over to
4828 * the buddy list. This is safe since pageset manipulation is done
4829 * with interrupts disabled.
4831 * The boot_pagesets must be kept even after bootup is complete for
4832 * unused processors and/or zones. They do play a role for bootstrapping
4833 * hotplugged processors.
4835 * zoneinfo_show() and maybe other functions do
4836 * not check if the processor is online before following the pageset pointer.
4837 * Other parts of the kernel may not check if the zone is available.
4839 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4840 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4841 static void setup_zone_pageset(struct zone *zone);
4844 * Global mutex to protect against size modification of zonelists
4845 * as well as to serialize pageset setup for the new populated zone.
4847 DEFINE_MUTEX(zonelists_mutex);
4849 /* return values int ....just for stop_machine() */
4850 static int __build_all_zonelists(void *data)
4854 pg_data_t *self = data;
4857 memset(node_load, 0, sizeof(node_load));
4860 if (self && !node_online(self->node_id)) {
4861 build_zonelists(self);
4864 for_each_online_node(nid) {
4865 pg_data_t *pgdat = NODE_DATA(nid);
4867 build_zonelists(pgdat);
4871 * Initialize the boot_pagesets that are going to be used
4872 * for bootstrapping processors. The real pagesets for
4873 * each zone will be allocated later when the per cpu
4874 * allocator is available.
4876 * boot_pagesets are used also for bootstrapping offline
4877 * cpus if the system is already booted because the pagesets
4878 * are needed to initialize allocators on a specific cpu too.
4879 * F.e. the percpu allocator needs the page allocator which
4880 * needs the percpu allocator in order to allocate its pagesets
4881 * (a chicken-egg dilemma).
4883 for_each_possible_cpu(cpu) {
4884 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4886 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4888 * We now know the "local memory node" for each node--
4889 * i.e., the node of the first zone in the generic zonelist.
4890 * Set up numa_mem percpu variable for on-line cpus. During
4891 * boot, only the boot cpu should be on-line; we'll init the
4892 * secondary cpus' numa_mem as they come on-line. During
4893 * node/memory hotplug, we'll fixup all on-line cpus.
4895 if (cpu_online(cpu))
4896 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4903 static noinline void __init
4904 build_all_zonelists_init(void)
4906 __build_all_zonelists(NULL);
4907 mminit_verify_zonelist();
4908 cpuset_init_current_mems_allowed();
4912 * Called with zonelists_mutex held always
4913 * unless system_state == SYSTEM_BOOTING.
4915 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4916 * [we're only called with non-NULL zone through __meminit paths] and
4917 * (2) call of __init annotated helper build_all_zonelists_init
4918 * [protected by SYSTEM_BOOTING].
4920 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4922 set_zonelist_order();
4924 if (system_state == SYSTEM_BOOTING) {
4925 build_all_zonelists_init();
4927 #ifdef CONFIG_MEMORY_HOTPLUG
4929 setup_zone_pageset(zone);
4931 /* we have to stop all cpus to guarantee there is no user
4933 stop_machine(__build_all_zonelists, pgdat, NULL);
4934 /* cpuset refresh routine should be here */
4936 vm_total_pages = nr_free_pagecache_pages();
4938 * Disable grouping by mobility if the number of pages in the
4939 * system is too low to allow the mechanism to work. It would be
4940 * more accurate, but expensive to check per-zone. This check is
4941 * made on memory-hotadd so a system can start with mobility
4942 * disabled and enable it later
4944 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4945 page_group_by_mobility_disabled = 1;
4947 page_group_by_mobility_disabled = 0;
4949 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4951 zonelist_order_name[current_zonelist_order],
4952 page_group_by_mobility_disabled ? "off" : "on",
4955 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4960 * Helper functions to size the waitqueue hash table.
4961 * Essentially these want to choose hash table sizes sufficiently
4962 * large so that collisions trying to wait on pages are rare.
4963 * But in fact, the number of active page waitqueues on typical
4964 * systems is ridiculously low, less than 200. So this is even
4965 * conservative, even though it seems large.
4967 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4968 * waitqueues, i.e. the size of the waitq table given the number of pages.
4970 #define PAGES_PER_WAITQUEUE 256
4972 #ifndef CONFIG_MEMORY_HOTPLUG
4973 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4975 unsigned long size = 1;
4977 pages /= PAGES_PER_WAITQUEUE;
4979 while (size < pages)
4983 * Once we have dozens or even hundreds of threads sleeping
4984 * on IO we've got bigger problems than wait queue collision.
4985 * Limit the size of the wait table to a reasonable size.
4987 size = min(size, 4096UL);
4989 return max(size, 4UL);
4993 * A zone's size might be changed by hot-add, so it is not possible to determine
4994 * a suitable size for its wait_table. So we use the maximum size now.
4996 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4998 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4999 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5000 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
5002 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5003 * or more by the traditional way. (See above). It equals:
5005 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
5006 * ia64(16K page size) : = ( 8G + 4M)byte.
5007 * powerpc (64K page size) : = (32G +16M)byte.
5009 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5016 * This is an integer logarithm so that shifts can be used later
5017 * to extract the more random high bits from the multiplicative
5018 * hash function before the remainder is taken.
5020 static inline unsigned long wait_table_bits(unsigned long size)
5026 * Initially all pages are reserved - free ones are freed
5027 * up by free_all_bootmem() once the early boot process is
5028 * done. Non-atomic initialization, single-pass.
5030 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5031 unsigned long start_pfn, enum memmap_context context)
5033 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5034 unsigned long end_pfn = start_pfn + size;
5035 pg_data_t *pgdat = NODE_DATA(nid);
5037 unsigned long nr_initialised = 0;
5038 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5039 struct memblock_region *r = NULL, *tmp;
5042 if (highest_memmap_pfn < end_pfn - 1)
5043 highest_memmap_pfn = end_pfn - 1;
5046 * Honor reservation requested by the driver for this ZONE_DEVICE
5049 if (altmap && start_pfn == altmap->base_pfn)
5050 start_pfn += altmap->reserve;
5052 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5054 * There can be holes in boot-time mem_map[]s handed to this
5055 * function. They do not exist on hotplugged memory.
5057 if (context != MEMMAP_EARLY)
5060 if (!early_pfn_valid(pfn))
5062 if (!early_pfn_in_nid(pfn, nid))
5064 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5067 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5069 * Check given memblock attribute by firmware which can affect
5070 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5071 * mirrored, it's an overlapped memmap init. skip it.
5073 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5074 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5075 for_each_memblock(memory, tmp)
5076 if (pfn < memblock_region_memory_end_pfn(tmp))
5080 if (pfn >= memblock_region_memory_base_pfn(r) &&
5081 memblock_is_mirror(r)) {
5082 /* already initialized as NORMAL */
5083 pfn = memblock_region_memory_end_pfn(r);
5091 * Mark the block movable so that blocks are reserved for
5092 * movable at startup. This will force kernel allocations
5093 * to reserve their blocks rather than leaking throughout
5094 * the address space during boot when many long-lived
5095 * kernel allocations are made.
5097 * bitmap is created for zone's valid pfn range. but memmap
5098 * can be created for invalid pages (for alignment)
5099 * check here not to call set_pageblock_migratetype() against
5102 if (!(pfn & (pageblock_nr_pages - 1))) {
5103 struct page *page = pfn_to_page(pfn);
5105 __init_single_page(page, pfn, zone, nid);
5106 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5108 __init_single_pfn(pfn, zone, nid);
5113 static void __meminit zone_init_free_lists(struct zone *zone)
5115 unsigned int order, t;
5116 for_each_migratetype_order(order, t) {
5117 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5118 zone->free_area[order].nr_free = 0;
5122 #ifndef __HAVE_ARCH_MEMMAP_INIT
5123 #define memmap_init(size, nid, zone, start_pfn) \
5124 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5127 static int zone_batchsize(struct zone *zone)
5133 * The per-cpu-pages pools are set to around 1000th of the
5134 * size of the zone. But no more than 1/2 of a meg.
5136 * OK, so we don't know how big the cache is. So guess.
5138 batch = zone->managed_pages / 1024;
5139 if (batch * PAGE_SIZE > 512 * 1024)
5140 batch = (512 * 1024) / PAGE_SIZE;
5141 batch /= 4; /* We effectively *= 4 below */
5146 * Clamp the batch to a 2^n - 1 value. Having a power
5147 * of 2 value was found to be more likely to have
5148 * suboptimal cache aliasing properties in some cases.
5150 * For example if 2 tasks are alternately allocating
5151 * batches of pages, one task can end up with a lot
5152 * of pages of one half of the possible page colors
5153 * and the other with pages of the other colors.
5155 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5160 /* The deferral and batching of frees should be suppressed under NOMMU
5163 * The problem is that NOMMU needs to be able to allocate large chunks
5164 * of contiguous memory as there's no hardware page translation to
5165 * assemble apparent contiguous memory from discontiguous pages.
5167 * Queueing large contiguous runs of pages for batching, however,
5168 * causes the pages to actually be freed in smaller chunks. As there
5169 * can be a significant delay between the individual batches being
5170 * recycled, this leads to the once large chunks of space being
5171 * fragmented and becoming unavailable for high-order allocations.
5178 * pcp->high and pcp->batch values are related and dependent on one another:
5179 * ->batch must never be higher then ->high.
5180 * The following function updates them in a safe manner without read side
5183 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5184 * those fields changing asynchronously (acording the the above rule).
5186 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5187 * outside of boot time (or some other assurance that no concurrent updaters
5190 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5191 unsigned long batch)
5193 /* start with a fail safe value for batch */
5197 /* Update high, then batch, in order */
5204 /* a companion to pageset_set_high() */
5205 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5207 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5210 static void pageset_init(struct per_cpu_pageset *p)
5212 struct per_cpu_pages *pcp;
5215 memset(p, 0, sizeof(*p));
5219 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5220 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5223 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5226 pageset_set_batch(p, batch);
5230 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5231 * to the value high for the pageset p.
5233 static void pageset_set_high(struct per_cpu_pageset *p,
5236 unsigned long batch = max(1UL, high / 4);
5237 if ((high / 4) > (PAGE_SHIFT * 8))
5238 batch = PAGE_SHIFT * 8;
5240 pageset_update(&p->pcp, high, batch);
5243 static void pageset_set_high_and_batch(struct zone *zone,
5244 struct per_cpu_pageset *pcp)
5246 if (percpu_pagelist_fraction)
5247 pageset_set_high(pcp,
5248 (zone->managed_pages /
5249 percpu_pagelist_fraction));
5251 pageset_set_batch(pcp, zone_batchsize(zone));
5254 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5256 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5259 pageset_set_high_and_batch(zone, pcp);
5262 static void __meminit setup_zone_pageset(struct zone *zone)
5265 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5266 for_each_possible_cpu(cpu)
5267 zone_pageset_init(zone, cpu);
5271 * Allocate per cpu pagesets and initialize them.
5272 * Before this call only boot pagesets were available.
5274 void __init setup_per_cpu_pageset(void)
5276 struct pglist_data *pgdat;
5279 for_each_populated_zone(zone)
5280 setup_zone_pageset(zone);
5282 for_each_online_pgdat(pgdat)
5283 pgdat->per_cpu_nodestats =
5284 alloc_percpu(struct per_cpu_nodestat);
5287 static noinline __ref
5288 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5294 * The per-page waitqueue mechanism uses hashed waitqueues
5297 zone->wait_table_hash_nr_entries =
5298 wait_table_hash_nr_entries(zone_size_pages);
5299 zone->wait_table_bits =
5300 wait_table_bits(zone->wait_table_hash_nr_entries);
5301 alloc_size = zone->wait_table_hash_nr_entries
5302 * sizeof(wait_queue_head_t);
5304 if (!slab_is_available()) {
5305 zone->wait_table = (wait_queue_head_t *)
5306 memblock_virt_alloc_node_nopanic(
5307 alloc_size, zone->zone_pgdat->node_id);
5310 * This case means that a zone whose size was 0 gets new memory
5311 * via memory hot-add.
5312 * But it may be the case that a new node was hot-added. In
5313 * this case vmalloc() will not be able to use this new node's
5314 * memory - this wait_table must be initialized to use this new
5315 * node itself as well.
5316 * To use this new node's memory, further consideration will be
5319 zone->wait_table = vmalloc(alloc_size);
5321 if (!zone->wait_table)
5324 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5325 init_waitqueue_head(zone->wait_table + i);
5330 static __meminit void zone_pcp_init(struct zone *zone)
5333 * per cpu subsystem is not up at this point. The following code
5334 * relies on the ability of the linker to provide the
5335 * offset of a (static) per cpu variable into the per cpu area.
5337 zone->pageset = &boot_pageset;
5339 if (populated_zone(zone))
5340 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5341 zone->name, zone->present_pages,
5342 zone_batchsize(zone));
5345 int __meminit init_currently_empty_zone(struct zone *zone,
5346 unsigned long zone_start_pfn,
5349 struct pglist_data *pgdat = zone->zone_pgdat;
5351 ret = zone_wait_table_init(zone, size);
5354 pgdat->nr_zones = zone_idx(zone) + 1;
5356 zone->zone_start_pfn = zone_start_pfn;
5358 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5359 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5361 (unsigned long)zone_idx(zone),
5362 zone_start_pfn, (zone_start_pfn + size));
5364 zone_init_free_lists(zone);
5369 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5370 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5373 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5375 int __meminit __early_pfn_to_nid(unsigned long pfn,
5376 struct mminit_pfnnid_cache *state)
5378 unsigned long start_pfn, end_pfn;
5381 if (state->last_start <= pfn && pfn < state->last_end)
5382 return state->last_nid;
5384 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5386 state->last_start = start_pfn;
5387 state->last_end = end_pfn;
5388 state->last_nid = nid;
5393 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5396 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5397 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5398 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5400 * If an architecture guarantees that all ranges registered contain no holes
5401 * and may be freed, this this function may be used instead of calling
5402 * memblock_free_early_nid() manually.
5404 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5406 unsigned long start_pfn, end_pfn;
5409 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5410 start_pfn = min(start_pfn, max_low_pfn);
5411 end_pfn = min(end_pfn, max_low_pfn);
5413 if (start_pfn < end_pfn)
5414 memblock_free_early_nid(PFN_PHYS(start_pfn),
5415 (end_pfn - start_pfn) << PAGE_SHIFT,
5421 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5422 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5424 * If an architecture guarantees that all ranges registered contain no holes and may
5425 * be freed, this function may be used instead of calling memory_present() manually.
5427 void __init sparse_memory_present_with_active_regions(int nid)
5429 unsigned long start_pfn, end_pfn;
5432 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5433 memory_present(this_nid, start_pfn, end_pfn);
5437 * get_pfn_range_for_nid - Return the start and end page frames for a node
5438 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5439 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5440 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5442 * It returns the start and end page frame of a node based on information
5443 * provided by memblock_set_node(). If called for a node
5444 * with no available memory, a warning is printed and the start and end
5447 void __meminit get_pfn_range_for_nid(unsigned int nid,
5448 unsigned long *start_pfn, unsigned long *end_pfn)
5450 unsigned long this_start_pfn, this_end_pfn;
5456 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5457 *start_pfn = min(*start_pfn, this_start_pfn);
5458 *end_pfn = max(*end_pfn, this_end_pfn);
5461 if (*start_pfn == -1UL)
5466 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5467 * assumption is made that zones within a node are ordered in monotonic
5468 * increasing memory addresses so that the "highest" populated zone is used
5470 static void __init find_usable_zone_for_movable(void)
5473 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5474 if (zone_index == ZONE_MOVABLE)
5477 if (arch_zone_highest_possible_pfn[zone_index] >
5478 arch_zone_lowest_possible_pfn[zone_index])
5482 VM_BUG_ON(zone_index == -1);
5483 movable_zone = zone_index;
5487 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5488 * because it is sized independent of architecture. Unlike the other zones,
5489 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5490 * in each node depending on the size of each node and how evenly kernelcore
5491 * is distributed. This helper function adjusts the zone ranges
5492 * provided by the architecture for a given node by using the end of the
5493 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5494 * zones within a node are in order of monotonic increases memory addresses
5496 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5497 unsigned long zone_type,
5498 unsigned long node_start_pfn,
5499 unsigned long node_end_pfn,
5500 unsigned long *zone_start_pfn,
5501 unsigned long *zone_end_pfn)
5503 /* Only adjust if ZONE_MOVABLE is on this node */
5504 if (zone_movable_pfn[nid]) {
5505 /* Size ZONE_MOVABLE */
5506 if (zone_type == ZONE_MOVABLE) {
5507 *zone_start_pfn = zone_movable_pfn[nid];
5508 *zone_end_pfn = min(node_end_pfn,
5509 arch_zone_highest_possible_pfn[movable_zone]);
5511 /* Adjust for ZONE_MOVABLE starting within this range */
5512 } else if (!mirrored_kernelcore &&
5513 *zone_start_pfn < zone_movable_pfn[nid] &&
5514 *zone_end_pfn > zone_movable_pfn[nid]) {
5515 *zone_end_pfn = zone_movable_pfn[nid];
5517 /* Check if this whole range is within ZONE_MOVABLE */
5518 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5519 *zone_start_pfn = *zone_end_pfn;
5524 * Return the number of pages a zone spans in a node, including holes
5525 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5527 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5528 unsigned long zone_type,
5529 unsigned long node_start_pfn,
5530 unsigned long node_end_pfn,
5531 unsigned long *zone_start_pfn,
5532 unsigned long *zone_end_pfn,
5533 unsigned long *ignored)
5535 /* When hotadd a new node from cpu_up(), the node should be empty */
5536 if (!node_start_pfn && !node_end_pfn)
5539 /* Get the start and end of the zone */
5540 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5541 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5542 adjust_zone_range_for_zone_movable(nid, zone_type,
5543 node_start_pfn, node_end_pfn,
5544 zone_start_pfn, zone_end_pfn);
5546 /* Check that this node has pages within the zone's required range */
5547 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5550 /* Move the zone boundaries inside the node if necessary */
5551 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5552 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5554 /* Return the spanned pages */
5555 return *zone_end_pfn - *zone_start_pfn;
5559 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5560 * then all holes in the requested range will be accounted for.
5562 unsigned long __meminit __absent_pages_in_range(int nid,
5563 unsigned long range_start_pfn,
5564 unsigned long range_end_pfn)
5566 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5567 unsigned long start_pfn, end_pfn;
5570 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5571 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5572 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5573 nr_absent -= end_pfn - start_pfn;
5579 * absent_pages_in_range - Return number of page frames in holes within a range
5580 * @start_pfn: The start PFN to start searching for holes
5581 * @end_pfn: The end PFN to stop searching for holes
5583 * It returns the number of pages frames in memory holes within a range.
5585 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5586 unsigned long end_pfn)
5588 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5591 /* Return the number of page frames in holes in a zone on a node */
5592 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5593 unsigned long zone_type,
5594 unsigned long node_start_pfn,
5595 unsigned long node_end_pfn,
5596 unsigned long *ignored)
5598 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5599 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5600 unsigned long zone_start_pfn, zone_end_pfn;
5601 unsigned long nr_absent;
5603 /* When hotadd a new node from cpu_up(), the node should be empty */
5604 if (!node_start_pfn && !node_end_pfn)
5607 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5608 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5610 adjust_zone_range_for_zone_movable(nid, zone_type,
5611 node_start_pfn, node_end_pfn,
5612 &zone_start_pfn, &zone_end_pfn);
5613 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5616 * ZONE_MOVABLE handling.
5617 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5620 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5621 unsigned long start_pfn, end_pfn;
5622 struct memblock_region *r;
5624 for_each_memblock(memory, r) {
5625 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5626 zone_start_pfn, zone_end_pfn);
5627 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5628 zone_start_pfn, zone_end_pfn);
5630 if (zone_type == ZONE_MOVABLE &&
5631 memblock_is_mirror(r))
5632 nr_absent += end_pfn - start_pfn;
5634 if (zone_type == ZONE_NORMAL &&
5635 !memblock_is_mirror(r))
5636 nr_absent += end_pfn - start_pfn;
5643 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5644 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5645 unsigned long zone_type,
5646 unsigned long node_start_pfn,
5647 unsigned long node_end_pfn,
5648 unsigned long *zone_start_pfn,
5649 unsigned long *zone_end_pfn,
5650 unsigned long *zones_size)
5654 *zone_start_pfn = node_start_pfn;
5655 for (zone = 0; zone < zone_type; zone++)
5656 *zone_start_pfn += zones_size[zone];
5658 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5660 return zones_size[zone_type];
5663 static inline unsigned long __meminit zone_absent_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 *zholes_size)
5672 return zholes_size[zone_type];
5675 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5677 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5678 unsigned long node_start_pfn,
5679 unsigned long node_end_pfn,
5680 unsigned long *zones_size,
5681 unsigned long *zholes_size)
5683 unsigned long realtotalpages = 0, totalpages = 0;
5686 for (i = 0; i < MAX_NR_ZONES; i++) {
5687 struct zone *zone = pgdat->node_zones + i;
5688 unsigned long zone_start_pfn, zone_end_pfn;
5689 unsigned long size, real_size;
5691 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5697 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5698 node_start_pfn, node_end_pfn,
5701 zone->zone_start_pfn = zone_start_pfn;
5703 zone->zone_start_pfn = 0;
5704 zone->spanned_pages = size;
5705 zone->present_pages = real_size;
5708 realtotalpages += real_size;
5711 pgdat->node_spanned_pages = totalpages;
5712 pgdat->node_present_pages = realtotalpages;
5713 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5717 #ifndef CONFIG_SPARSEMEM
5719 * Calculate the size of the zone->blockflags rounded to an unsigned long
5720 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5721 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5722 * round what is now in bits to nearest long in bits, then return it in
5725 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5727 unsigned long usemapsize;
5729 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5730 usemapsize = roundup(zonesize, pageblock_nr_pages);
5731 usemapsize = usemapsize >> pageblock_order;
5732 usemapsize *= NR_PAGEBLOCK_BITS;
5733 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5735 return usemapsize / 8;
5738 static void __init setup_usemap(struct pglist_data *pgdat,
5740 unsigned long zone_start_pfn,
5741 unsigned long zonesize)
5743 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5744 zone->pageblock_flags = NULL;
5746 zone->pageblock_flags =
5747 memblock_virt_alloc_node_nopanic(usemapsize,
5751 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5752 unsigned long zone_start_pfn, unsigned long zonesize) {}
5753 #endif /* CONFIG_SPARSEMEM */
5755 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5757 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5758 void __paginginit set_pageblock_order(void)
5762 /* Check that pageblock_nr_pages has not already been setup */
5763 if (pageblock_order)
5766 if (HPAGE_SHIFT > PAGE_SHIFT)
5767 order = HUGETLB_PAGE_ORDER;
5769 order = MAX_ORDER - 1;
5772 * Assume the largest contiguous order of interest is a huge page.
5773 * This value may be variable depending on boot parameters on IA64 and
5776 pageblock_order = order;
5778 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5781 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5782 * is unused as pageblock_order is set at compile-time. See
5783 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5786 void __paginginit set_pageblock_order(void)
5790 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5792 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5793 unsigned long present_pages)
5795 unsigned long pages = spanned_pages;
5798 * Provide a more accurate estimation if there are holes within
5799 * the zone and SPARSEMEM is in use. If there are holes within the
5800 * zone, each populated memory region may cost us one or two extra
5801 * memmap pages due to alignment because memmap pages for each
5802 * populated regions may not naturally algined on page boundary.
5803 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5805 if (spanned_pages > present_pages + (present_pages >> 4) &&
5806 IS_ENABLED(CONFIG_SPARSEMEM))
5807 pages = present_pages;
5809 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5813 * Set up the zone data structures:
5814 * - mark all pages reserved
5815 * - mark all memory queues empty
5816 * - clear the memory bitmaps
5818 * NOTE: pgdat should get zeroed by caller.
5820 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5823 int nid = pgdat->node_id;
5826 pgdat_resize_init(pgdat);
5827 #ifdef CONFIG_NUMA_BALANCING
5828 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5829 pgdat->numabalancing_migrate_nr_pages = 0;
5830 pgdat->numabalancing_migrate_next_window = jiffies;
5832 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5833 spin_lock_init(&pgdat->split_queue_lock);
5834 INIT_LIST_HEAD(&pgdat->split_queue);
5835 pgdat->split_queue_len = 0;
5837 init_waitqueue_head(&pgdat->kswapd_wait);
5838 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5839 #ifdef CONFIG_COMPACTION
5840 init_waitqueue_head(&pgdat->kcompactd_wait);
5842 pgdat_page_ext_init(pgdat);
5843 spin_lock_init(&pgdat->lru_lock);
5844 lruvec_init(node_lruvec(pgdat));
5846 for (j = 0; j < MAX_NR_ZONES; j++) {
5847 struct zone *zone = pgdat->node_zones + j;
5848 unsigned long size, realsize, freesize, memmap_pages;
5849 unsigned long zone_start_pfn = zone->zone_start_pfn;
5851 size = zone->spanned_pages;
5852 realsize = freesize = zone->present_pages;
5855 * Adjust freesize so that it accounts for how much memory
5856 * is used by this zone for memmap. This affects the watermark
5857 * and per-cpu initialisations
5859 memmap_pages = calc_memmap_size(size, realsize);
5860 if (!is_highmem_idx(j)) {
5861 if (freesize >= memmap_pages) {
5862 freesize -= memmap_pages;
5865 " %s zone: %lu pages used for memmap\n",
5866 zone_names[j], memmap_pages);
5868 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5869 zone_names[j], memmap_pages, freesize);
5872 /* Account for reserved pages */
5873 if (j == 0 && freesize > dma_reserve) {
5874 freesize -= dma_reserve;
5875 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5876 zone_names[0], dma_reserve);
5879 if (!is_highmem_idx(j))
5880 nr_kernel_pages += freesize;
5881 /* Charge for highmem memmap if there are enough kernel pages */
5882 else if (nr_kernel_pages > memmap_pages * 2)
5883 nr_kernel_pages -= memmap_pages;
5884 nr_all_pages += freesize;
5887 * Set an approximate value for lowmem here, it will be adjusted
5888 * when the bootmem allocator frees pages into the buddy system.
5889 * And all highmem pages will be managed by the buddy system.
5891 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5895 zone->name = zone_names[j];
5896 zone->zone_pgdat = pgdat;
5897 spin_lock_init(&zone->lock);
5898 zone_seqlock_init(zone);
5899 zone_pcp_init(zone);
5904 set_pageblock_order();
5905 setup_usemap(pgdat, zone, zone_start_pfn, size);
5906 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5908 memmap_init(size, nid, j, zone_start_pfn);
5912 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5914 unsigned long __maybe_unused start = 0;
5915 unsigned long __maybe_unused offset = 0;
5917 /* Skip empty nodes */
5918 if (!pgdat->node_spanned_pages)
5921 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5922 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5923 offset = pgdat->node_start_pfn - start;
5924 /* ia64 gets its own node_mem_map, before this, without bootmem */
5925 if (!pgdat->node_mem_map) {
5926 unsigned long size, end;
5930 * The zone's endpoints aren't required to be MAX_ORDER
5931 * aligned but the node_mem_map endpoints must be in order
5932 * for the buddy allocator to function correctly.
5934 end = pgdat_end_pfn(pgdat);
5935 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5936 size = (end - start) * sizeof(struct page);
5937 map = alloc_remap(pgdat->node_id, size);
5939 map = memblock_virt_alloc_node_nopanic(size,
5941 pgdat->node_mem_map = map + offset;
5943 #ifndef CONFIG_NEED_MULTIPLE_NODES
5945 * With no DISCONTIG, the global mem_map is just set as node 0's
5947 if (pgdat == NODE_DATA(0)) {
5948 mem_map = NODE_DATA(0)->node_mem_map;
5949 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5950 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5952 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5955 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5958 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5959 unsigned long node_start_pfn, unsigned long *zholes_size)
5961 pg_data_t *pgdat = NODE_DATA(nid);
5962 unsigned long start_pfn = 0;
5963 unsigned long end_pfn = 0;
5965 /* pg_data_t should be reset to zero when it's allocated */
5966 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5968 reset_deferred_meminit(pgdat);
5969 pgdat->node_id = nid;
5970 pgdat->node_start_pfn = node_start_pfn;
5971 pgdat->per_cpu_nodestats = NULL;
5972 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5973 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5974 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5975 (u64)start_pfn << PAGE_SHIFT,
5976 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5978 start_pfn = node_start_pfn;
5980 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5981 zones_size, zholes_size);
5983 alloc_node_mem_map(pgdat);
5984 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5985 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5986 nid, (unsigned long)pgdat,
5987 (unsigned long)pgdat->node_mem_map);
5990 free_area_init_core(pgdat);
5993 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5995 #if MAX_NUMNODES > 1
5997 * Figure out the number of possible node ids.
5999 void __init setup_nr_node_ids(void)
6001 unsigned int highest;
6003 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6004 nr_node_ids = highest + 1;
6009 * node_map_pfn_alignment - determine the maximum internode alignment
6011 * This function should be called after node map is populated and sorted.
6012 * It calculates the maximum power of two alignment which can distinguish
6015 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6016 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6017 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6018 * shifted, 1GiB is enough and this function will indicate so.
6020 * This is used to test whether pfn -> nid mapping of the chosen memory
6021 * model has fine enough granularity to avoid incorrect mapping for the
6022 * populated node map.
6024 * Returns the determined alignment in pfn's. 0 if there is no alignment
6025 * requirement (single node).
6027 unsigned long __init node_map_pfn_alignment(void)
6029 unsigned long accl_mask = 0, last_end = 0;
6030 unsigned long start, end, mask;
6034 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6035 if (!start || last_nid < 0 || last_nid == nid) {
6042 * Start with a mask granular enough to pin-point to the
6043 * start pfn and tick off bits one-by-one until it becomes
6044 * too coarse to separate the current node from the last.
6046 mask = ~((1 << __ffs(start)) - 1);
6047 while (mask && last_end <= (start & (mask << 1)))
6050 /* accumulate all internode masks */
6054 /* convert mask to number of pages */
6055 return ~accl_mask + 1;
6058 /* Find the lowest pfn for a node */
6059 static unsigned long __init find_min_pfn_for_node(int nid)
6061 unsigned long min_pfn = ULONG_MAX;
6062 unsigned long start_pfn;
6065 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6066 min_pfn = min(min_pfn, start_pfn);
6068 if (min_pfn == ULONG_MAX) {
6069 pr_warn("Could not find start_pfn for node %d\n", nid);
6077 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6079 * It returns the minimum PFN based on information provided via
6080 * memblock_set_node().
6082 unsigned long __init find_min_pfn_with_active_regions(void)
6084 return find_min_pfn_for_node(MAX_NUMNODES);
6088 * early_calculate_totalpages()
6089 * Sum pages in active regions for movable zone.
6090 * Populate N_MEMORY for calculating usable_nodes.
6092 static unsigned long __init early_calculate_totalpages(void)
6094 unsigned long totalpages = 0;
6095 unsigned long start_pfn, end_pfn;
6098 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6099 unsigned long pages = end_pfn - start_pfn;
6101 totalpages += pages;
6103 node_set_state(nid, N_MEMORY);
6109 * Find the PFN the Movable zone begins in each node. Kernel memory
6110 * is spread evenly between nodes as long as the nodes have enough
6111 * memory. When they don't, some nodes will have more kernelcore than
6114 static void __init find_zone_movable_pfns_for_nodes(void)
6117 unsigned long usable_startpfn;
6118 unsigned long kernelcore_node, kernelcore_remaining;
6119 /* save the state before borrow the nodemask */
6120 nodemask_t saved_node_state = node_states[N_MEMORY];
6121 unsigned long totalpages = early_calculate_totalpages();
6122 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6123 struct memblock_region *r;
6125 /* Need to find movable_zone earlier when movable_node is specified. */
6126 find_usable_zone_for_movable();
6129 * If movable_node is specified, ignore kernelcore and movablecore
6132 if (movable_node_is_enabled()) {
6133 for_each_memblock(memory, r) {
6134 if (!memblock_is_hotpluggable(r))
6139 usable_startpfn = PFN_DOWN(r->base);
6140 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6141 min(usable_startpfn, zone_movable_pfn[nid]) :
6149 * If kernelcore=mirror is specified, ignore movablecore option
6151 if (mirrored_kernelcore) {
6152 bool mem_below_4gb_not_mirrored = false;
6154 for_each_memblock(memory, r) {
6155 if (memblock_is_mirror(r))
6160 usable_startpfn = memblock_region_memory_base_pfn(r);
6162 if (usable_startpfn < 0x100000) {
6163 mem_below_4gb_not_mirrored = true;
6167 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6168 min(usable_startpfn, zone_movable_pfn[nid]) :
6172 if (mem_below_4gb_not_mirrored)
6173 pr_warn("This configuration results in unmirrored kernel memory.");
6179 * If movablecore=nn[KMG] was specified, calculate what size of
6180 * kernelcore that corresponds so that memory usable for
6181 * any allocation type is evenly spread. If both kernelcore
6182 * and movablecore are specified, then the value of kernelcore
6183 * will be used for required_kernelcore if it's greater than
6184 * what movablecore would have allowed.
6186 if (required_movablecore) {
6187 unsigned long corepages;
6190 * Round-up so that ZONE_MOVABLE is at least as large as what
6191 * was requested by the user
6193 required_movablecore =
6194 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6195 required_movablecore = min(totalpages, required_movablecore);
6196 corepages = totalpages - required_movablecore;
6198 required_kernelcore = max(required_kernelcore, corepages);
6202 * If kernelcore was not specified or kernelcore size is larger
6203 * than totalpages, there is no ZONE_MOVABLE.
6205 if (!required_kernelcore || required_kernelcore >= totalpages)
6208 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6209 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6212 /* Spread kernelcore memory as evenly as possible throughout nodes */
6213 kernelcore_node = required_kernelcore / usable_nodes;
6214 for_each_node_state(nid, N_MEMORY) {
6215 unsigned long start_pfn, end_pfn;
6218 * Recalculate kernelcore_node if the division per node
6219 * now exceeds what is necessary to satisfy the requested
6220 * amount of memory for the kernel
6222 if (required_kernelcore < kernelcore_node)
6223 kernelcore_node = required_kernelcore / usable_nodes;
6226 * As the map is walked, we track how much memory is usable
6227 * by the kernel using kernelcore_remaining. When it is
6228 * 0, the rest of the node is usable by ZONE_MOVABLE
6230 kernelcore_remaining = kernelcore_node;
6232 /* Go through each range of PFNs within this node */
6233 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6234 unsigned long size_pages;
6236 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6237 if (start_pfn >= end_pfn)
6240 /* Account for what is only usable for kernelcore */
6241 if (start_pfn < usable_startpfn) {
6242 unsigned long kernel_pages;
6243 kernel_pages = min(end_pfn, usable_startpfn)
6246 kernelcore_remaining -= min(kernel_pages,
6247 kernelcore_remaining);
6248 required_kernelcore -= min(kernel_pages,
6249 required_kernelcore);
6251 /* Continue if range is now fully accounted */
6252 if (end_pfn <= usable_startpfn) {
6255 * Push zone_movable_pfn to the end so
6256 * that if we have to rebalance
6257 * kernelcore across nodes, we will
6258 * not double account here
6260 zone_movable_pfn[nid] = end_pfn;
6263 start_pfn = usable_startpfn;
6267 * The usable PFN range for ZONE_MOVABLE is from
6268 * start_pfn->end_pfn. Calculate size_pages as the
6269 * number of pages used as kernelcore
6271 size_pages = end_pfn - start_pfn;
6272 if (size_pages > kernelcore_remaining)
6273 size_pages = kernelcore_remaining;
6274 zone_movable_pfn[nid] = start_pfn + size_pages;
6277 * Some kernelcore has been met, update counts and
6278 * break if the kernelcore for this node has been
6281 required_kernelcore -= min(required_kernelcore,
6283 kernelcore_remaining -= size_pages;
6284 if (!kernelcore_remaining)
6290 * If there is still required_kernelcore, we do another pass with one
6291 * less node in the count. This will push zone_movable_pfn[nid] further
6292 * along on the nodes that still have memory until kernelcore is
6296 if (usable_nodes && required_kernelcore > usable_nodes)
6300 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6301 for (nid = 0; nid < MAX_NUMNODES; nid++)
6302 zone_movable_pfn[nid] =
6303 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6306 /* restore the node_state */
6307 node_states[N_MEMORY] = saved_node_state;
6310 /* Any regular or high memory on that node ? */
6311 static void check_for_memory(pg_data_t *pgdat, int nid)
6313 enum zone_type zone_type;
6315 if (N_MEMORY == N_NORMAL_MEMORY)
6318 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6319 struct zone *zone = &pgdat->node_zones[zone_type];
6320 if (populated_zone(zone)) {
6321 node_set_state(nid, N_HIGH_MEMORY);
6322 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6323 zone_type <= ZONE_NORMAL)
6324 node_set_state(nid, N_NORMAL_MEMORY);
6331 * free_area_init_nodes - Initialise all pg_data_t and zone data
6332 * @max_zone_pfn: an array of max PFNs for each zone
6334 * This will call free_area_init_node() for each active node in the system.
6335 * Using the page ranges provided by memblock_set_node(), the size of each
6336 * zone in each node and their holes is calculated. If the maximum PFN
6337 * between two adjacent zones match, it is assumed that the zone is empty.
6338 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6339 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6340 * starts where the previous one ended. For example, ZONE_DMA32 starts
6341 * at arch_max_dma_pfn.
6343 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6345 unsigned long start_pfn, end_pfn;
6348 /* Record where the zone boundaries are */
6349 memset(arch_zone_lowest_possible_pfn, 0,
6350 sizeof(arch_zone_lowest_possible_pfn));
6351 memset(arch_zone_highest_possible_pfn, 0,
6352 sizeof(arch_zone_highest_possible_pfn));
6354 start_pfn = find_min_pfn_with_active_regions();
6356 for (i = 0; i < MAX_NR_ZONES; i++) {
6357 if (i == ZONE_MOVABLE)
6360 end_pfn = max(max_zone_pfn[i], start_pfn);
6361 arch_zone_lowest_possible_pfn[i] = start_pfn;
6362 arch_zone_highest_possible_pfn[i] = end_pfn;
6364 start_pfn = end_pfn;
6366 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6367 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6369 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6370 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6371 find_zone_movable_pfns_for_nodes();
6373 /* Print out the zone ranges */
6374 pr_info("Zone ranges:\n");
6375 for (i = 0; i < MAX_NR_ZONES; i++) {
6376 if (i == ZONE_MOVABLE)
6378 pr_info(" %-8s ", zone_names[i]);
6379 if (arch_zone_lowest_possible_pfn[i] ==
6380 arch_zone_highest_possible_pfn[i])
6383 pr_cont("[mem %#018Lx-%#018Lx]\n",
6384 (u64)arch_zone_lowest_possible_pfn[i]
6386 ((u64)arch_zone_highest_possible_pfn[i]
6387 << PAGE_SHIFT) - 1);
6390 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6391 pr_info("Movable zone start for each node\n");
6392 for (i = 0; i < MAX_NUMNODES; i++) {
6393 if (zone_movable_pfn[i])
6394 pr_info(" Node %d: %#018Lx\n", i,
6395 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6398 /* Print out the early node map */
6399 pr_info("Early memory node ranges\n");
6400 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6401 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6402 (u64)start_pfn << PAGE_SHIFT,
6403 ((u64)end_pfn << PAGE_SHIFT) - 1);
6405 /* Initialise every node */
6406 mminit_verify_pageflags_layout();
6407 setup_nr_node_ids();
6408 for_each_online_node(nid) {
6409 pg_data_t *pgdat = NODE_DATA(nid);
6410 free_area_init_node(nid, NULL,
6411 find_min_pfn_for_node(nid), NULL);
6413 /* Any memory on that node */
6414 if (pgdat->node_present_pages)
6415 node_set_state(nid, N_MEMORY);
6416 check_for_memory(pgdat, nid);
6420 static int __init cmdline_parse_core(char *p, unsigned long *core)
6422 unsigned long long coremem;
6426 coremem = memparse(p, &p);
6427 *core = coremem >> PAGE_SHIFT;
6429 /* Paranoid check that UL is enough for the coremem value */
6430 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6436 * kernelcore=size sets the amount of memory for use for allocations that
6437 * cannot be reclaimed or migrated.
6439 static int __init cmdline_parse_kernelcore(char *p)
6441 /* parse kernelcore=mirror */
6442 if (parse_option_str(p, "mirror")) {
6443 mirrored_kernelcore = true;
6447 return cmdline_parse_core(p, &required_kernelcore);
6451 * movablecore=size sets the amount of memory for use for allocations that
6452 * can be reclaimed or migrated.
6454 static int __init cmdline_parse_movablecore(char *p)
6456 return cmdline_parse_core(p, &required_movablecore);
6459 early_param("kernelcore", cmdline_parse_kernelcore);
6460 early_param("movablecore", cmdline_parse_movablecore);
6462 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6464 void adjust_managed_page_count(struct page *page, long count)
6466 spin_lock(&managed_page_count_lock);
6467 page_zone(page)->managed_pages += count;
6468 totalram_pages += count;
6469 #ifdef CONFIG_HIGHMEM
6470 if (PageHighMem(page))
6471 totalhigh_pages += count;
6473 spin_unlock(&managed_page_count_lock);
6475 EXPORT_SYMBOL(adjust_managed_page_count);
6477 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6480 unsigned long pages = 0;
6482 start = (void *)PAGE_ALIGN((unsigned long)start);
6483 end = (void *)((unsigned long)end & PAGE_MASK);
6484 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6485 if ((unsigned int)poison <= 0xFF)
6486 memset(pos, poison, PAGE_SIZE);
6487 free_reserved_page(virt_to_page(pos));
6491 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6492 s, pages << (PAGE_SHIFT - 10), start, end);
6496 EXPORT_SYMBOL(free_reserved_area);
6498 #ifdef CONFIG_HIGHMEM
6499 void free_highmem_page(struct page *page)
6501 __free_reserved_page(page);
6503 page_zone(page)->managed_pages++;
6509 void __init mem_init_print_info(const char *str)
6511 unsigned long physpages, codesize, datasize, rosize, bss_size;
6512 unsigned long init_code_size, init_data_size;
6514 physpages = get_num_physpages();
6515 codesize = _etext - _stext;
6516 datasize = _edata - _sdata;
6517 rosize = __end_rodata - __start_rodata;
6518 bss_size = __bss_stop - __bss_start;
6519 init_data_size = __init_end - __init_begin;
6520 init_code_size = _einittext - _sinittext;
6523 * Detect special cases and adjust section sizes accordingly:
6524 * 1) .init.* may be embedded into .data sections
6525 * 2) .init.text.* may be out of [__init_begin, __init_end],
6526 * please refer to arch/tile/kernel/vmlinux.lds.S.
6527 * 3) .rodata.* may be embedded into .text or .data sections.
6529 #define adj_init_size(start, end, size, pos, adj) \
6531 if (start <= pos && pos < end && size > adj) \
6535 adj_init_size(__init_begin, __init_end, init_data_size,
6536 _sinittext, init_code_size);
6537 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6538 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6539 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6540 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6542 #undef adj_init_size
6544 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6545 #ifdef CONFIG_HIGHMEM
6549 nr_free_pages() << (PAGE_SHIFT - 10),
6550 physpages << (PAGE_SHIFT - 10),
6551 codesize >> 10, datasize >> 10, rosize >> 10,
6552 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6553 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6554 totalcma_pages << (PAGE_SHIFT - 10),
6555 #ifdef CONFIG_HIGHMEM
6556 totalhigh_pages << (PAGE_SHIFT - 10),
6558 str ? ", " : "", str ? str : "");
6562 * set_dma_reserve - set the specified number of pages reserved in the first zone
6563 * @new_dma_reserve: The number of pages to mark reserved
6565 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6566 * In the DMA zone, a significant percentage may be consumed by kernel image
6567 * and other unfreeable allocations which can skew the watermarks badly. This
6568 * function may optionally be used to account for unfreeable pages in the
6569 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6570 * smaller per-cpu batchsize.
6572 void __init set_dma_reserve(unsigned long new_dma_reserve)
6574 dma_reserve = new_dma_reserve;
6577 void __init free_area_init(unsigned long *zones_size)
6579 free_area_init_node(0, zones_size,
6580 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6583 static int page_alloc_cpu_notify(struct notifier_block *self,
6584 unsigned long action, void *hcpu)
6586 int cpu = (unsigned long)hcpu;
6588 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6589 lru_add_drain_cpu(cpu);
6593 * Spill the event counters of the dead processor
6594 * into the current processors event counters.
6595 * This artificially elevates the count of the current
6598 vm_events_fold_cpu(cpu);
6601 * Zero the differential counters of the dead processor
6602 * so that the vm statistics are consistent.
6604 * This is only okay since the processor is dead and cannot
6605 * race with what we are doing.
6607 cpu_vm_stats_fold(cpu);
6612 void __init page_alloc_init(void)
6614 hotcpu_notifier(page_alloc_cpu_notify, 0);
6618 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6619 * or min_free_kbytes changes.
6621 static void calculate_totalreserve_pages(void)
6623 struct pglist_data *pgdat;
6624 unsigned long reserve_pages = 0;
6625 enum zone_type i, j;
6627 for_each_online_pgdat(pgdat) {
6629 pgdat->totalreserve_pages = 0;
6631 for (i = 0; i < MAX_NR_ZONES; i++) {
6632 struct zone *zone = pgdat->node_zones + i;
6635 /* Find valid and maximum lowmem_reserve in the zone */
6636 for (j = i; j < MAX_NR_ZONES; j++) {
6637 if (zone->lowmem_reserve[j] > max)
6638 max = zone->lowmem_reserve[j];
6641 /* we treat the high watermark as reserved pages. */
6642 max += high_wmark_pages(zone);
6644 if (max > zone->managed_pages)
6645 max = zone->managed_pages;
6647 pgdat->totalreserve_pages += max;
6649 reserve_pages += max;
6652 totalreserve_pages = reserve_pages;
6656 * setup_per_zone_lowmem_reserve - called whenever
6657 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6658 * has a correct pages reserved value, so an adequate number of
6659 * pages are left in the zone after a successful __alloc_pages().
6661 static void setup_per_zone_lowmem_reserve(void)
6663 struct pglist_data *pgdat;
6664 enum zone_type j, idx;
6666 for_each_online_pgdat(pgdat) {
6667 for (j = 0; j < MAX_NR_ZONES; j++) {
6668 struct zone *zone = pgdat->node_zones + j;
6669 unsigned long managed_pages = zone->managed_pages;
6671 zone->lowmem_reserve[j] = 0;
6675 struct zone *lower_zone;
6679 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6680 sysctl_lowmem_reserve_ratio[idx] = 1;
6682 lower_zone = pgdat->node_zones + idx;
6683 lower_zone->lowmem_reserve[j] = managed_pages /
6684 sysctl_lowmem_reserve_ratio[idx];
6685 managed_pages += lower_zone->managed_pages;
6690 /* update totalreserve_pages */
6691 calculate_totalreserve_pages();
6694 static void __setup_per_zone_wmarks(void)
6696 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6697 unsigned long lowmem_pages = 0;
6699 unsigned long flags;
6701 /* Calculate total number of !ZONE_HIGHMEM pages */
6702 for_each_zone(zone) {
6703 if (!is_highmem(zone))
6704 lowmem_pages += zone->managed_pages;
6707 for_each_zone(zone) {
6710 spin_lock_irqsave(&zone->lock, flags);
6711 tmp = (u64)pages_min * zone->managed_pages;
6712 do_div(tmp, lowmem_pages);
6713 if (is_highmem(zone)) {
6715 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6716 * need highmem pages, so cap pages_min to a small
6719 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6720 * deltas control asynch page reclaim, and so should
6721 * not be capped for highmem.
6723 unsigned long min_pages;
6725 min_pages = zone->managed_pages / 1024;
6726 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6727 zone->watermark[WMARK_MIN] = min_pages;
6730 * If it's a lowmem zone, reserve a number of pages
6731 * proportionate to the zone's size.
6733 zone->watermark[WMARK_MIN] = tmp;
6737 * Set the kswapd watermarks distance according to the
6738 * scale factor in proportion to available memory, but
6739 * ensure a minimum size on small systems.
6741 tmp = max_t(u64, tmp >> 2,
6742 mult_frac(zone->managed_pages,
6743 watermark_scale_factor, 10000));
6745 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6746 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6748 spin_unlock_irqrestore(&zone->lock, flags);
6751 /* update totalreserve_pages */
6752 calculate_totalreserve_pages();
6756 * setup_per_zone_wmarks - called when min_free_kbytes changes
6757 * or when memory is hot-{added|removed}
6759 * Ensures that the watermark[min,low,high] values for each zone are set
6760 * correctly with respect to min_free_kbytes.
6762 void setup_per_zone_wmarks(void)
6764 mutex_lock(&zonelists_mutex);
6765 __setup_per_zone_wmarks();
6766 mutex_unlock(&zonelists_mutex);
6770 * Initialise min_free_kbytes.
6772 * For small machines we want it small (128k min). For large machines
6773 * we want it large (64MB max). But it is not linear, because network
6774 * bandwidth does not increase linearly with machine size. We use
6776 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6777 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6793 int __meminit init_per_zone_wmark_min(void)
6795 unsigned long lowmem_kbytes;
6796 int new_min_free_kbytes;
6798 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6799 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6801 if (new_min_free_kbytes > user_min_free_kbytes) {
6802 min_free_kbytes = new_min_free_kbytes;
6803 if (min_free_kbytes < 128)
6804 min_free_kbytes = 128;
6805 if (min_free_kbytes > 65536)
6806 min_free_kbytes = 65536;
6808 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6809 new_min_free_kbytes, user_min_free_kbytes);
6811 setup_per_zone_wmarks();
6812 refresh_zone_stat_thresholds();
6813 setup_per_zone_lowmem_reserve();
6816 setup_min_unmapped_ratio();
6817 setup_min_slab_ratio();
6822 core_initcall(init_per_zone_wmark_min)
6825 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6826 * that we can call two helper functions whenever min_free_kbytes
6829 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6830 void __user *buffer, size_t *length, loff_t *ppos)
6834 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6839 user_min_free_kbytes = min_free_kbytes;
6840 setup_per_zone_wmarks();
6845 int watermark_scale_factor_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 setup_per_zone_wmarks();
6861 static void setup_min_unmapped_ratio(void)
6866 for_each_online_pgdat(pgdat)
6867 pgdat->min_unmapped_pages = 0;
6870 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6871 sysctl_min_unmapped_ratio) / 100;
6875 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6876 void __user *buffer, size_t *length, loff_t *ppos)
6880 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6884 setup_min_unmapped_ratio();
6889 static void setup_min_slab_ratio(void)
6894 for_each_online_pgdat(pgdat)
6895 pgdat->min_slab_pages = 0;
6898 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6899 sysctl_min_slab_ratio) / 100;
6902 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6903 void __user *buffer, size_t *length, loff_t *ppos)
6907 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6911 setup_min_slab_ratio();
6918 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6919 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6920 * whenever sysctl_lowmem_reserve_ratio changes.
6922 * The reserve ratio obviously has absolutely no relation with the
6923 * minimum watermarks. The lowmem reserve ratio can only make sense
6924 * if in function of the boot time zone sizes.
6926 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6927 void __user *buffer, size_t *length, loff_t *ppos)
6929 proc_dointvec_minmax(table, write, buffer, length, ppos);
6930 setup_per_zone_lowmem_reserve();
6935 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6936 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6937 * pagelist can have before it gets flushed back to buddy allocator.
6939 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6940 void __user *buffer, size_t *length, loff_t *ppos)
6943 int old_percpu_pagelist_fraction;
6946 mutex_lock(&pcp_batch_high_lock);
6947 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6949 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6950 if (!write || ret < 0)
6953 /* Sanity checking to avoid pcp imbalance */
6954 if (percpu_pagelist_fraction &&
6955 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6956 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6962 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6965 for_each_populated_zone(zone) {
6968 for_each_possible_cpu(cpu)
6969 pageset_set_high_and_batch(zone,
6970 per_cpu_ptr(zone->pageset, cpu));
6973 mutex_unlock(&pcp_batch_high_lock);
6978 int hashdist = HASHDIST_DEFAULT;
6980 static int __init set_hashdist(char *str)
6984 hashdist = simple_strtoul(str, &str, 0);
6987 __setup("hashdist=", set_hashdist);
6990 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6992 * Returns the number of pages that arch has reserved but
6993 * is not known to alloc_large_system_hash().
6995 static unsigned long __init arch_reserved_kernel_pages(void)
7002 * allocate a large system hash table from bootmem
7003 * - it is assumed that the hash table must contain an exact power-of-2
7004 * quantity of entries
7005 * - limit is the number of hash buckets, not the total allocation size
7007 void *__init alloc_large_system_hash(const char *tablename,
7008 unsigned long bucketsize,
7009 unsigned long numentries,
7012 unsigned int *_hash_shift,
7013 unsigned int *_hash_mask,
7014 unsigned long low_limit,
7015 unsigned long high_limit)
7017 unsigned long long max = high_limit;
7018 unsigned long log2qty, size;
7021 /* allow the kernel cmdline to have a say */
7023 /* round applicable memory size up to nearest megabyte */
7024 numentries = nr_kernel_pages;
7025 numentries -= arch_reserved_kernel_pages();
7027 /* It isn't necessary when PAGE_SIZE >= 1MB */
7028 if (PAGE_SHIFT < 20)
7029 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7031 /* limit to 1 bucket per 2^scale bytes of low memory */
7032 if (scale > PAGE_SHIFT)
7033 numentries >>= (scale - PAGE_SHIFT);
7035 numentries <<= (PAGE_SHIFT - scale);
7037 /* Make sure we've got at least a 0-order allocation.. */
7038 if (unlikely(flags & HASH_SMALL)) {
7039 /* Makes no sense without HASH_EARLY */
7040 WARN_ON(!(flags & HASH_EARLY));
7041 if (!(numentries >> *_hash_shift)) {
7042 numentries = 1UL << *_hash_shift;
7043 BUG_ON(!numentries);
7045 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7046 numentries = PAGE_SIZE / bucketsize;
7048 numentries = roundup_pow_of_two(numentries);
7050 /* limit allocation size to 1/16 total memory by default */
7052 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7053 do_div(max, bucketsize);
7055 max = min(max, 0x80000000ULL);
7057 if (numentries < low_limit)
7058 numentries = low_limit;
7059 if (numentries > max)
7062 log2qty = ilog2(numentries);
7065 size = bucketsize << log2qty;
7066 if (flags & HASH_EARLY)
7067 table = memblock_virt_alloc_nopanic(size, 0);
7069 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7072 * If bucketsize is not a power-of-two, we may free
7073 * some pages at the end of hash table which
7074 * alloc_pages_exact() automatically does
7076 if (get_order(size) < MAX_ORDER) {
7077 table = alloc_pages_exact(size, GFP_ATOMIC);
7078 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7081 } while (!table && size > PAGE_SIZE && --log2qty);
7084 panic("Failed to allocate %s hash table\n", tablename);
7086 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7087 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7090 *_hash_shift = log2qty;
7092 *_hash_mask = (1 << log2qty) - 1;
7098 * This function checks whether pageblock includes unmovable pages or not.
7099 * If @count is not zero, it is okay to include less @count unmovable pages
7101 * PageLRU check without isolation or lru_lock could race so that
7102 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7103 * expect this function should be exact.
7105 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7106 bool skip_hwpoisoned_pages)
7108 unsigned long pfn, iter, found;
7112 * For avoiding noise data, lru_add_drain_all() should be called
7113 * If ZONE_MOVABLE, the zone never contains unmovable pages
7115 if (zone_idx(zone) == ZONE_MOVABLE)
7117 mt = get_pageblock_migratetype(page);
7118 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7121 pfn = page_to_pfn(page);
7122 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7123 unsigned long check = pfn + iter;
7125 if (!pfn_valid_within(check))
7128 page = pfn_to_page(check);
7131 * Hugepages are not in LRU lists, but they're movable.
7132 * We need not scan over tail pages bacause we don't
7133 * handle each tail page individually in migration.
7135 if (PageHuge(page)) {
7136 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7141 * We can't use page_count without pin a page
7142 * because another CPU can free compound page.
7143 * This check already skips compound tails of THP
7144 * because their page->_refcount is zero at all time.
7146 if (!page_ref_count(page)) {
7147 if (PageBuddy(page))
7148 iter += (1 << page_order(page)) - 1;
7153 * The HWPoisoned page may be not in buddy system, and
7154 * page_count() is not 0.
7156 if (skip_hwpoisoned_pages && PageHWPoison(page))
7162 * If there are RECLAIMABLE pages, we need to check
7163 * it. But now, memory offline itself doesn't call
7164 * shrink_node_slabs() and it still to be fixed.
7167 * If the page is not RAM, page_count()should be 0.
7168 * we don't need more check. This is an _used_ not-movable page.
7170 * The problematic thing here is PG_reserved pages. PG_reserved
7171 * is set to both of a memory hole page and a _used_ kernel
7180 bool is_pageblock_removable_nolock(struct page *page)
7186 * We have to be careful here because we are iterating over memory
7187 * sections which are not zone aware so we might end up outside of
7188 * the zone but still within the section.
7189 * We have to take care about the node as well. If the node is offline
7190 * its NODE_DATA will be NULL - see page_zone.
7192 if (!node_online(page_to_nid(page)))
7195 zone = page_zone(page);
7196 pfn = page_to_pfn(page);
7197 if (!zone_spans_pfn(zone, pfn))
7200 return !has_unmovable_pages(zone, page, 0, true);
7203 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7205 static unsigned long pfn_max_align_down(unsigned long pfn)
7207 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7208 pageblock_nr_pages) - 1);
7211 static unsigned long pfn_max_align_up(unsigned long pfn)
7213 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7214 pageblock_nr_pages));
7217 /* [start, end) must belong to a single zone. */
7218 static int __alloc_contig_migrate_range(struct compact_control *cc,
7219 unsigned long start, unsigned long end)
7221 /* This function is based on compact_zone() from compaction.c. */
7222 unsigned long nr_reclaimed;
7223 unsigned long pfn = start;
7224 unsigned int tries = 0;
7229 while (pfn < end || !list_empty(&cc->migratepages)) {
7230 if (fatal_signal_pending(current)) {
7235 if (list_empty(&cc->migratepages)) {
7236 cc->nr_migratepages = 0;
7237 pfn = isolate_migratepages_range(cc, pfn, end);
7243 } else if (++tries == 5) {
7244 ret = ret < 0 ? ret : -EBUSY;
7248 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7250 cc->nr_migratepages -= nr_reclaimed;
7252 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7253 NULL, 0, cc->mode, MR_CMA);
7256 putback_movable_pages(&cc->migratepages);
7263 * alloc_contig_range() -- tries to allocate given range of pages
7264 * @start: start PFN to allocate
7265 * @end: one-past-the-last PFN to allocate
7266 * @migratetype: migratetype of the underlaying pageblocks (either
7267 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7268 * in range must have the same migratetype and it must
7269 * be either of the two.
7271 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7272 * aligned, however it's the caller's responsibility to guarantee that
7273 * we are the only thread that changes migrate type of pageblocks the
7276 * The PFN range must belong to a single zone.
7278 * Returns zero on success or negative error code. On success all
7279 * pages which PFN is in [start, end) are allocated for the caller and
7280 * need to be freed with free_contig_range().
7282 int alloc_contig_range(unsigned long start, unsigned long end,
7283 unsigned migratetype)
7285 unsigned long outer_start, outer_end;
7289 struct compact_control cc = {
7290 .nr_migratepages = 0,
7292 .zone = page_zone(pfn_to_page(start)),
7293 .mode = MIGRATE_SYNC,
7294 .ignore_skip_hint = true,
7296 INIT_LIST_HEAD(&cc.migratepages);
7299 * What we do here is we mark all pageblocks in range as
7300 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7301 * have different sizes, and due to the way page allocator
7302 * work, we align the range to biggest of the two pages so
7303 * that page allocator won't try to merge buddies from
7304 * different pageblocks and change MIGRATE_ISOLATE to some
7305 * other migration type.
7307 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7308 * migrate the pages from an unaligned range (ie. pages that
7309 * we are interested in). This will put all the pages in
7310 * range back to page allocator as MIGRATE_ISOLATE.
7312 * When this is done, we take the pages in range from page
7313 * allocator removing them from the buddy system. This way
7314 * page allocator will never consider using them.
7316 * This lets us mark the pageblocks back as
7317 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7318 * aligned range but not in the unaligned, original range are
7319 * put back to page allocator so that buddy can use them.
7322 ret = start_isolate_page_range(pfn_max_align_down(start),
7323 pfn_max_align_up(end), migratetype,
7329 * In case of -EBUSY, we'd like to know which page causes problem.
7330 * So, just fall through. We will check it in test_pages_isolated().
7332 ret = __alloc_contig_migrate_range(&cc, start, end);
7333 if (ret && ret != -EBUSY)
7337 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7338 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7339 * more, all pages in [start, end) are free in page allocator.
7340 * What we are going to do is to allocate all pages from
7341 * [start, end) (that is remove them from page allocator).
7343 * The only problem is that pages at the beginning and at the
7344 * end of interesting range may be not aligned with pages that
7345 * page allocator holds, ie. they can be part of higher order
7346 * pages. Because of this, we reserve the bigger range and
7347 * once this is done free the pages we are not interested in.
7349 * We don't have to hold zone->lock here because the pages are
7350 * isolated thus they won't get removed from buddy.
7353 lru_add_drain_all();
7354 drain_all_pages(cc.zone);
7357 outer_start = start;
7358 while (!PageBuddy(pfn_to_page(outer_start))) {
7359 if (++order >= MAX_ORDER) {
7360 outer_start = start;
7363 outer_start &= ~0UL << order;
7366 if (outer_start != start) {
7367 order = page_order(pfn_to_page(outer_start));
7370 * outer_start page could be small order buddy page and
7371 * it doesn't include start page. Adjust outer_start
7372 * in this case to report failed page properly
7373 * on tracepoint in test_pages_isolated()
7375 if (outer_start + (1UL << order) <= start)
7376 outer_start = start;
7379 /* Make sure the range is really isolated. */
7380 if (test_pages_isolated(outer_start, end, false)) {
7381 pr_info("%s: [%lx, %lx) PFNs busy\n",
7382 __func__, outer_start, end);
7387 /* Grab isolated pages from freelists. */
7388 outer_end = isolate_freepages_range(&cc, outer_start, end);
7394 /* Free head and tail (if any) */
7395 if (start != outer_start)
7396 free_contig_range(outer_start, start - outer_start);
7397 if (end != outer_end)
7398 free_contig_range(end, outer_end - end);
7401 undo_isolate_page_range(pfn_max_align_down(start),
7402 pfn_max_align_up(end), migratetype);
7406 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7408 unsigned int count = 0;
7410 for (; nr_pages--; pfn++) {
7411 struct page *page = pfn_to_page(pfn);
7413 count += page_count(page) != 1;
7416 WARN(count != 0, "%d pages are still in use!\n", count);
7420 #ifdef CONFIG_MEMORY_HOTPLUG
7422 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7423 * page high values need to be recalulated.
7425 void __meminit zone_pcp_update(struct zone *zone)
7428 mutex_lock(&pcp_batch_high_lock);
7429 for_each_possible_cpu(cpu)
7430 pageset_set_high_and_batch(zone,
7431 per_cpu_ptr(zone->pageset, cpu));
7432 mutex_unlock(&pcp_batch_high_lock);
7436 void zone_pcp_reset(struct zone *zone)
7438 unsigned long flags;
7440 struct per_cpu_pageset *pset;
7442 /* avoid races with drain_pages() */
7443 local_irq_save(flags);
7444 if (zone->pageset != &boot_pageset) {
7445 for_each_online_cpu(cpu) {
7446 pset = per_cpu_ptr(zone->pageset, cpu);
7447 drain_zonestat(zone, pset);
7449 free_percpu(zone->pageset);
7450 zone->pageset = &boot_pageset;
7452 local_irq_restore(flags);
7455 #ifdef CONFIG_MEMORY_HOTREMOVE
7457 * All pages in the range must be in a single zone and isolated
7458 * before calling this.
7461 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7465 unsigned int order, i;
7467 unsigned long flags;
7468 /* find the first valid pfn */
7469 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7474 zone = page_zone(pfn_to_page(pfn));
7475 spin_lock_irqsave(&zone->lock, flags);
7477 while (pfn < end_pfn) {
7478 if (!pfn_valid(pfn)) {
7482 page = pfn_to_page(pfn);
7484 * The HWPoisoned page may be not in buddy system, and
7485 * page_count() is not 0.
7487 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7489 SetPageReserved(page);
7493 BUG_ON(page_count(page));
7494 BUG_ON(!PageBuddy(page));
7495 order = page_order(page);
7496 #ifdef CONFIG_DEBUG_VM
7497 pr_info("remove from free list %lx %d %lx\n",
7498 pfn, 1 << order, end_pfn);
7500 list_del(&page->lru);
7501 rmv_page_order(page);
7502 zone->free_area[order].nr_free--;
7503 for (i = 0; i < (1 << order); i++)
7504 SetPageReserved((page+i));
7505 pfn += (1 << order);
7507 spin_unlock_irqrestore(&zone->lock, flags);
7511 bool is_free_buddy_page(struct page *page)
7513 struct zone *zone = page_zone(page);
7514 unsigned long pfn = page_to_pfn(page);
7515 unsigned long flags;
7518 spin_lock_irqsave(&zone->lock, flags);
7519 for (order = 0; order < MAX_ORDER; order++) {
7520 struct page *page_head = page - (pfn & ((1 << order) - 1));
7522 if (PageBuddy(page_head) && page_order(page_head) >= order)
7525 spin_unlock_irqrestore(&zone->lock, flags);
7527 return order < MAX_ORDER;