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/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
87 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
88 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
89 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
90 * defined in <linux/topology.h>.
92 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
93 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
94 int _node_numa_mem_[MAX_NUMNODES];
97 /* work_structs for global per-cpu drains */
98 DEFINE_MUTEX(pcpu_drain_mutex);
99 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
101 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
102 volatile unsigned long latent_entropy __latent_entropy;
103 EXPORT_SYMBOL(latent_entropy);
107 * Array of node states.
109 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
110 [N_POSSIBLE] = NODE_MASK_ALL,
111 [N_ONLINE] = { { [0] = 1UL } },
113 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
114 #ifdef CONFIG_HIGHMEM
115 [N_HIGH_MEMORY] = { { [0] = 1UL } },
117 [N_MEMORY] = { { [0] = 1UL } },
118 [N_CPU] = { { [0] = 1UL } },
121 EXPORT_SYMBOL(node_states);
123 /* Protect totalram_pages and zone->managed_pages */
124 static DEFINE_SPINLOCK(managed_page_count_lock);
126 unsigned long totalram_pages __read_mostly;
127 unsigned long totalreserve_pages __read_mostly;
128 unsigned long totalcma_pages __read_mostly;
130 int percpu_pagelist_fraction;
131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page *page)
146 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
148 page->index = migratetype;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with pm_mutex held (gfp_allowed_mask also should
157 * only be modified with pm_mutex held, unless the suspend/hibernate code is
158 * guaranteed not to run in parallel with that modification).
161 static gfp_t saved_gfp_mask;
163 void pm_restore_gfp_mask(void)
165 WARN_ON(!mutex_is_locked(&pm_mutex));
166 if (saved_gfp_mask) {
167 gfp_allowed_mask = saved_gfp_mask;
172 void pm_restrict_gfp_mask(void)
174 WARN_ON(!mutex_is_locked(&pm_mutex));
175 WARN_ON(saved_gfp_mask);
176 saved_gfp_mask = gfp_allowed_mask;
177 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
180 bool pm_suspended_storage(void)
182 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
186 #endif /* CONFIG_PM_SLEEP */
188 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
189 unsigned int pageblock_order __read_mostly;
192 static void __free_pages_ok(struct page *page, unsigned int order);
195 * results with 256, 32 in the lowmem_reserve sysctl:
196 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
197 * 1G machine -> (16M dma, 784M normal, 224M high)
198 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
199 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
200 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
202 * TBD: should special case ZONE_DMA32 machines here - in those we normally
203 * don't need any ZONE_NORMAL reservation
205 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
206 #ifdef CONFIG_ZONE_DMA
209 #ifdef CONFIG_ZONE_DMA32
212 #ifdef CONFIG_HIGHMEM
218 EXPORT_SYMBOL(totalram_pages);
220 static char * const zone_names[MAX_NR_ZONES] = {
221 #ifdef CONFIG_ZONE_DMA
224 #ifdef CONFIG_ZONE_DMA32
228 #ifdef CONFIG_HIGHMEM
232 #ifdef CONFIG_ZONE_DEVICE
237 char * const migratetype_names[MIGRATE_TYPES] = {
245 #ifdef CONFIG_MEMORY_ISOLATION
250 compound_page_dtor * const compound_page_dtors[] = {
253 #ifdef CONFIG_HUGETLB_PAGE
256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
261 int min_free_kbytes = 1024;
262 int user_min_free_kbytes = -1;
263 int watermark_scale_factor = 10;
265 static unsigned long __meminitdata nr_kernel_pages;
266 static unsigned long __meminitdata nr_all_pages;
267 static unsigned long __meminitdata dma_reserve;
269 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
270 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
271 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
272 static unsigned long __initdata required_kernelcore;
273 static unsigned long __initdata required_movablecore;
274 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
275 static bool mirrored_kernelcore;
277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
279 EXPORT_SYMBOL(movable_zone);
280 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
283 int nr_node_ids __read_mostly = MAX_NUMNODES;
284 int nr_online_nodes __read_mostly = 1;
285 EXPORT_SYMBOL(nr_node_ids);
286 EXPORT_SYMBOL(nr_online_nodes);
289 int page_group_by_mobility_disabled __read_mostly;
291 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
294 * Determine how many pages need to be initialized durig early boot
295 * (non-deferred initialization).
296 * The value of first_deferred_pfn will be set later, once non-deferred pages
297 * are initialized, but for now set it ULONG_MAX.
299 static inline void reset_deferred_meminit(pg_data_t *pgdat)
301 phys_addr_t start_addr, end_addr;
302 unsigned long max_pgcnt;
303 unsigned long reserved;
306 * Initialise at least 2G of a node but also take into account that
307 * two large system hashes that can take up 1GB for 0.25TB/node.
309 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
310 (pgdat->node_spanned_pages >> 8));
313 * Compensate the all the memblock reservations (e.g. crash kernel)
314 * from the initial estimation to make sure we will initialize enough
317 start_addr = PFN_PHYS(pgdat->node_start_pfn);
318 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
319 reserved = memblock_reserved_memory_within(start_addr, end_addr);
320 max_pgcnt += PHYS_PFN(reserved);
322 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
323 pgdat->first_deferred_pfn = ULONG_MAX;
326 /* Returns true if the struct page for the pfn is uninitialised */
327 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
329 int nid = early_pfn_to_nid(pfn);
331 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
338 * Returns false when the remaining initialisation should be deferred until
339 * later in the boot cycle when it can be parallelised.
341 static inline bool update_defer_init(pg_data_t *pgdat,
342 unsigned long pfn, unsigned long zone_end,
343 unsigned long *nr_initialised)
345 /* Always populate low zones for address-contrained allocations */
346 if (zone_end < pgdat_end_pfn(pgdat))
349 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
350 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
351 pgdat->first_deferred_pfn = pfn;
358 static inline void reset_deferred_meminit(pg_data_t *pgdat)
362 static inline bool early_page_uninitialised(unsigned long pfn)
367 static inline bool update_defer_init(pg_data_t *pgdat,
368 unsigned long pfn, unsigned long zone_end,
369 unsigned long *nr_initialised)
375 /* Return a pointer to the bitmap storing bits affecting a block of pages */
376 static inline unsigned long *get_pageblock_bitmap(struct page *page,
379 #ifdef CONFIG_SPARSEMEM
380 return __pfn_to_section(pfn)->pageblock_flags;
382 return page_zone(page)->pageblock_flags;
383 #endif /* CONFIG_SPARSEMEM */
386 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
388 #ifdef CONFIG_SPARSEMEM
389 pfn &= (PAGES_PER_SECTION-1);
390 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
392 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
393 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
394 #endif /* CONFIG_SPARSEMEM */
398 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
399 * @page: The page within the block of interest
400 * @pfn: The target page frame number
401 * @end_bitidx: The last bit of interest to retrieve
402 * @mask: mask of bits that the caller is interested in
404 * Return: pageblock_bits flags
406 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
408 unsigned long end_bitidx,
411 unsigned long *bitmap;
412 unsigned long bitidx, word_bitidx;
415 bitmap = get_pageblock_bitmap(page, pfn);
416 bitidx = pfn_to_bitidx(page, pfn);
417 word_bitidx = bitidx / BITS_PER_LONG;
418 bitidx &= (BITS_PER_LONG-1);
420 word = bitmap[word_bitidx];
421 bitidx += end_bitidx;
422 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
425 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
426 unsigned long end_bitidx,
429 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
432 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
434 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
438 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
439 * @page: The page within the block of interest
440 * @flags: The flags to set
441 * @pfn: The target page frame number
442 * @end_bitidx: The last bit of interest
443 * @mask: mask of bits that the caller is interested in
445 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
447 unsigned long end_bitidx,
450 unsigned long *bitmap;
451 unsigned long bitidx, word_bitidx;
452 unsigned long old_word, word;
454 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
456 bitmap = get_pageblock_bitmap(page, pfn);
457 bitidx = pfn_to_bitidx(page, pfn);
458 word_bitidx = bitidx / BITS_PER_LONG;
459 bitidx &= (BITS_PER_LONG-1);
461 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
463 bitidx += end_bitidx;
464 mask <<= (BITS_PER_LONG - bitidx - 1);
465 flags <<= (BITS_PER_LONG - bitidx - 1);
467 word = READ_ONCE(bitmap[word_bitidx]);
469 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
470 if (word == old_word)
476 void set_pageblock_migratetype(struct page *page, int migratetype)
478 if (unlikely(page_group_by_mobility_disabled &&
479 migratetype < MIGRATE_PCPTYPES))
480 migratetype = MIGRATE_UNMOVABLE;
482 set_pageblock_flags_group(page, (unsigned long)migratetype,
483 PB_migrate, PB_migrate_end);
486 #ifdef CONFIG_DEBUG_VM
487 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
491 unsigned long pfn = page_to_pfn(page);
492 unsigned long sp, start_pfn;
495 seq = zone_span_seqbegin(zone);
496 start_pfn = zone->zone_start_pfn;
497 sp = zone->spanned_pages;
498 if (!zone_spans_pfn(zone, pfn))
500 } while (zone_span_seqretry(zone, seq));
503 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
504 pfn, zone_to_nid(zone), zone->name,
505 start_pfn, start_pfn + sp);
510 static int page_is_consistent(struct zone *zone, struct page *page)
512 if (!pfn_valid_within(page_to_pfn(page)))
514 if (zone != page_zone(page))
520 * Temporary debugging check for pages not lying within a given zone.
522 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
524 if (page_outside_zone_boundaries(zone, page))
526 if (!page_is_consistent(zone, page))
532 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
538 static void bad_page(struct page *page, const char *reason,
539 unsigned long bad_flags)
541 static unsigned long resume;
542 static unsigned long nr_shown;
543 static unsigned long nr_unshown;
546 * Allow a burst of 60 reports, then keep quiet for that minute;
547 * or allow a steady drip of one report per second.
549 if (nr_shown == 60) {
550 if (time_before(jiffies, resume)) {
556 "BUG: Bad page state: %lu messages suppressed\n",
563 resume = jiffies + 60 * HZ;
565 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
566 current->comm, page_to_pfn(page));
567 __dump_page(page, reason);
568 bad_flags &= page->flags;
570 pr_alert("bad because of flags: %#lx(%pGp)\n",
571 bad_flags, &bad_flags);
572 dump_page_owner(page);
577 /* Leave bad fields for debug, except PageBuddy could make trouble */
578 page_mapcount_reset(page); /* remove PageBuddy */
579 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
583 * Higher-order pages are called "compound pages". They are structured thusly:
585 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
587 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
588 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
590 * The first tail page's ->compound_dtor holds the offset in array of compound
591 * page destructors. See compound_page_dtors.
593 * The first tail page's ->compound_order holds the order of allocation.
594 * This usage means that zero-order pages may not be compound.
597 void free_compound_page(struct page *page)
599 __free_pages_ok(page, compound_order(page));
602 void prep_compound_page(struct page *page, unsigned int order)
605 int nr_pages = 1 << order;
607 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
608 set_compound_order(page, order);
610 for (i = 1; i < nr_pages; i++) {
611 struct page *p = page + i;
612 set_page_count(p, 0);
613 p->mapping = TAIL_MAPPING;
614 set_compound_head(p, page);
616 atomic_set(compound_mapcount_ptr(page), -1);
619 #ifdef CONFIG_DEBUG_PAGEALLOC
620 unsigned int _debug_guardpage_minorder;
621 bool _debug_pagealloc_enabled __read_mostly
622 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
623 EXPORT_SYMBOL(_debug_pagealloc_enabled);
624 bool _debug_guardpage_enabled __read_mostly;
626 static int __init early_debug_pagealloc(char *buf)
630 return kstrtobool(buf, &_debug_pagealloc_enabled);
632 early_param("debug_pagealloc", early_debug_pagealloc);
634 static bool need_debug_guardpage(void)
636 /* If we don't use debug_pagealloc, we don't need guard page */
637 if (!debug_pagealloc_enabled())
640 if (!debug_guardpage_minorder())
646 static void init_debug_guardpage(void)
648 if (!debug_pagealloc_enabled())
651 if (!debug_guardpage_minorder())
654 _debug_guardpage_enabled = true;
657 struct page_ext_operations debug_guardpage_ops = {
658 .need = need_debug_guardpage,
659 .init = init_debug_guardpage,
662 static int __init debug_guardpage_minorder_setup(char *buf)
666 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
667 pr_err("Bad debug_guardpage_minorder value\n");
670 _debug_guardpage_minorder = res;
671 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
674 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
676 static inline bool set_page_guard(struct zone *zone, struct page *page,
677 unsigned int order, int migratetype)
679 struct page_ext *page_ext;
681 if (!debug_guardpage_enabled())
684 if (order >= debug_guardpage_minorder())
687 page_ext = lookup_page_ext(page);
688 if (unlikely(!page_ext))
691 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
693 INIT_LIST_HEAD(&page->lru);
694 set_page_private(page, order);
695 /* Guard pages are not available for any usage */
696 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
701 static inline void clear_page_guard(struct zone *zone, struct page *page,
702 unsigned int order, int migratetype)
704 struct page_ext *page_ext;
706 if (!debug_guardpage_enabled())
709 page_ext = lookup_page_ext(page);
710 if (unlikely(!page_ext))
713 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
715 set_page_private(page, 0);
716 if (!is_migrate_isolate(migratetype))
717 __mod_zone_freepage_state(zone, (1 << order), migratetype);
720 struct page_ext_operations debug_guardpage_ops;
721 static inline bool set_page_guard(struct zone *zone, struct page *page,
722 unsigned int order, int migratetype) { return false; }
723 static inline void clear_page_guard(struct zone *zone, struct page *page,
724 unsigned int order, int migratetype) {}
727 static inline void set_page_order(struct page *page, unsigned int order)
729 set_page_private(page, order);
730 __SetPageBuddy(page);
733 static inline void rmv_page_order(struct page *page)
735 __ClearPageBuddy(page);
736 set_page_private(page, 0);
740 * This function checks whether a page is free && is the buddy
741 * we can do coalesce a page and its buddy if
742 * (a) the buddy is not in a hole (check before calling!) &&
743 * (b) the buddy is in the buddy system &&
744 * (c) a page and its buddy have the same order &&
745 * (d) a page and its buddy are in the same zone.
747 * For recording whether a page is in the buddy system, we set ->_mapcount
748 * PAGE_BUDDY_MAPCOUNT_VALUE.
749 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
750 * serialized by zone->lock.
752 * For recording page's order, we use page_private(page).
754 static inline int page_is_buddy(struct page *page, struct page *buddy,
757 if (page_is_guard(buddy) && page_order(buddy) == order) {
758 if (page_zone_id(page) != page_zone_id(buddy))
761 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
766 if (PageBuddy(buddy) && page_order(buddy) == order) {
768 * zone check is done late to avoid uselessly
769 * calculating zone/node ids for pages that could
772 if (page_zone_id(page) != page_zone_id(buddy))
775 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
783 * Freeing function for a buddy system allocator.
785 * The concept of a buddy system is to maintain direct-mapped table
786 * (containing bit values) for memory blocks of various "orders".
787 * The bottom level table contains the map for the smallest allocatable
788 * units of memory (here, pages), and each level above it describes
789 * pairs of units from the levels below, hence, "buddies".
790 * At a high level, all that happens here is marking the table entry
791 * at the bottom level available, and propagating the changes upward
792 * as necessary, plus some accounting needed to play nicely with other
793 * parts of the VM system.
794 * At each level, we keep a list of pages, which are heads of continuous
795 * free pages of length of (1 << order) and marked with _mapcount
796 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
798 * So when we are allocating or freeing one, we can derive the state of the
799 * other. That is, if we allocate a small block, and both were
800 * free, the remainder of the region must be split into blocks.
801 * If a block is freed, and its buddy is also free, then this
802 * triggers coalescing into a block of larger size.
807 static inline void __free_one_page(struct page *page,
809 struct zone *zone, unsigned int order,
812 unsigned long combined_pfn;
813 unsigned long uninitialized_var(buddy_pfn);
815 unsigned int max_order;
817 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
819 VM_BUG_ON(!zone_is_initialized(zone));
820 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
822 VM_BUG_ON(migratetype == -1);
823 if (likely(!is_migrate_isolate(migratetype)))
824 __mod_zone_freepage_state(zone, 1 << order, migratetype);
826 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
827 VM_BUG_ON_PAGE(bad_range(zone, page), page);
830 while (order < max_order - 1) {
831 buddy_pfn = __find_buddy_pfn(pfn, order);
832 buddy = page + (buddy_pfn - pfn);
834 if (!pfn_valid_within(buddy_pfn))
836 if (!page_is_buddy(page, buddy, order))
839 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
840 * merge with it and move up one order.
842 if (page_is_guard(buddy)) {
843 clear_page_guard(zone, buddy, order, migratetype);
845 list_del(&buddy->lru);
846 zone->free_area[order].nr_free--;
847 rmv_page_order(buddy);
849 combined_pfn = buddy_pfn & pfn;
850 page = page + (combined_pfn - pfn);
854 if (max_order < MAX_ORDER) {
855 /* If we are here, it means order is >= pageblock_order.
856 * We want to prevent merge between freepages on isolate
857 * pageblock and normal pageblock. Without this, pageblock
858 * isolation could cause incorrect freepage or CMA accounting.
860 * We don't want to hit this code for the more frequent
863 if (unlikely(has_isolate_pageblock(zone))) {
866 buddy_pfn = __find_buddy_pfn(pfn, order);
867 buddy = page + (buddy_pfn - pfn);
868 buddy_mt = get_pageblock_migratetype(buddy);
870 if (migratetype != buddy_mt
871 && (is_migrate_isolate(migratetype) ||
872 is_migrate_isolate(buddy_mt)))
876 goto continue_merging;
880 set_page_order(page, order);
883 * If this is not the largest possible page, check if the buddy
884 * of the next-highest order is free. If it is, it's possible
885 * that pages are being freed that will coalesce soon. In case,
886 * that is happening, add the free page to the tail of the list
887 * so it's less likely to be used soon and more likely to be merged
888 * as a higher order page
890 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
891 struct page *higher_page, *higher_buddy;
892 combined_pfn = buddy_pfn & pfn;
893 higher_page = page + (combined_pfn - pfn);
894 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
895 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
896 if (pfn_valid_within(buddy_pfn) &&
897 page_is_buddy(higher_page, higher_buddy, order + 1)) {
898 list_add_tail(&page->lru,
899 &zone->free_area[order].free_list[migratetype]);
904 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
906 zone->free_area[order].nr_free++;
910 * A bad page could be due to a number of fields. Instead of multiple branches,
911 * try and check multiple fields with one check. The caller must do a detailed
912 * check if necessary.
914 static inline bool page_expected_state(struct page *page,
915 unsigned long check_flags)
917 if (unlikely(atomic_read(&page->_mapcount) != -1))
920 if (unlikely((unsigned long)page->mapping |
921 page_ref_count(page) |
923 (unsigned long)page->mem_cgroup |
925 (page->flags & check_flags)))
931 static void free_pages_check_bad(struct page *page)
933 const char *bad_reason;
934 unsigned long bad_flags;
939 if (unlikely(atomic_read(&page->_mapcount) != -1))
940 bad_reason = "nonzero mapcount";
941 if (unlikely(page->mapping != NULL))
942 bad_reason = "non-NULL mapping";
943 if (unlikely(page_ref_count(page) != 0))
944 bad_reason = "nonzero _refcount";
945 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
946 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
947 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
950 if (unlikely(page->mem_cgroup))
951 bad_reason = "page still charged to cgroup";
953 bad_page(page, bad_reason, bad_flags);
956 static inline int free_pages_check(struct page *page)
958 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
961 /* Something has gone sideways, find it */
962 free_pages_check_bad(page);
966 static int free_tail_pages_check(struct page *head_page, struct page *page)
971 * We rely page->lru.next never has bit 0 set, unless the page
972 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
974 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
976 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
980 switch (page - head_page) {
982 /* the first tail page: ->mapping is compound_mapcount() */
983 if (unlikely(compound_mapcount(page))) {
984 bad_page(page, "nonzero compound_mapcount", 0);
990 * the second tail page: ->mapping is
991 * page_deferred_list().next -- ignore value.
995 if (page->mapping != TAIL_MAPPING) {
996 bad_page(page, "corrupted mapping in tail page", 0);
1001 if (unlikely(!PageTail(page))) {
1002 bad_page(page, "PageTail not set", 0);
1005 if (unlikely(compound_head(page) != head_page)) {
1006 bad_page(page, "compound_head not consistent", 0);
1011 page->mapping = NULL;
1012 clear_compound_head(page);
1016 static __always_inline bool free_pages_prepare(struct page *page,
1017 unsigned int order, bool check_free)
1021 VM_BUG_ON_PAGE(PageTail(page), page);
1023 trace_mm_page_free(page, order);
1026 * Check tail pages before head page information is cleared to
1027 * avoid checking PageCompound for order-0 pages.
1029 if (unlikely(order)) {
1030 bool compound = PageCompound(page);
1033 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1036 ClearPageDoubleMap(page);
1037 for (i = 1; i < (1 << order); i++) {
1039 bad += free_tail_pages_check(page, page + i);
1040 if (unlikely(free_pages_check(page + i))) {
1044 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1047 if (PageMappingFlags(page))
1048 page->mapping = NULL;
1049 if (memcg_kmem_enabled() && PageKmemcg(page))
1050 memcg_kmem_uncharge(page, order);
1052 bad += free_pages_check(page);
1056 page_cpupid_reset_last(page);
1057 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1058 reset_page_owner(page, order);
1060 if (!PageHighMem(page)) {
1061 debug_check_no_locks_freed(page_address(page),
1062 PAGE_SIZE << order);
1063 debug_check_no_obj_freed(page_address(page),
1064 PAGE_SIZE << order);
1066 arch_free_page(page, order);
1067 kernel_poison_pages(page, 1 << order, 0);
1068 kernel_map_pages(page, 1 << order, 0);
1069 kasan_free_pages(page, order);
1074 #ifdef CONFIG_DEBUG_VM
1075 static inline bool free_pcp_prepare(struct page *page)
1077 return free_pages_prepare(page, 0, true);
1080 static inline bool bulkfree_pcp_prepare(struct page *page)
1085 static bool free_pcp_prepare(struct page *page)
1087 return free_pages_prepare(page, 0, false);
1090 static bool bulkfree_pcp_prepare(struct page *page)
1092 return free_pages_check(page);
1094 #endif /* CONFIG_DEBUG_VM */
1097 * Frees a number of pages from the PCP lists
1098 * Assumes all pages on list are in same zone, and of same order.
1099 * count is the number of pages to free.
1101 * If the zone was previously in an "all pages pinned" state then look to
1102 * see if this freeing clears that state.
1104 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1105 * pinned" detection logic.
1107 static void free_pcppages_bulk(struct zone *zone, int count,
1108 struct per_cpu_pages *pcp)
1110 int migratetype = 0;
1112 bool isolated_pageblocks;
1114 spin_lock(&zone->lock);
1115 isolated_pageblocks = has_isolate_pageblock(zone);
1119 struct list_head *list;
1122 * Remove pages from lists in a round-robin fashion. A
1123 * batch_free count is maintained that is incremented when an
1124 * empty list is encountered. This is so more pages are freed
1125 * off fuller lists instead of spinning excessively around empty
1130 if (++migratetype == MIGRATE_PCPTYPES)
1132 list = &pcp->lists[migratetype];
1133 } while (list_empty(list));
1135 /* This is the only non-empty list. Free them all. */
1136 if (batch_free == MIGRATE_PCPTYPES)
1140 int mt; /* migratetype of the to-be-freed page */
1142 page = list_last_entry(list, struct page, lru);
1143 /* must delete as __free_one_page list manipulates */
1144 list_del(&page->lru);
1146 mt = get_pcppage_migratetype(page);
1147 /* MIGRATE_ISOLATE page should not go to pcplists */
1148 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1149 /* Pageblock could have been isolated meanwhile */
1150 if (unlikely(isolated_pageblocks))
1151 mt = get_pageblock_migratetype(page);
1153 if (bulkfree_pcp_prepare(page))
1156 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1157 trace_mm_page_pcpu_drain(page, 0, mt);
1158 } while (--count && --batch_free && !list_empty(list));
1160 spin_unlock(&zone->lock);
1163 static void free_one_page(struct zone *zone,
1164 struct page *page, unsigned long pfn,
1168 spin_lock(&zone->lock);
1169 if (unlikely(has_isolate_pageblock(zone) ||
1170 is_migrate_isolate(migratetype))) {
1171 migratetype = get_pfnblock_migratetype(page, pfn);
1173 __free_one_page(page, pfn, zone, order, migratetype);
1174 spin_unlock(&zone->lock);
1177 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1178 unsigned long zone, int nid)
1180 set_page_links(page, zone, nid, pfn);
1181 init_page_count(page);
1182 page_mapcount_reset(page);
1183 page_cpupid_reset_last(page);
1185 INIT_LIST_HEAD(&page->lru);
1186 #ifdef WANT_PAGE_VIRTUAL
1187 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1188 if (!is_highmem_idx(zone))
1189 set_page_address(page, __va(pfn << PAGE_SHIFT));
1193 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1196 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1199 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1200 static void __meminit init_reserved_page(unsigned long pfn)
1205 if (!early_page_uninitialised(pfn))
1208 nid = early_pfn_to_nid(pfn);
1209 pgdat = NODE_DATA(nid);
1211 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1212 struct zone *zone = &pgdat->node_zones[zid];
1214 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1217 __init_single_pfn(pfn, zid, nid);
1220 static inline void init_reserved_page(unsigned long pfn)
1223 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1226 * Initialised pages do not have PageReserved set. This function is
1227 * called for each range allocated by the bootmem allocator and
1228 * marks the pages PageReserved. The remaining valid pages are later
1229 * sent to the buddy page allocator.
1231 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1233 unsigned long start_pfn = PFN_DOWN(start);
1234 unsigned long end_pfn = PFN_UP(end);
1236 for (; start_pfn < end_pfn; start_pfn++) {
1237 if (pfn_valid(start_pfn)) {
1238 struct page *page = pfn_to_page(start_pfn);
1240 init_reserved_page(start_pfn);
1242 /* Avoid false-positive PageTail() */
1243 INIT_LIST_HEAD(&page->lru);
1245 SetPageReserved(page);
1250 static void __free_pages_ok(struct page *page, unsigned int order)
1252 unsigned long flags;
1254 unsigned long pfn = page_to_pfn(page);
1256 if (!free_pages_prepare(page, order, true))
1259 migratetype = get_pfnblock_migratetype(page, pfn);
1260 local_irq_save(flags);
1261 __count_vm_events(PGFREE, 1 << order);
1262 free_one_page(page_zone(page), page, pfn, order, migratetype);
1263 local_irq_restore(flags);
1266 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1268 unsigned int nr_pages = 1 << order;
1269 struct page *p = page;
1273 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1275 __ClearPageReserved(p);
1276 set_page_count(p, 0);
1278 __ClearPageReserved(p);
1279 set_page_count(p, 0);
1281 page_zone(page)->managed_pages += nr_pages;
1282 set_page_refcounted(page);
1283 __free_pages(page, order);
1286 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1287 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1289 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1291 int __meminit early_pfn_to_nid(unsigned long pfn)
1293 static DEFINE_SPINLOCK(early_pfn_lock);
1296 spin_lock(&early_pfn_lock);
1297 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1299 nid = first_online_node;
1300 spin_unlock(&early_pfn_lock);
1306 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1307 static inline bool __meminit __maybe_unused
1308 meminit_pfn_in_nid(unsigned long pfn, int node,
1309 struct mminit_pfnnid_cache *state)
1313 nid = __early_pfn_to_nid(pfn, state);
1314 if (nid >= 0 && nid != node)
1319 /* Only safe to use early in boot when initialisation is single-threaded */
1320 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1322 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1327 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1331 static inline bool __meminit __maybe_unused
1332 meminit_pfn_in_nid(unsigned long pfn, int node,
1333 struct mminit_pfnnid_cache *state)
1340 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1343 if (early_page_uninitialised(pfn))
1345 return __free_pages_boot_core(page, order);
1349 * Check that the whole (or subset of) a pageblock given by the interval of
1350 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1351 * with the migration of free compaction scanner. The scanners then need to
1352 * use only pfn_valid_within() check for arches that allow holes within
1355 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1357 * It's possible on some configurations to have a setup like node0 node1 node0
1358 * i.e. it's possible that all pages within a zones range of pages do not
1359 * belong to a single zone. We assume that a border between node0 and node1
1360 * can occur within a single pageblock, but not a node0 node1 node0
1361 * interleaving within a single pageblock. It is therefore sufficient to check
1362 * the first and last page of a pageblock and avoid checking each individual
1363 * page in a pageblock.
1365 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1366 unsigned long end_pfn, struct zone *zone)
1368 struct page *start_page;
1369 struct page *end_page;
1371 /* end_pfn is one past the range we are checking */
1374 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1377 start_page = pfn_to_online_page(start_pfn);
1381 if (page_zone(start_page) != zone)
1384 end_page = pfn_to_page(end_pfn);
1386 /* This gives a shorter code than deriving page_zone(end_page) */
1387 if (page_zone_id(start_page) != page_zone_id(end_page))
1393 void set_zone_contiguous(struct zone *zone)
1395 unsigned long block_start_pfn = zone->zone_start_pfn;
1396 unsigned long block_end_pfn;
1398 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1399 for (; block_start_pfn < zone_end_pfn(zone);
1400 block_start_pfn = block_end_pfn,
1401 block_end_pfn += pageblock_nr_pages) {
1403 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1405 if (!__pageblock_pfn_to_page(block_start_pfn,
1406 block_end_pfn, zone))
1410 /* We confirm that there is no hole */
1411 zone->contiguous = true;
1414 void clear_zone_contiguous(struct zone *zone)
1416 zone->contiguous = false;
1419 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1420 static void __init deferred_free_range(struct page *page,
1421 unsigned long pfn, int nr_pages)
1428 /* Free a large naturally-aligned chunk if possible */
1429 if (nr_pages == pageblock_nr_pages &&
1430 (pfn & (pageblock_nr_pages - 1)) == 0) {
1431 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1432 __free_pages_boot_core(page, pageblock_order);
1436 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1437 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1438 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1439 __free_pages_boot_core(page, 0);
1443 /* Completion tracking for deferred_init_memmap() threads */
1444 static atomic_t pgdat_init_n_undone __initdata;
1445 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1447 static inline void __init pgdat_init_report_one_done(void)
1449 if (atomic_dec_and_test(&pgdat_init_n_undone))
1450 complete(&pgdat_init_all_done_comp);
1453 /* Initialise remaining memory on a node */
1454 static int __init deferred_init_memmap(void *data)
1456 pg_data_t *pgdat = data;
1457 int nid = pgdat->node_id;
1458 struct mminit_pfnnid_cache nid_init_state = { };
1459 unsigned long start = jiffies;
1460 unsigned long nr_pages = 0;
1461 unsigned long walk_start, walk_end;
1464 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1465 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1467 if (first_init_pfn == ULONG_MAX) {
1468 pgdat_init_report_one_done();
1472 /* Bind memory initialisation thread to a local node if possible */
1473 if (!cpumask_empty(cpumask))
1474 set_cpus_allowed_ptr(current, cpumask);
1476 /* Sanity check boundaries */
1477 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1478 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1479 pgdat->first_deferred_pfn = ULONG_MAX;
1481 /* Only the highest zone is deferred so find it */
1482 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1483 zone = pgdat->node_zones + zid;
1484 if (first_init_pfn < zone_end_pfn(zone))
1488 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1489 unsigned long pfn, end_pfn;
1490 struct page *page = NULL;
1491 struct page *free_base_page = NULL;
1492 unsigned long free_base_pfn = 0;
1495 end_pfn = min(walk_end, zone_end_pfn(zone));
1496 pfn = first_init_pfn;
1497 if (pfn < walk_start)
1499 if (pfn < zone->zone_start_pfn)
1500 pfn = zone->zone_start_pfn;
1502 for (; pfn < end_pfn; pfn++) {
1503 if (!pfn_valid_within(pfn))
1507 * Ensure pfn_valid is checked every
1508 * pageblock_nr_pages for memory holes
1510 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1511 if (!pfn_valid(pfn)) {
1517 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1522 /* Minimise pfn page lookups and scheduler checks */
1523 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1526 nr_pages += nr_to_free;
1527 deferred_free_range(free_base_page,
1528 free_base_pfn, nr_to_free);
1529 free_base_page = NULL;
1530 free_base_pfn = nr_to_free = 0;
1532 page = pfn_to_page(pfn);
1537 VM_BUG_ON(page_zone(page) != zone);
1541 __init_single_page(page, pfn, zid, nid);
1542 if (!free_base_page) {
1543 free_base_page = page;
1544 free_base_pfn = pfn;
1549 /* Where possible, batch up pages for a single free */
1552 /* Free the current block of pages to allocator */
1553 nr_pages += nr_to_free;
1554 deferred_free_range(free_base_page, free_base_pfn,
1556 free_base_page = NULL;
1557 free_base_pfn = nr_to_free = 0;
1559 /* Free the last block of pages to allocator */
1560 nr_pages += nr_to_free;
1561 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1563 first_init_pfn = max(end_pfn, first_init_pfn);
1566 /* Sanity check that the next zone really is unpopulated */
1567 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1569 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1570 jiffies_to_msecs(jiffies - start));
1572 pgdat_init_report_one_done();
1575 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1577 void __init page_alloc_init_late(void)
1581 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1584 /* There will be num_node_state(N_MEMORY) threads */
1585 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1586 for_each_node_state(nid, N_MEMORY) {
1587 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1590 /* Block until all are initialised */
1591 wait_for_completion(&pgdat_init_all_done_comp);
1593 /* Reinit limits that are based on free pages after the kernel is up */
1594 files_maxfiles_init();
1596 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1597 /* Discard memblock private memory */
1601 for_each_populated_zone(zone)
1602 set_zone_contiguous(zone);
1606 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1607 void __init init_cma_reserved_pageblock(struct page *page)
1609 unsigned i = pageblock_nr_pages;
1610 struct page *p = page;
1613 __ClearPageReserved(p);
1614 set_page_count(p, 0);
1617 set_pageblock_migratetype(page, MIGRATE_CMA);
1619 if (pageblock_order >= MAX_ORDER) {
1620 i = pageblock_nr_pages;
1623 set_page_refcounted(p);
1624 __free_pages(p, MAX_ORDER - 1);
1625 p += MAX_ORDER_NR_PAGES;
1626 } while (i -= MAX_ORDER_NR_PAGES);
1628 set_page_refcounted(page);
1629 __free_pages(page, pageblock_order);
1632 adjust_managed_page_count(page, pageblock_nr_pages);
1637 * The order of subdivision here is critical for the IO subsystem.
1638 * Please do not alter this order without good reasons and regression
1639 * testing. Specifically, as large blocks of memory are subdivided,
1640 * the order in which smaller blocks are delivered depends on the order
1641 * they're subdivided in this function. This is the primary factor
1642 * influencing the order in which pages are delivered to the IO
1643 * subsystem according to empirical testing, and this is also justified
1644 * by considering the behavior of a buddy system containing a single
1645 * large block of memory acted on by a series of small allocations.
1646 * This behavior is a critical factor in sglist merging's success.
1650 static inline void expand(struct zone *zone, struct page *page,
1651 int low, int high, struct free_area *area,
1654 unsigned long size = 1 << high;
1656 while (high > low) {
1660 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1663 * Mark as guard pages (or page), that will allow to
1664 * merge back to allocator when buddy will be freed.
1665 * Corresponding page table entries will not be touched,
1666 * pages will stay not present in virtual address space
1668 if (set_page_guard(zone, &page[size], high, migratetype))
1671 list_add(&page[size].lru, &area->free_list[migratetype]);
1673 set_page_order(&page[size], high);
1677 static void check_new_page_bad(struct page *page)
1679 const char *bad_reason = NULL;
1680 unsigned long bad_flags = 0;
1682 if (unlikely(atomic_read(&page->_mapcount) != -1))
1683 bad_reason = "nonzero mapcount";
1684 if (unlikely(page->mapping != NULL))
1685 bad_reason = "non-NULL mapping";
1686 if (unlikely(page_ref_count(page) != 0))
1687 bad_reason = "nonzero _count";
1688 if (unlikely(page->flags & __PG_HWPOISON)) {
1689 bad_reason = "HWPoisoned (hardware-corrupted)";
1690 bad_flags = __PG_HWPOISON;
1691 /* Don't complain about hwpoisoned pages */
1692 page_mapcount_reset(page); /* remove PageBuddy */
1695 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1696 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1697 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1700 if (unlikely(page->mem_cgroup))
1701 bad_reason = "page still charged to cgroup";
1703 bad_page(page, bad_reason, bad_flags);
1707 * This page is about to be returned from the page allocator
1709 static inline int check_new_page(struct page *page)
1711 if (likely(page_expected_state(page,
1712 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1715 check_new_page_bad(page);
1719 static inline bool free_pages_prezeroed(void)
1721 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1722 page_poisoning_enabled();
1725 #ifdef CONFIG_DEBUG_VM
1726 static bool check_pcp_refill(struct page *page)
1731 static bool check_new_pcp(struct page *page)
1733 return check_new_page(page);
1736 static bool check_pcp_refill(struct page *page)
1738 return check_new_page(page);
1740 static bool check_new_pcp(struct page *page)
1744 #endif /* CONFIG_DEBUG_VM */
1746 static bool check_new_pages(struct page *page, unsigned int order)
1749 for (i = 0; i < (1 << order); i++) {
1750 struct page *p = page + i;
1752 if (unlikely(check_new_page(p)))
1759 inline void post_alloc_hook(struct page *page, unsigned int order,
1762 set_page_private(page, 0);
1763 set_page_refcounted(page);
1765 arch_alloc_page(page, order);
1766 kernel_map_pages(page, 1 << order, 1);
1767 kernel_poison_pages(page, 1 << order, 1);
1768 kasan_alloc_pages(page, order);
1769 set_page_owner(page, order, gfp_flags);
1772 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1773 unsigned int alloc_flags)
1777 post_alloc_hook(page, order, gfp_flags);
1779 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1780 for (i = 0; i < (1 << order); i++)
1781 clear_highpage(page + i);
1783 if (order && (gfp_flags & __GFP_COMP))
1784 prep_compound_page(page, order);
1787 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1788 * allocate the page. The expectation is that the caller is taking
1789 * steps that will free more memory. The caller should avoid the page
1790 * being used for !PFMEMALLOC purposes.
1792 if (alloc_flags & ALLOC_NO_WATERMARKS)
1793 set_page_pfmemalloc(page);
1795 clear_page_pfmemalloc(page);
1799 * Go through the free lists for the given migratetype and remove
1800 * the smallest available page from the freelists
1803 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1806 unsigned int current_order;
1807 struct free_area *area;
1810 /* Find a page of the appropriate size in the preferred list */
1811 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1812 area = &(zone->free_area[current_order]);
1813 page = list_first_entry_or_null(&area->free_list[migratetype],
1817 list_del(&page->lru);
1818 rmv_page_order(page);
1820 expand(zone, page, order, current_order, area, migratetype);
1821 set_pcppage_migratetype(page, migratetype);
1830 * This array describes the order lists are fallen back to when
1831 * the free lists for the desirable migrate type are depleted
1833 static int fallbacks[MIGRATE_TYPES][4] = {
1834 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1835 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1836 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1838 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1840 #ifdef CONFIG_MEMORY_ISOLATION
1841 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1846 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1849 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1852 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1853 unsigned int order) { return NULL; }
1857 * Move the free pages in a range to the free lists of the requested type.
1858 * Note that start_page and end_pages are not aligned on a pageblock
1859 * boundary. If alignment is required, use move_freepages_block()
1861 static int move_freepages(struct zone *zone,
1862 struct page *start_page, struct page *end_page,
1863 int migratetype, int *num_movable)
1867 int pages_moved = 0;
1869 #ifndef CONFIG_HOLES_IN_ZONE
1871 * page_zone is not safe to call in this context when
1872 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1873 * anyway as we check zone boundaries in move_freepages_block().
1874 * Remove at a later date when no bug reports exist related to
1875 * grouping pages by mobility
1877 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1883 for (page = start_page; page <= end_page;) {
1884 if (!pfn_valid_within(page_to_pfn(page))) {
1889 /* Make sure we are not inadvertently changing nodes */
1890 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1892 if (!PageBuddy(page)) {
1894 * We assume that pages that could be isolated for
1895 * migration are movable. But we don't actually try
1896 * isolating, as that would be expensive.
1899 (PageLRU(page) || __PageMovable(page)))
1906 order = page_order(page);
1907 list_move(&page->lru,
1908 &zone->free_area[order].free_list[migratetype]);
1910 pages_moved += 1 << order;
1916 int move_freepages_block(struct zone *zone, struct page *page,
1917 int migratetype, int *num_movable)
1919 unsigned long start_pfn, end_pfn;
1920 struct page *start_page, *end_page;
1922 start_pfn = page_to_pfn(page);
1923 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1924 start_page = pfn_to_page(start_pfn);
1925 end_page = start_page + pageblock_nr_pages - 1;
1926 end_pfn = start_pfn + pageblock_nr_pages - 1;
1928 /* Do not cross zone boundaries */
1929 if (!zone_spans_pfn(zone, start_pfn))
1931 if (!zone_spans_pfn(zone, end_pfn))
1934 return move_freepages(zone, start_page, end_page, migratetype,
1938 static void change_pageblock_range(struct page *pageblock_page,
1939 int start_order, int migratetype)
1941 int nr_pageblocks = 1 << (start_order - pageblock_order);
1943 while (nr_pageblocks--) {
1944 set_pageblock_migratetype(pageblock_page, migratetype);
1945 pageblock_page += pageblock_nr_pages;
1950 * When we are falling back to another migratetype during allocation, try to
1951 * steal extra free pages from the same pageblocks to satisfy further
1952 * allocations, instead of polluting multiple pageblocks.
1954 * If we are stealing a relatively large buddy page, it is likely there will
1955 * be more free pages in the pageblock, so try to steal them all. For
1956 * reclaimable and unmovable allocations, we steal regardless of page size,
1957 * as fragmentation caused by those allocations polluting movable pageblocks
1958 * is worse than movable allocations stealing from unmovable and reclaimable
1961 static bool can_steal_fallback(unsigned int order, int start_mt)
1964 * Leaving this order check is intended, although there is
1965 * relaxed order check in next check. The reason is that
1966 * we can actually steal whole pageblock if this condition met,
1967 * but, below check doesn't guarantee it and that is just heuristic
1968 * so could be changed anytime.
1970 if (order >= pageblock_order)
1973 if (order >= pageblock_order / 2 ||
1974 start_mt == MIGRATE_RECLAIMABLE ||
1975 start_mt == MIGRATE_UNMOVABLE ||
1976 page_group_by_mobility_disabled)
1983 * This function implements actual steal behaviour. If order is large enough,
1984 * we can steal whole pageblock. If not, we first move freepages in this
1985 * pageblock to our migratetype and determine how many already-allocated pages
1986 * are there in the pageblock with a compatible migratetype. If at least half
1987 * of pages are free or compatible, we can change migratetype of the pageblock
1988 * itself, so pages freed in the future will be put on the correct free list.
1990 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1991 int start_type, bool whole_block)
1993 unsigned int current_order = page_order(page);
1994 struct free_area *area;
1995 int free_pages, movable_pages, alike_pages;
1998 old_block_type = get_pageblock_migratetype(page);
2001 * This can happen due to races and we want to prevent broken
2002 * highatomic accounting.
2004 if (is_migrate_highatomic(old_block_type))
2007 /* Take ownership for orders >= pageblock_order */
2008 if (current_order >= pageblock_order) {
2009 change_pageblock_range(page, current_order, start_type);
2013 /* We are not allowed to try stealing from the whole block */
2017 free_pages = move_freepages_block(zone, page, start_type,
2020 * Determine how many pages are compatible with our allocation.
2021 * For movable allocation, it's the number of movable pages which
2022 * we just obtained. For other types it's a bit more tricky.
2024 if (start_type == MIGRATE_MOVABLE) {
2025 alike_pages = movable_pages;
2028 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2029 * to MOVABLE pageblock, consider all non-movable pages as
2030 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2031 * vice versa, be conservative since we can't distinguish the
2032 * exact migratetype of non-movable pages.
2034 if (old_block_type == MIGRATE_MOVABLE)
2035 alike_pages = pageblock_nr_pages
2036 - (free_pages + movable_pages);
2041 /* moving whole block can fail due to zone boundary conditions */
2046 * If a sufficient number of pages in the block are either free or of
2047 * comparable migratability as our allocation, claim the whole block.
2049 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2050 page_group_by_mobility_disabled)
2051 set_pageblock_migratetype(page, start_type);
2056 area = &zone->free_area[current_order];
2057 list_move(&page->lru, &area->free_list[start_type]);
2061 * Check whether there is a suitable fallback freepage with requested order.
2062 * If only_stealable is true, this function returns fallback_mt only if
2063 * we can steal other freepages all together. This would help to reduce
2064 * fragmentation due to mixed migratetype pages in one pageblock.
2066 int find_suitable_fallback(struct free_area *area, unsigned int order,
2067 int migratetype, bool only_stealable, bool *can_steal)
2072 if (area->nr_free == 0)
2077 fallback_mt = fallbacks[migratetype][i];
2078 if (fallback_mt == MIGRATE_TYPES)
2081 if (list_empty(&area->free_list[fallback_mt]))
2084 if (can_steal_fallback(order, migratetype))
2087 if (!only_stealable)
2098 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2099 * there are no empty page blocks that contain a page with a suitable order
2101 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2102 unsigned int alloc_order)
2105 unsigned long max_managed, flags;
2108 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2109 * Check is race-prone but harmless.
2111 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2112 if (zone->nr_reserved_highatomic >= max_managed)
2115 spin_lock_irqsave(&zone->lock, flags);
2117 /* Recheck the nr_reserved_highatomic limit under the lock */
2118 if (zone->nr_reserved_highatomic >= max_managed)
2122 mt = get_pageblock_migratetype(page);
2123 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2124 && !is_migrate_cma(mt)) {
2125 zone->nr_reserved_highatomic += pageblock_nr_pages;
2126 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2127 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2131 spin_unlock_irqrestore(&zone->lock, flags);
2135 * Used when an allocation is about to fail under memory pressure. This
2136 * potentially hurts the reliability of high-order allocations when under
2137 * intense memory pressure but failed atomic allocations should be easier
2138 * to recover from than an OOM.
2140 * If @force is true, try to unreserve a pageblock even though highatomic
2141 * pageblock is exhausted.
2143 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2146 struct zonelist *zonelist = ac->zonelist;
2147 unsigned long flags;
2154 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2157 * Preserve at least one pageblock unless memory pressure
2160 if (!force && zone->nr_reserved_highatomic <=
2164 spin_lock_irqsave(&zone->lock, flags);
2165 for (order = 0; order < MAX_ORDER; order++) {
2166 struct free_area *area = &(zone->free_area[order]);
2168 page = list_first_entry_or_null(
2169 &area->free_list[MIGRATE_HIGHATOMIC],
2175 * In page freeing path, migratetype change is racy so
2176 * we can counter several free pages in a pageblock
2177 * in this loop althoug we changed the pageblock type
2178 * from highatomic to ac->migratetype. So we should
2179 * adjust the count once.
2181 if (is_migrate_highatomic_page(page)) {
2183 * It should never happen but changes to
2184 * locking could inadvertently allow a per-cpu
2185 * drain to add pages to MIGRATE_HIGHATOMIC
2186 * while unreserving so be safe and watch for
2189 zone->nr_reserved_highatomic -= min(
2191 zone->nr_reserved_highatomic);
2195 * Convert to ac->migratetype and avoid the normal
2196 * pageblock stealing heuristics. Minimally, the caller
2197 * is doing the work and needs the pages. More
2198 * importantly, if the block was always converted to
2199 * MIGRATE_UNMOVABLE or another type then the number
2200 * of pageblocks that cannot be completely freed
2203 set_pageblock_migratetype(page, ac->migratetype);
2204 ret = move_freepages_block(zone, page, ac->migratetype,
2207 spin_unlock_irqrestore(&zone->lock, flags);
2211 spin_unlock_irqrestore(&zone->lock, flags);
2218 * Try finding a free buddy page on the fallback list and put it on the free
2219 * list of requested migratetype, possibly along with other pages from the same
2220 * block, depending on fragmentation avoidance heuristics. Returns true if
2221 * fallback was found so that __rmqueue_smallest() can grab it.
2223 * The use of signed ints for order and current_order is a deliberate
2224 * deviation from the rest of this file, to make the for loop
2225 * condition simpler.
2228 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2230 struct free_area *area;
2237 * Find the largest available free page in the other list. This roughly
2238 * approximates finding the pageblock with the most free pages, which
2239 * would be too costly to do exactly.
2241 for (current_order = MAX_ORDER - 1; current_order >= order;
2243 area = &(zone->free_area[current_order]);
2244 fallback_mt = find_suitable_fallback(area, current_order,
2245 start_migratetype, false, &can_steal);
2246 if (fallback_mt == -1)
2250 * We cannot steal all free pages from the pageblock and the
2251 * requested migratetype is movable. In that case it's better to
2252 * steal and split the smallest available page instead of the
2253 * largest available page, because even if the next movable
2254 * allocation falls back into a different pageblock than this
2255 * one, it won't cause permanent fragmentation.
2257 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2258 && current_order > order)
2267 for (current_order = order; current_order < MAX_ORDER;
2269 area = &(zone->free_area[current_order]);
2270 fallback_mt = find_suitable_fallback(area, current_order,
2271 start_migratetype, false, &can_steal);
2272 if (fallback_mt != -1)
2277 * This should not happen - we already found a suitable fallback
2278 * when looking for the largest page.
2280 VM_BUG_ON(current_order == MAX_ORDER);
2283 page = list_first_entry(&area->free_list[fallback_mt],
2286 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2288 trace_mm_page_alloc_extfrag(page, order, current_order,
2289 start_migratetype, fallback_mt);
2296 * Do the hard work of removing an element from the buddy allocator.
2297 * Call me with the zone->lock already held.
2299 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2305 page = __rmqueue_smallest(zone, order, migratetype);
2306 if (unlikely(!page)) {
2307 if (migratetype == MIGRATE_MOVABLE)
2308 page = __rmqueue_cma_fallback(zone, order);
2310 if (!page && __rmqueue_fallback(zone, order, migratetype))
2314 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2319 * Obtain a specified number of elements from the buddy allocator, all under
2320 * a single hold of the lock, for efficiency. Add them to the supplied list.
2321 * Returns the number of new pages which were placed at *list.
2323 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2324 unsigned long count, struct list_head *list,
2325 int migratetype, bool cold)
2329 spin_lock(&zone->lock);
2330 for (i = 0; i < count; ++i) {
2331 struct page *page = __rmqueue(zone, order, migratetype);
2332 if (unlikely(page == NULL))
2335 if (unlikely(check_pcp_refill(page)))
2339 * Split buddy pages returned by expand() are received here
2340 * in physical page order. The page is added to the callers and
2341 * list and the list head then moves forward. From the callers
2342 * perspective, the linked list is ordered by page number in
2343 * some conditions. This is useful for IO devices that can
2344 * merge IO requests if the physical pages are ordered
2348 list_add(&page->lru, list);
2350 list_add_tail(&page->lru, list);
2353 if (is_migrate_cma(get_pcppage_migratetype(page)))
2354 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2359 * i pages were removed from the buddy list even if some leak due
2360 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2361 * on i. Do not confuse with 'alloced' which is the number of
2362 * pages added to the pcp list.
2364 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2365 spin_unlock(&zone->lock);
2371 * Called from the vmstat counter updater to drain pagesets of this
2372 * currently executing processor on remote nodes after they have
2375 * Note that this function must be called with the thread pinned to
2376 * a single processor.
2378 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2380 unsigned long flags;
2381 int to_drain, batch;
2383 local_irq_save(flags);
2384 batch = READ_ONCE(pcp->batch);
2385 to_drain = min(pcp->count, batch);
2387 free_pcppages_bulk(zone, to_drain, pcp);
2388 pcp->count -= to_drain;
2390 local_irq_restore(flags);
2395 * Drain pcplists of the indicated processor and zone.
2397 * The processor must either be the current processor and the
2398 * thread pinned to the current processor or a processor that
2401 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2403 unsigned long flags;
2404 struct per_cpu_pageset *pset;
2405 struct per_cpu_pages *pcp;
2407 local_irq_save(flags);
2408 pset = per_cpu_ptr(zone->pageset, cpu);
2412 free_pcppages_bulk(zone, pcp->count, pcp);
2415 local_irq_restore(flags);
2419 * Drain pcplists of all zones on the indicated processor.
2421 * The processor must either be the current processor and the
2422 * thread pinned to the current processor or a processor that
2425 static void drain_pages(unsigned int cpu)
2429 for_each_populated_zone(zone) {
2430 drain_pages_zone(cpu, zone);
2435 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2437 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2438 * the single zone's pages.
2440 void drain_local_pages(struct zone *zone)
2442 int cpu = smp_processor_id();
2445 drain_pages_zone(cpu, zone);
2450 static void drain_local_pages_wq(struct work_struct *work)
2453 * drain_all_pages doesn't use proper cpu hotplug protection so
2454 * we can race with cpu offline when the WQ can move this from
2455 * a cpu pinned worker to an unbound one. We can operate on a different
2456 * cpu which is allright but we also have to make sure to not move to
2460 drain_local_pages(NULL);
2465 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2467 * When zone parameter is non-NULL, spill just the single zone's pages.
2469 * Note that this can be extremely slow as the draining happens in a workqueue.
2471 void drain_all_pages(struct zone *zone)
2476 * Allocate in the BSS so we wont require allocation in
2477 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2479 static cpumask_t cpus_with_pcps;
2482 * Make sure nobody triggers this path before mm_percpu_wq is fully
2485 if (WARN_ON_ONCE(!mm_percpu_wq))
2489 * Do not drain if one is already in progress unless it's specific to
2490 * a zone. Such callers are primarily CMA and memory hotplug and need
2491 * the drain to be complete when the call returns.
2493 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2496 mutex_lock(&pcpu_drain_mutex);
2500 * We don't care about racing with CPU hotplug event
2501 * as offline notification will cause the notified
2502 * cpu to drain that CPU pcps and on_each_cpu_mask
2503 * disables preemption as part of its processing
2505 for_each_online_cpu(cpu) {
2506 struct per_cpu_pageset *pcp;
2508 bool has_pcps = false;
2511 pcp = per_cpu_ptr(zone->pageset, cpu);
2515 for_each_populated_zone(z) {
2516 pcp = per_cpu_ptr(z->pageset, cpu);
2517 if (pcp->pcp.count) {
2525 cpumask_set_cpu(cpu, &cpus_with_pcps);
2527 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2530 for_each_cpu(cpu, &cpus_with_pcps) {
2531 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2532 INIT_WORK(work, drain_local_pages_wq);
2533 queue_work_on(cpu, mm_percpu_wq, work);
2535 for_each_cpu(cpu, &cpus_with_pcps)
2536 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2538 mutex_unlock(&pcpu_drain_mutex);
2541 #ifdef CONFIG_HIBERNATION
2544 * Touch the watchdog for every WD_PAGE_COUNT pages.
2546 #define WD_PAGE_COUNT (128*1024)
2548 void mark_free_pages(struct zone *zone)
2550 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2551 unsigned long flags;
2552 unsigned int order, t;
2555 if (zone_is_empty(zone))
2558 spin_lock_irqsave(&zone->lock, flags);
2560 max_zone_pfn = zone_end_pfn(zone);
2561 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2562 if (pfn_valid(pfn)) {
2563 page = pfn_to_page(pfn);
2565 if (!--page_count) {
2566 touch_nmi_watchdog();
2567 page_count = WD_PAGE_COUNT;
2570 if (page_zone(page) != zone)
2573 if (!swsusp_page_is_forbidden(page))
2574 swsusp_unset_page_free(page);
2577 for_each_migratetype_order(order, t) {
2578 list_for_each_entry(page,
2579 &zone->free_area[order].free_list[t], lru) {
2582 pfn = page_to_pfn(page);
2583 for (i = 0; i < (1UL << order); i++) {
2584 if (!--page_count) {
2585 touch_nmi_watchdog();
2586 page_count = WD_PAGE_COUNT;
2588 swsusp_set_page_free(pfn_to_page(pfn + i));
2592 spin_unlock_irqrestore(&zone->lock, flags);
2594 #endif /* CONFIG_PM */
2597 * Free a 0-order page
2598 * cold == true ? free a cold page : free a hot page
2600 void free_hot_cold_page(struct page *page, bool cold)
2602 struct zone *zone = page_zone(page);
2603 struct per_cpu_pages *pcp;
2604 unsigned long flags;
2605 unsigned long pfn = page_to_pfn(page);
2608 if (!free_pcp_prepare(page))
2611 migratetype = get_pfnblock_migratetype(page, pfn);
2612 set_pcppage_migratetype(page, migratetype);
2613 local_irq_save(flags);
2614 __count_vm_event(PGFREE);
2617 * We only track unmovable, reclaimable and movable on pcp lists.
2618 * Free ISOLATE pages back to the allocator because they are being
2619 * offlined but treat HIGHATOMIC as movable pages so we can get those
2620 * areas back if necessary. Otherwise, we may have to free
2621 * excessively into the page allocator
2623 if (migratetype >= MIGRATE_PCPTYPES) {
2624 if (unlikely(is_migrate_isolate(migratetype))) {
2625 free_one_page(zone, page, pfn, 0, migratetype);
2628 migratetype = MIGRATE_MOVABLE;
2631 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2633 list_add(&page->lru, &pcp->lists[migratetype]);
2635 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2637 if (pcp->count >= pcp->high) {
2638 unsigned long batch = READ_ONCE(pcp->batch);
2639 free_pcppages_bulk(zone, batch, pcp);
2640 pcp->count -= batch;
2644 local_irq_restore(flags);
2648 * Free a list of 0-order pages
2650 void free_hot_cold_page_list(struct list_head *list, bool cold)
2652 struct page *page, *next;
2654 list_for_each_entry_safe(page, next, list, lru) {
2655 trace_mm_page_free_batched(page, cold);
2656 free_hot_cold_page(page, cold);
2661 * split_page takes a non-compound higher-order page, and splits it into
2662 * n (1<<order) sub-pages: page[0..n]
2663 * Each sub-page must be freed individually.
2665 * Note: this is probably too low level an operation for use in drivers.
2666 * Please consult with lkml before using this in your driver.
2668 void split_page(struct page *page, unsigned int order)
2672 VM_BUG_ON_PAGE(PageCompound(page), page);
2673 VM_BUG_ON_PAGE(!page_count(page), page);
2675 for (i = 1; i < (1 << order); i++)
2676 set_page_refcounted(page + i);
2677 split_page_owner(page, order);
2679 EXPORT_SYMBOL_GPL(split_page);
2681 int __isolate_free_page(struct page *page, unsigned int order)
2683 unsigned long watermark;
2687 BUG_ON(!PageBuddy(page));
2689 zone = page_zone(page);
2690 mt = get_pageblock_migratetype(page);
2692 if (!is_migrate_isolate(mt)) {
2694 * Obey watermarks as if the page was being allocated. We can
2695 * emulate a high-order watermark check with a raised order-0
2696 * watermark, because we already know our high-order page
2699 watermark = min_wmark_pages(zone) + (1UL << order);
2700 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2703 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2706 /* Remove page from free list */
2707 list_del(&page->lru);
2708 zone->free_area[order].nr_free--;
2709 rmv_page_order(page);
2712 * Set the pageblock if the isolated page is at least half of a
2715 if (order >= pageblock_order - 1) {
2716 struct page *endpage = page + (1 << order) - 1;
2717 for (; page < endpage; page += pageblock_nr_pages) {
2718 int mt = get_pageblock_migratetype(page);
2719 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2720 && !is_migrate_highatomic(mt))
2721 set_pageblock_migratetype(page,
2727 return 1UL << order;
2731 * Update NUMA hit/miss statistics
2733 * Must be called with interrupts disabled.
2735 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2738 enum numa_stat_item local_stat = NUMA_LOCAL;
2740 if (z->node != numa_node_id())
2741 local_stat = NUMA_OTHER;
2743 if (z->node == preferred_zone->node)
2744 __inc_numa_state(z, NUMA_HIT);
2746 __inc_numa_state(z, NUMA_MISS);
2747 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2749 __inc_numa_state(z, local_stat);
2753 /* Remove page from the per-cpu list, caller must protect the list */
2754 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2755 bool cold, struct per_cpu_pages *pcp,
2756 struct list_head *list)
2761 if (list_empty(list)) {
2762 pcp->count += rmqueue_bulk(zone, 0,
2765 if (unlikely(list_empty(list)))
2770 page = list_last_entry(list, struct page, lru);
2772 page = list_first_entry(list, struct page, lru);
2774 list_del(&page->lru);
2776 } while (check_new_pcp(page));
2781 /* Lock and remove page from the per-cpu list */
2782 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2783 struct zone *zone, unsigned int order,
2784 gfp_t gfp_flags, int migratetype)
2786 struct per_cpu_pages *pcp;
2787 struct list_head *list;
2788 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2790 unsigned long flags;
2792 local_irq_save(flags);
2793 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2794 list = &pcp->lists[migratetype];
2795 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2797 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2798 zone_statistics(preferred_zone, zone);
2800 local_irq_restore(flags);
2805 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2808 struct page *rmqueue(struct zone *preferred_zone,
2809 struct zone *zone, unsigned int order,
2810 gfp_t gfp_flags, unsigned int alloc_flags,
2813 unsigned long flags;
2816 if (likely(order == 0)) {
2817 page = rmqueue_pcplist(preferred_zone, zone, order,
2818 gfp_flags, migratetype);
2823 * We most definitely don't want callers attempting to
2824 * allocate greater than order-1 page units with __GFP_NOFAIL.
2826 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2827 spin_lock_irqsave(&zone->lock, flags);
2831 if (alloc_flags & ALLOC_HARDER) {
2832 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2834 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2837 page = __rmqueue(zone, order, migratetype);
2838 } while (page && check_new_pages(page, order));
2839 spin_unlock(&zone->lock);
2842 __mod_zone_freepage_state(zone, -(1 << order),
2843 get_pcppage_migratetype(page));
2845 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2846 zone_statistics(preferred_zone, zone);
2847 local_irq_restore(flags);
2850 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2854 local_irq_restore(flags);
2858 #ifdef CONFIG_FAIL_PAGE_ALLOC
2861 struct fault_attr attr;
2863 bool ignore_gfp_highmem;
2864 bool ignore_gfp_reclaim;
2866 } fail_page_alloc = {
2867 .attr = FAULT_ATTR_INITIALIZER,
2868 .ignore_gfp_reclaim = true,
2869 .ignore_gfp_highmem = true,
2873 static int __init setup_fail_page_alloc(char *str)
2875 return setup_fault_attr(&fail_page_alloc.attr, str);
2877 __setup("fail_page_alloc=", setup_fail_page_alloc);
2879 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2881 if (order < fail_page_alloc.min_order)
2883 if (gfp_mask & __GFP_NOFAIL)
2885 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2887 if (fail_page_alloc.ignore_gfp_reclaim &&
2888 (gfp_mask & __GFP_DIRECT_RECLAIM))
2891 return should_fail(&fail_page_alloc.attr, 1 << order);
2894 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2896 static int __init fail_page_alloc_debugfs(void)
2898 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2901 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2902 &fail_page_alloc.attr);
2904 return PTR_ERR(dir);
2906 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2907 &fail_page_alloc.ignore_gfp_reclaim))
2909 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2910 &fail_page_alloc.ignore_gfp_highmem))
2912 if (!debugfs_create_u32("min-order", mode, dir,
2913 &fail_page_alloc.min_order))
2918 debugfs_remove_recursive(dir);
2923 late_initcall(fail_page_alloc_debugfs);
2925 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2927 #else /* CONFIG_FAIL_PAGE_ALLOC */
2929 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2934 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2937 * Return true if free base pages are above 'mark'. For high-order checks it
2938 * will return true of the order-0 watermark is reached and there is at least
2939 * one free page of a suitable size. Checking now avoids taking the zone lock
2940 * to check in the allocation paths if no pages are free.
2942 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2943 int classzone_idx, unsigned int alloc_flags,
2948 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2950 /* free_pages may go negative - that's OK */
2951 free_pages -= (1 << order) - 1;
2953 if (alloc_flags & ALLOC_HIGH)
2957 * If the caller does not have rights to ALLOC_HARDER then subtract
2958 * the high-atomic reserves. This will over-estimate the size of the
2959 * atomic reserve but it avoids a search.
2961 if (likely(!alloc_harder)) {
2962 free_pages -= z->nr_reserved_highatomic;
2965 * OOM victims can try even harder than normal ALLOC_HARDER
2966 * users on the grounds that it's definitely going to be in
2967 * the exit path shortly and free memory. Any allocation it
2968 * makes during the free path will be small and short-lived.
2970 if (alloc_flags & ALLOC_OOM)
2978 /* If allocation can't use CMA areas don't use free CMA pages */
2979 if (!(alloc_flags & ALLOC_CMA))
2980 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2984 * Check watermarks for an order-0 allocation request. If these
2985 * are not met, then a high-order request also cannot go ahead
2986 * even if a suitable page happened to be free.
2988 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2991 /* If this is an order-0 request then the watermark is fine */
2995 /* For a high-order request, check at least one suitable page is free */
2996 for (o = order; o < MAX_ORDER; o++) {
2997 struct free_area *area = &z->free_area[o];
3003 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3004 if (!list_empty(&area->free_list[mt]))
3009 if ((alloc_flags & ALLOC_CMA) &&
3010 !list_empty(&area->free_list[MIGRATE_CMA])) {
3015 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3021 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3022 int classzone_idx, unsigned int alloc_flags)
3024 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3025 zone_page_state(z, NR_FREE_PAGES));
3028 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3029 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3031 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3035 /* If allocation can't use CMA areas don't use free CMA pages */
3036 if (!(alloc_flags & ALLOC_CMA))
3037 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3041 * Fast check for order-0 only. If this fails then the reserves
3042 * need to be calculated. There is a corner case where the check
3043 * passes but only the high-order atomic reserve are free. If
3044 * the caller is !atomic then it'll uselessly search the free
3045 * list. That corner case is then slower but it is harmless.
3047 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3050 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3054 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3055 unsigned long mark, int classzone_idx)
3057 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3059 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3060 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3062 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3067 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3069 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3072 #else /* CONFIG_NUMA */
3073 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3077 #endif /* CONFIG_NUMA */
3080 * get_page_from_freelist goes through the zonelist trying to allocate
3083 static struct page *
3084 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3085 const struct alloc_context *ac)
3087 struct zoneref *z = ac->preferred_zoneref;
3089 struct pglist_data *last_pgdat_dirty_limit = NULL;
3092 * Scan zonelist, looking for a zone with enough free.
3093 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3095 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3100 if (cpusets_enabled() &&
3101 (alloc_flags & ALLOC_CPUSET) &&
3102 !__cpuset_zone_allowed(zone, gfp_mask))
3105 * When allocating a page cache page for writing, we
3106 * want to get it from a node that is within its dirty
3107 * limit, such that no single node holds more than its
3108 * proportional share of globally allowed dirty pages.
3109 * The dirty limits take into account the node's
3110 * lowmem reserves and high watermark so that kswapd
3111 * should be able to balance it without having to
3112 * write pages from its LRU list.
3114 * XXX: For now, allow allocations to potentially
3115 * exceed the per-node dirty limit in the slowpath
3116 * (spread_dirty_pages unset) before going into reclaim,
3117 * which is important when on a NUMA setup the allowed
3118 * nodes are together not big enough to reach the
3119 * global limit. The proper fix for these situations
3120 * will require awareness of nodes in the
3121 * dirty-throttling and the flusher threads.
3123 if (ac->spread_dirty_pages) {
3124 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3127 if (!node_dirty_ok(zone->zone_pgdat)) {
3128 last_pgdat_dirty_limit = zone->zone_pgdat;
3133 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3134 if (!zone_watermark_fast(zone, order, mark,
3135 ac_classzone_idx(ac), alloc_flags)) {
3138 /* Checked here to keep the fast path fast */
3139 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3140 if (alloc_flags & ALLOC_NO_WATERMARKS)
3143 if (node_reclaim_mode == 0 ||
3144 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3147 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3149 case NODE_RECLAIM_NOSCAN:
3152 case NODE_RECLAIM_FULL:
3153 /* scanned but unreclaimable */
3156 /* did we reclaim enough */
3157 if (zone_watermark_ok(zone, order, mark,
3158 ac_classzone_idx(ac), alloc_flags))
3166 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3167 gfp_mask, alloc_flags, ac->migratetype);
3169 prep_new_page(page, order, gfp_mask, alloc_flags);
3172 * If this is a high-order atomic allocation then check
3173 * if the pageblock should be reserved for the future
3175 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3176 reserve_highatomic_pageblock(page, zone, order);
3186 * Large machines with many possible nodes should not always dump per-node
3187 * meminfo in irq context.
3189 static inline bool should_suppress_show_mem(void)
3194 ret = in_interrupt();
3199 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3201 unsigned int filter = SHOW_MEM_FILTER_NODES;
3202 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3204 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3208 * This documents exceptions given to allocations in certain
3209 * contexts that are allowed to allocate outside current's set
3212 if (!(gfp_mask & __GFP_NOMEMALLOC))
3213 if (tsk_is_oom_victim(current) ||
3214 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3215 filter &= ~SHOW_MEM_FILTER_NODES;
3216 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3217 filter &= ~SHOW_MEM_FILTER_NODES;
3219 show_mem(filter, nodemask);
3222 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3224 struct va_format vaf;
3226 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3227 DEFAULT_RATELIMIT_BURST);
3229 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3232 pr_warn("%s: ", current->comm);
3234 va_start(args, fmt);
3237 pr_cont("%pV", &vaf);
3240 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3242 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3244 pr_cont("(null)\n");
3246 cpuset_print_current_mems_allowed();
3249 warn_alloc_show_mem(gfp_mask, nodemask);
3252 static inline struct page *
3253 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3254 unsigned int alloc_flags,
3255 const struct alloc_context *ac)
3259 page = get_page_from_freelist(gfp_mask, order,
3260 alloc_flags|ALLOC_CPUSET, ac);
3262 * fallback to ignore cpuset restriction if our nodes
3266 page = get_page_from_freelist(gfp_mask, order,
3272 static inline struct page *
3273 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3274 const struct alloc_context *ac, unsigned long *did_some_progress)
3276 struct oom_control oc = {
3277 .zonelist = ac->zonelist,
3278 .nodemask = ac->nodemask,
3280 .gfp_mask = gfp_mask,
3285 *did_some_progress = 0;
3288 * Acquire the oom lock. If that fails, somebody else is
3289 * making progress for us.
3291 if (!mutex_trylock(&oom_lock)) {
3292 *did_some_progress = 1;
3293 schedule_timeout_uninterruptible(1);
3298 * Go through the zonelist yet one more time, keep very high watermark
3299 * here, this is only to catch a parallel oom killing, we must fail if
3300 * we're still under heavy pressure. But make sure that this reclaim
3301 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3302 * allocation which will never fail due to oom_lock already held.
3304 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3305 ~__GFP_DIRECT_RECLAIM, order,
3306 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3310 /* Coredumps can quickly deplete all memory reserves */
3311 if (current->flags & PF_DUMPCORE)
3313 /* The OOM killer will not help higher order allocs */
3314 if (order > PAGE_ALLOC_COSTLY_ORDER)
3317 * We have already exhausted all our reclaim opportunities without any
3318 * success so it is time to admit defeat. We will skip the OOM killer
3319 * because it is very likely that the caller has a more reasonable
3320 * fallback than shooting a random task.
3322 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3324 /* The OOM killer does not needlessly kill tasks for lowmem */
3325 if (ac->high_zoneidx < ZONE_NORMAL)
3327 if (pm_suspended_storage())
3330 * XXX: GFP_NOFS allocations should rather fail than rely on
3331 * other request to make a forward progress.
3332 * We are in an unfortunate situation where out_of_memory cannot
3333 * do much for this context but let's try it to at least get
3334 * access to memory reserved if the current task is killed (see
3335 * out_of_memory). Once filesystems are ready to handle allocation
3336 * failures more gracefully we should just bail out here.
3339 /* The OOM killer may not free memory on a specific node */
3340 if (gfp_mask & __GFP_THISNODE)
3343 /* Exhausted what can be done so it's blamo time */
3344 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3345 *did_some_progress = 1;
3348 * Help non-failing allocations by giving them access to memory
3351 if (gfp_mask & __GFP_NOFAIL)
3352 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3353 ALLOC_NO_WATERMARKS, ac);
3356 mutex_unlock(&oom_lock);
3361 * Maximum number of compaction retries wit a progress before OOM
3362 * killer is consider as the only way to move forward.
3364 #define MAX_COMPACT_RETRIES 16
3366 #ifdef CONFIG_COMPACTION
3367 /* Try memory compaction for high-order allocations before reclaim */
3368 static struct page *
3369 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3370 unsigned int alloc_flags, const struct alloc_context *ac,
3371 enum compact_priority prio, enum compact_result *compact_result)
3374 unsigned int noreclaim_flag;
3379 noreclaim_flag = memalloc_noreclaim_save();
3380 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3382 memalloc_noreclaim_restore(noreclaim_flag);
3384 if (*compact_result <= COMPACT_INACTIVE)
3388 * At least in one zone compaction wasn't deferred or skipped, so let's
3389 * count a compaction stall
3391 count_vm_event(COMPACTSTALL);
3393 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3396 struct zone *zone = page_zone(page);
3398 zone->compact_blockskip_flush = false;
3399 compaction_defer_reset(zone, order, true);
3400 count_vm_event(COMPACTSUCCESS);
3405 * It's bad if compaction run occurs and fails. The most likely reason
3406 * is that pages exist, but not enough to satisfy watermarks.
3408 count_vm_event(COMPACTFAIL);
3416 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3417 enum compact_result compact_result,
3418 enum compact_priority *compact_priority,
3419 int *compaction_retries)
3421 int max_retries = MAX_COMPACT_RETRIES;
3424 int retries = *compaction_retries;
3425 enum compact_priority priority = *compact_priority;
3430 if (compaction_made_progress(compact_result))
3431 (*compaction_retries)++;
3434 * compaction considers all the zone as desperately out of memory
3435 * so it doesn't really make much sense to retry except when the
3436 * failure could be caused by insufficient priority
3438 if (compaction_failed(compact_result))
3439 goto check_priority;
3442 * make sure the compaction wasn't deferred or didn't bail out early
3443 * due to locks contention before we declare that we should give up.
3444 * But do not retry if the given zonelist is not suitable for
3447 if (compaction_withdrawn(compact_result)) {
3448 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3453 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3454 * costly ones because they are de facto nofail and invoke OOM
3455 * killer to move on while costly can fail and users are ready
3456 * to cope with that. 1/4 retries is rather arbitrary but we
3457 * would need much more detailed feedback from compaction to
3458 * make a better decision.
3460 if (order > PAGE_ALLOC_COSTLY_ORDER)
3462 if (*compaction_retries <= max_retries) {
3468 * Make sure there are attempts at the highest priority if we exhausted
3469 * all retries or failed at the lower priorities.
3472 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3473 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3475 if (*compact_priority > min_priority) {
3476 (*compact_priority)--;
3477 *compaction_retries = 0;
3481 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3485 static inline struct page *
3486 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3487 unsigned int alloc_flags, const struct alloc_context *ac,
3488 enum compact_priority prio, enum compact_result *compact_result)
3490 *compact_result = COMPACT_SKIPPED;
3495 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3496 enum compact_result compact_result,
3497 enum compact_priority *compact_priority,
3498 int *compaction_retries)
3503 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3507 * There are setups with compaction disabled which would prefer to loop
3508 * inside the allocator rather than hit the oom killer prematurely.
3509 * Let's give them a good hope and keep retrying while the order-0
3510 * watermarks are OK.
3512 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3514 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3515 ac_classzone_idx(ac), alloc_flags))
3520 #endif /* CONFIG_COMPACTION */
3522 #ifdef CONFIG_LOCKDEP
3523 struct lockdep_map __fs_reclaim_map =
3524 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3526 static bool __need_fs_reclaim(gfp_t gfp_mask)
3528 gfp_mask = current_gfp_context(gfp_mask);
3530 /* no reclaim without waiting on it */
3531 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3534 /* this guy won't enter reclaim */
3535 if (current->flags & PF_MEMALLOC)
3538 /* We're only interested __GFP_FS allocations for now */
3539 if (!(gfp_mask & __GFP_FS))
3542 if (gfp_mask & __GFP_NOLOCKDEP)
3548 void fs_reclaim_acquire(gfp_t gfp_mask)
3550 if (__need_fs_reclaim(gfp_mask))
3551 lock_map_acquire(&__fs_reclaim_map);
3553 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3555 void fs_reclaim_release(gfp_t gfp_mask)
3557 if (__need_fs_reclaim(gfp_mask))
3558 lock_map_release(&__fs_reclaim_map);
3560 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3563 /* Perform direct synchronous page reclaim */
3565 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3566 const struct alloc_context *ac)
3568 struct reclaim_state reclaim_state;
3570 unsigned int noreclaim_flag;
3574 /* We now go into synchronous reclaim */
3575 cpuset_memory_pressure_bump();
3576 noreclaim_flag = memalloc_noreclaim_save();
3577 fs_reclaim_acquire(gfp_mask);
3578 reclaim_state.reclaimed_slab = 0;
3579 current->reclaim_state = &reclaim_state;
3581 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3584 current->reclaim_state = NULL;
3585 fs_reclaim_release(gfp_mask);
3586 memalloc_noreclaim_restore(noreclaim_flag);
3593 /* The really slow allocator path where we enter direct reclaim */
3594 static inline struct page *
3595 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3596 unsigned int alloc_flags, const struct alloc_context *ac,
3597 unsigned long *did_some_progress)
3599 struct page *page = NULL;
3600 bool drained = false;
3602 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3603 if (unlikely(!(*did_some_progress)))
3607 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3610 * If an allocation failed after direct reclaim, it could be because
3611 * pages are pinned on the per-cpu lists or in high alloc reserves.
3612 * Shrink them them and try again
3614 if (!page && !drained) {
3615 unreserve_highatomic_pageblock(ac, false);
3616 drain_all_pages(NULL);
3624 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3628 pg_data_t *last_pgdat = NULL;
3630 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3631 ac->high_zoneidx, ac->nodemask) {
3632 if (last_pgdat != zone->zone_pgdat)
3633 wakeup_kswapd(zone, order, ac->high_zoneidx);
3634 last_pgdat = zone->zone_pgdat;
3638 static inline unsigned int
3639 gfp_to_alloc_flags(gfp_t gfp_mask)
3641 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3643 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3644 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3647 * The caller may dip into page reserves a bit more if the caller
3648 * cannot run direct reclaim, or if the caller has realtime scheduling
3649 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3650 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3652 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3654 if (gfp_mask & __GFP_ATOMIC) {
3656 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3657 * if it can't schedule.
3659 if (!(gfp_mask & __GFP_NOMEMALLOC))
3660 alloc_flags |= ALLOC_HARDER;
3662 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3663 * comment for __cpuset_node_allowed().
3665 alloc_flags &= ~ALLOC_CPUSET;
3666 } else if (unlikely(rt_task(current)) && !in_interrupt())
3667 alloc_flags |= ALLOC_HARDER;
3670 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3671 alloc_flags |= ALLOC_CMA;
3676 static bool oom_reserves_allowed(struct task_struct *tsk)
3678 if (!tsk_is_oom_victim(tsk))
3682 * !MMU doesn't have oom reaper so give access to memory reserves
3683 * only to the thread with TIF_MEMDIE set
3685 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3692 * Distinguish requests which really need access to full memory
3693 * reserves from oom victims which can live with a portion of it
3695 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3697 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3699 if (gfp_mask & __GFP_MEMALLOC)
3700 return ALLOC_NO_WATERMARKS;
3701 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3702 return ALLOC_NO_WATERMARKS;
3703 if (!in_interrupt()) {
3704 if (current->flags & PF_MEMALLOC)
3705 return ALLOC_NO_WATERMARKS;
3706 else if (oom_reserves_allowed(current))
3713 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3715 return !!__gfp_pfmemalloc_flags(gfp_mask);
3719 * Checks whether it makes sense to retry the reclaim to make a forward progress
3720 * for the given allocation request.
3722 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3723 * without success, or when we couldn't even meet the watermark if we
3724 * reclaimed all remaining pages on the LRU lists.
3726 * Returns true if a retry is viable or false to enter the oom path.
3729 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3730 struct alloc_context *ac, int alloc_flags,
3731 bool did_some_progress, int *no_progress_loops)
3737 * Costly allocations might have made a progress but this doesn't mean
3738 * their order will become available due to high fragmentation so
3739 * always increment the no progress counter for them
3741 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3742 *no_progress_loops = 0;
3744 (*no_progress_loops)++;
3747 * Make sure we converge to OOM if we cannot make any progress
3748 * several times in the row.
3750 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3751 /* Before OOM, exhaust highatomic_reserve */
3752 return unreserve_highatomic_pageblock(ac, true);
3756 * Keep reclaiming pages while there is a chance this will lead
3757 * somewhere. If none of the target zones can satisfy our allocation
3758 * request even if all reclaimable pages are considered then we are
3759 * screwed and have to go OOM.
3761 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3763 unsigned long available;
3764 unsigned long reclaimable;
3765 unsigned long min_wmark = min_wmark_pages(zone);
3768 available = reclaimable = zone_reclaimable_pages(zone);
3769 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3772 * Would the allocation succeed if we reclaimed all
3773 * reclaimable pages?
3775 wmark = __zone_watermark_ok(zone, order, min_wmark,
3776 ac_classzone_idx(ac), alloc_flags, available);
3777 trace_reclaim_retry_zone(z, order, reclaimable,
3778 available, min_wmark, *no_progress_loops, wmark);
3781 * If we didn't make any progress and have a lot of
3782 * dirty + writeback pages then we should wait for
3783 * an IO to complete to slow down the reclaim and
3784 * prevent from pre mature OOM
3786 if (!did_some_progress) {
3787 unsigned long write_pending;
3789 write_pending = zone_page_state_snapshot(zone,
3790 NR_ZONE_WRITE_PENDING);
3792 if (2 * write_pending > reclaimable) {
3793 congestion_wait(BLK_RW_ASYNC, HZ/10);
3799 * Memory allocation/reclaim might be called from a WQ
3800 * context and the current implementation of the WQ
3801 * concurrency control doesn't recognize that
3802 * a particular WQ is congested if the worker thread is
3803 * looping without ever sleeping. Therefore we have to
3804 * do a short sleep here rather than calling
3807 if (current->flags & PF_WQ_WORKER)
3808 schedule_timeout_uninterruptible(1);
3820 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3823 * It's possible that cpuset's mems_allowed and the nodemask from
3824 * mempolicy don't intersect. This should be normally dealt with by
3825 * policy_nodemask(), but it's possible to race with cpuset update in
3826 * such a way the check therein was true, and then it became false
3827 * before we got our cpuset_mems_cookie here.
3828 * This assumes that for all allocations, ac->nodemask can come only
3829 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3830 * when it does not intersect with the cpuset restrictions) or the
3831 * caller can deal with a violated nodemask.
3833 if (cpusets_enabled() && ac->nodemask &&
3834 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3835 ac->nodemask = NULL;
3840 * When updating a task's mems_allowed or mempolicy nodemask, it is
3841 * possible to race with parallel threads in such a way that our
3842 * allocation can fail while the mask is being updated. If we are about
3843 * to fail, check if the cpuset changed during allocation and if so,
3846 if (read_mems_allowed_retry(cpuset_mems_cookie))
3852 static inline struct page *
3853 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3854 struct alloc_context *ac)
3856 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3857 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3858 struct page *page = NULL;
3859 unsigned int alloc_flags;
3860 unsigned long did_some_progress;
3861 enum compact_priority compact_priority;
3862 enum compact_result compact_result;
3863 int compaction_retries;
3864 int no_progress_loops;
3865 unsigned long alloc_start = jiffies;
3866 unsigned int stall_timeout = 10 * HZ;
3867 unsigned int cpuset_mems_cookie;
3871 * In the slowpath, we sanity check order to avoid ever trying to
3872 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3873 * be using allocators in order of preference for an area that is
3876 if (order >= MAX_ORDER) {
3877 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3882 * We also sanity check to catch abuse of atomic reserves being used by
3883 * callers that are not in atomic context.
3885 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3886 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3887 gfp_mask &= ~__GFP_ATOMIC;
3890 compaction_retries = 0;
3891 no_progress_loops = 0;
3892 compact_priority = DEF_COMPACT_PRIORITY;
3893 cpuset_mems_cookie = read_mems_allowed_begin();
3896 * The fast path uses conservative alloc_flags to succeed only until
3897 * kswapd needs to be woken up, and to avoid the cost of setting up
3898 * alloc_flags precisely. So we do that now.
3900 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3903 * We need to recalculate the starting point for the zonelist iterator
3904 * because we might have used different nodemask in the fast path, or
3905 * there was a cpuset modification and we are retrying - otherwise we
3906 * could end up iterating over non-eligible zones endlessly.
3908 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3909 ac->high_zoneidx, ac->nodemask);
3910 if (!ac->preferred_zoneref->zone)
3913 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3914 wake_all_kswapds(order, ac);
3917 * The adjusted alloc_flags might result in immediate success, so try
3920 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3925 * For costly allocations, try direct compaction first, as it's likely
3926 * that we have enough base pages and don't need to reclaim. For non-
3927 * movable high-order allocations, do that as well, as compaction will
3928 * try prevent permanent fragmentation by migrating from blocks of the
3930 * Don't try this for allocations that are allowed to ignore
3931 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3933 if (can_direct_reclaim &&
3935 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3936 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3937 page = __alloc_pages_direct_compact(gfp_mask, order,
3939 INIT_COMPACT_PRIORITY,
3945 * Checks for costly allocations with __GFP_NORETRY, which
3946 * includes THP page fault allocations
3948 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3950 * If compaction is deferred for high-order allocations,
3951 * it is because sync compaction recently failed. If
3952 * this is the case and the caller requested a THP
3953 * allocation, we do not want to heavily disrupt the
3954 * system, so we fail the allocation instead of entering
3957 if (compact_result == COMPACT_DEFERRED)
3961 * Looks like reclaim/compaction is worth trying, but
3962 * sync compaction could be very expensive, so keep
3963 * using async compaction.
3965 compact_priority = INIT_COMPACT_PRIORITY;
3970 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3971 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3972 wake_all_kswapds(order, ac);
3974 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
3976 alloc_flags = reserve_flags;
3979 * Reset the zonelist iterators if memory policies can be ignored.
3980 * These allocations are high priority and system rather than user
3983 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
3984 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3985 ac->high_zoneidx, ac->nodemask);
3988 /* Attempt with potentially adjusted zonelist and alloc_flags */
3989 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3993 /* Caller is not willing to reclaim, we can't balance anything */
3994 if (!can_direct_reclaim)
3997 /* Make sure we know about allocations which stall for too long */
3998 if (time_after(jiffies, alloc_start + stall_timeout)) {
3999 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
4000 "page allocation stalls for %ums, order:%u",
4001 jiffies_to_msecs(jiffies-alloc_start), order);
4002 stall_timeout += 10 * HZ;
4005 /* Avoid recursion of direct reclaim */
4006 if (current->flags & PF_MEMALLOC)
4009 /* Try direct reclaim and then allocating */
4010 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4011 &did_some_progress);
4015 /* Try direct compaction and then allocating */
4016 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4017 compact_priority, &compact_result);
4021 /* Do not loop if specifically requested */
4022 if (gfp_mask & __GFP_NORETRY)
4026 * Do not retry costly high order allocations unless they are
4027 * __GFP_RETRY_MAYFAIL
4029 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4032 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4033 did_some_progress > 0, &no_progress_loops))
4037 * It doesn't make any sense to retry for the compaction if the order-0
4038 * reclaim is not able to make any progress because the current
4039 * implementation of the compaction depends on the sufficient amount
4040 * of free memory (see __compaction_suitable)
4042 if (did_some_progress > 0 &&
4043 should_compact_retry(ac, order, alloc_flags,
4044 compact_result, &compact_priority,
4045 &compaction_retries))
4049 /* Deal with possible cpuset update races before we start OOM killing */
4050 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4053 /* Reclaim has failed us, start killing things */
4054 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4058 /* Avoid allocations with no watermarks from looping endlessly */
4059 if (tsk_is_oom_victim(current) &&
4060 (alloc_flags == ALLOC_OOM ||
4061 (gfp_mask & __GFP_NOMEMALLOC)))
4064 /* Retry as long as the OOM killer is making progress */
4065 if (did_some_progress) {
4066 no_progress_loops = 0;
4071 /* Deal with possible cpuset update races before we fail */
4072 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4076 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4079 if (gfp_mask & __GFP_NOFAIL) {
4081 * All existing users of the __GFP_NOFAIL are blockable, so warn
4082 * of any new users that actually require GFP_NOWAIT
4084 if (WARN_ON_ONCE(!can_direct_reclaim))
4088 * PF_MEMALLOC request from this context is rather bizarre
4089 * because we cannot reclaim anything and only can loop waiting
4090 * for somebody to do a work for us
4092 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4095 * non failing costly orders are a hard requirement which we
4096 * are not prepared for much so let's warn about these users
4097 * so that we can identify them and convert them to something
4100 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4103 * Help non-failing allocations by giving them access to memory
4104 * reserves but do not use ALLOC_NO_WATERMARKS because this
4105 * could deplete whole memory reserves which would just make
4106 * the situation worse
4108 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4116 warn_alloc(gfp_mask, ac->nodemask,
4117 "page allocation failure: order:%u", order);
4122 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4123 int preferred_nid, nodemask_t *nodemask,
4124 struct alloc_context *ac, gfp_t *alloc_mask,
4125 unsigned int *alloc_flags)
4127 ac->high_zoneidx = gfp_zone(gfp_mask);
4128 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4129 ac->nodemask = nodemask;
4130 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4132 if (cpusets_enabled()) {
4133 *alloc_mask |= __GFP_HARDWALL;
4135 ac->nodemask = &cpuset_current_mems_allowed;
4137 *alloc_flags |= ALLOC_CPUSET;
4140 fs_reclaim_acquire(gfp_mask);
4141 fs_reclaim_release(gfp_mask);
4143 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4145 if (should_fail_alloc_page(gfp_mask, order))
4148 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4149 *alloc_flags |= ALLOC_CMA;
4154 /* Determine whether to spread dirty pages and what the first usable zone */
4155 static inline void finalise_ac(gfp_t gfp_mask,
4156 unsigned int order, struct alloc_context *ac)
4158 /* Dirty zone balancing only done in the fast path */
4159 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4162 * The preferred zone is used for statistics but crucially it is
4163 * also used as the starting point for the zonelist iterator. It
4164 * may get reset for allocations that ignore memory policies.
4166 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4167 ac->high_zoneidx, ac->nodemask);
4171 * This is the 'heart' of the zoned buddy allocator.
4174 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4175 nodemask_t *nodemask)
4178 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4179 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4180 struct alloc_context ac = { };
4182 gfp_mask &= gfp_allowed_mask;
4183 alloc_mask = gfp_mask;
4184 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4187 finalise_ac(gfp_mask, order, &ac);
4189 /* First allocation attempt */
4190 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4195 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4196 * resp. GFP_NOIO which has to be inherited for all allocation requests
4197 * from a particular context which has been marked by
4198 * memalloc_no{fs,io}_{save,restore}.
4200 alloc_mask = current_gfp_context(gfp_mask);
4201 ac.spread_dirty_pages = false;
4204 * Restore the original nodemask if it was potentially replaced with
4205 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4207 if (unlikely(ac.nodemask != nodemask))
4208 ac.nodemask = nodemask;
4210 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4213 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4214 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4215 __free_pages(page, order);
4219 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4223 EXPORT_SYMBOL(__alloc_pages_nodemask);
4226 * Common helper functions.
4228 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4233 * __get_free_pages() returns a 32-bit address, which cannot represent
4236 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4238 page = alloc_pages(gfp_mask, order);
4241 return (unsigned long) page_address(page);
4243 EXPORT_SYMBOL(__get_free_pages);
4245 unsigned long get_zeroed_page(gfp_t gfp_mask)
4247 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4249 EXPORT_SYMBOL(get_zeroed_page);
4251 void __free_pages(struct page *page, unsigned int order)
4253 if (put_page_testzero(page)) {
4255 free_hot_cold_page(page, false);
4257 __free_pages_ok(page, order);
4261 EXPORT_SYMBOL(__free_pages);
4263 void free_pages(unsigned long addr, unsigned int order)
4266 VM_BUG_ON(!virt_addr_valid((void *)addr));
4267 __free_pages(virt_to_page((void *)addr), order);
4271 EXPORT_SYMBOL(free_pages);
4275 * An arbitrary-length arbitrary-offset area of memory which resides
4276 * within a 0 or higher order page. Multiple fragments within that page
4277 * are individually refcounted, in the page's reference counter.
4279 * The page_frag functions below provide a simple allocation framework for
4280 * page fragments. This is used by the network stack and network device
4281 * drivers to provide a backing region of memory for use as either an
4282 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4284 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4287 struct page *page = NULL;
4288 gfp_t gfp = gfp_mask;
4290 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4291 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4293 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4294 PAGE_FRAG_CACHE_MAX_ORDER);
4295 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4297 if (unlikely(!page))
4298 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4300 nc->va = page ? page_address(page) : NULL;
4305 void __page_frag_cache_drain(struct page *page, unsigned int count)
4307 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4309 if (page_ref_sub_and_test(page, count)) {
4310 unsigned int order = compound_order(page);
4313 free_hot_cold_page(page, false);
4315 __free_pages_ok(page, order);
4318 EXPORT_SYMBOL(__page_frag_cache_drain);
4320 void *page_frag_alloc(struct page_frag_cache *nc,
4321 unsigned int fragsz, gfp_t gfp_mask)
4323 unsigned int size = PAGE_SIZE;
4327 if (unlikely(!nc->va)) {
4329 page = __page_frag_cache_refill(nc, gfp_mask);
4333 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4334 /* if size can vary use size else just use PAGE_SIZE */
4337 /* Even if we own the page, we do not use atomic_set().
4338 * This would break get_page_unless_zero() users.
4340 page_ref_add(page, size - 1);
4342 /* reset page count bias and offset to start of new frag */
4343 nc->pfmemalloc = page_is_pfmemalloc(page);
4344 nc->pagecnt_bias = size;
4348 offset = nc->offset - fragsz;
4349 if (unlikely(offset < 0)) {
4350 page = virt_to_page(nc->va);
4352 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4355 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4356 /* if size can vary use size else just use PAGE_SIZE */
4359 /* OK, page count is 0, we can safely set it */
4360 set_page_count(page, size);
4362 /* reset page count bias and offset to start of new frag */
4363 nc->pagecnt_bias = size;
4364 offset = size - fragsz;
4368 nc->offset = offset;
4370 return nc->va + offset;
4372 EXPORT_SYMBOL(page_frag_alloc);
4375 * Frees a page fragment allocated out of either a compound or order 0 page.
4377 void page_frag_free(void *addr)
4379 struct page *page = virt_to_head_page(addr);
4381 if (unlikely(put_page_testzero(page)))
4382 __free_pages_ok(page, compound_order(page));
4384 EXPORT_SYMBOL(page_frag_free);
4386 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4390 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4391 unsigned long used = addr + PAGE_ALIGN(size);
4393 split_page(virt_to_page((void *)addr), order);
4394 while (used < alloc_end) {
4399 return (void *)addr;
4403 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4404 * @size: the number of bytes to allocate
4405 * @gfp_mask: GFP flags for the allocation
4407 * This function is similar to alloc_pages(), except that it allocates the
4408 * minimum number of pages to satisfy the request. alloc_pages() can only
4409 * allocate memory in power-of-two pages.
4411 * This function is also limited by MAX_ORDER.
4413 * Memory allocated by this function must be released by free_pages_exact().
4415 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4417 unsigned int order = get_order(size);
4420 addr = __get_free_pages(gfp_mask, order);
4421 return make_alloc_exact(addr, order, size);
4423 EXPORT_SYMBOL(alloc_pages_exact);
4426 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4428 * @nid: the preferred node ID where memory should be allocated
4429 * @size: the number of bytes to allocate
4430 * @gfp_mask: GFP flags for the allocation
4432 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4435 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4437 unsigned int order = get_order(size);
4438 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4441 return make_alloc_exact((unsigned long)page_address(p), order, size);
4445 * free_pages_exact - release memory allocated via alloc_pages_exact()
4446 * @virt: the value returned by alloc_pages_exact.
4447 * @size: size of allocation, same value as passed to alloc_pages_exact().
4449 * Release the memory allocated by a previous call to alloc_pages_exact.
4451 void free_pages_exact(void *virt, size_t size)
4453 unsigned long addr = (unsigned long)virt;
4454 unsigned long end = addr + PAGE_ALIGN(size);
4456 while (addr < end) {
4461 EXPORT_SYMBOL(free_pages_exact);
4464 * nr_free_zone_pages - count number of pages beyond high watermark
4465 * @offset: The zone index of the highest zone
4467 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4468 * high watermark within all zones at or below a given zone index. For each
4469 * zone, the number of pages is calculated as:
4471 * nr_free_zone_pages = managed_pages - high_pages
4473 static unsigned long nr_free_zone_pages(int offset)
4478 /* Just pick one node, since fallback list is circular */
4479 unsigned long sum = 0;
4481 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4483 for_each_zone_zonelist(zone, z, zonelist, offset) {
4484 unsigned long size = zone->managed_pages;
4485 unsigned long high = high_wmark_pages(zone);
4494 * nr_free_buffer_pages - count number of pages beyond high watermark
4496 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4497 * watermark within ZONE_DMA and ZONE_NORMAL.
4499 unsigned long nr_free_buffer_pages(void)
4501 return nr_free_zone_pages(gfp_zone(GFP_USER));
4503 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4506 * nr_free_pagecache_pages - count number of pages beyond high watermark
4508 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4509 * high watermark within all zones.
4511 unsigned long nr_free_pagecache_pages(void)
4513 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4516 static inline void show_node(struct zone *zone)
4518 if (IS_ENABLED(CONFIG_NUMA))
4519 printk("Node %d ", zone_to_nid(zone));
4522 long si_mem_available(void)
4525 unsigned long pagecache;
4526 unsigned long wmark_low = 0;
4527 unsigned long pages[NR_LRU_LISTS];
4531 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4532 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4535 wmark_low += zone->watermark[WMARK_LOW];
4538 * Estimate the amount of memory available for userspace allocations,
4539 * without causing swapping.
4541 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4544 * Not all the page cache can be freed, otherwise the system will
4545 * start swapping. Assume at least half of the page cache, or the
4546 * low watermark worth of cache, needs to stay.
4548 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4549 pagecache -= min(pagecache / 2, wmark_low);
4550 available += pagecache;
4553 * Part of the reclaimable slab consists of items that are in use,
4554 * and cannot be freed. Cap this estimate at the low watermark.
4556 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4557 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4564 EXPORT_SYMBOL_GPL(si_mem_available);
4566 void si_meminfo(struct sysinfo *val)
4568 val->totalram = totalram_pages;
4569 val->sharedram = global_node_page_state(NR_SHMEM);
4570 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4571 val->bufferram = nr_blockdev_pages();
4572 val->totalhigh = totalhigh_pages;
4573 val->freehigh = nr_free_highpages();
4574 val->mem_unit = PAGE_SIZE;
4577 EXPORT_SYMBOL(si_meminfo);
4580 void si_meminfo_node(struct sysinfo *val, int nid)
4582 int zone_type; /* needs to be signed */
4583 unsigned long managed_pages = 0;
4584 unsigned long managed_highpages = 0;
4585 unsigned long free_highpages = 0;
4586 pg_data_t *pgdat = NODE_DATA(nid);
4588 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4589 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4590 val->totalram = managed_pages;
4591 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4592 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4593 #ifdef CONFIG_HIGHMEM
4594 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4595 struct zone *zone = &pgdat->node_zones[zone_type];
4597 if (is_highmem(zone)) {
4598 managed_highpages += zone->managed_pages;
4599 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4602 val->totalhigh = managed_highpages;
4603 val->freehigh = free_highpages;
4605 val->totalhigh = managed_highpages;
4606 val->freehigh = free_highpages;
4608 val->mem_unit = PAGE_SIZE;
4613 * Determine whether the node should be displayed or not, depending on whether
4614 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4616 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4618 if (!(flags & SHOW_MEM_FILTER_NODES))
4622 * no node mask - aka implicit memory numa policy. Do not bother with
4623 * the synchronization - read_mems_allowed_begin - because we do not
4624 * have to be precise here.
4627 nodemask = &cpuset_current_mems_allowed;
4629 return !node_isset(nid, *nodemask);
4632 #define K(x) ((x) << (PAGE_SHIFT-10))
4634 static void show_migration_types(unsigned char type)
4636 static const char types[MIGRATE_TYPES] = {
4637 [MIGRATE_UNMOVABLE] = 'U',
4638 [MIGRATE_MOVABLE] = 'M',
4639 [MIGRATE_RECLAIMABLE] = 'E',
4640 [MIGRATE_HIGHATOMIC] = 'H',
4642 [MIGRATE_CMA] = 'C',
4644 #ifdef CONFIG_MEMORY_ISOLATION
4645 [MIGRATE_ISOLATE] = 'I',
4648 char tmp[MIGRATE_TYPES + 1];
4652 for (i = 0; i < MIGRATE_TYPES; i++) {
4653 if (type & (1 << i))
4658 printk(KERN_CONT "(%s) ", tmp);
4662 * Show free area list (used inside shift_scroll-lock stuff)
4663 * We also calculate the percentage fragmentation. We do this by counting the
4664 * memory on each free list with the exception of the first item on the list.
4667 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4670 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4672 unsigned long free_pcp = 0;
4677 for_each_populated_zone(zone) {
4678 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4681 for_each_online_cpu(cpu)
4682 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4685 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4686 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4687 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4688 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4689 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4690 " free:%lu free_pcp:%lu free_cma:%lu\n",
4691 global_node_page_state(NR_ACTIVE_ANON),
4692 global_node_page_state(NR_INACTIVE_ANON),
4693 global_node_page_state(NR_ISOLATED_ANON),
4694 global_node_page_state(NR_ACTIVE_FILE),
4695 global_node_page_state(NR_INACTIVE_FILE),
4696 global_node_page_state(NR_ISOLATED_FILE),
4697 global_node_page_state(NR_UNEVICTABLE),
4698 global_node_page_state(NR_FILE_DIRTY),
4699 global_node_page_state(NR_WRITEBACK),
4700 global_node_page_state(NR_UNSTABLE_NFS),
4701 global_node_page_state(NR_SLAB_RECLAIMABLE),
4702 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4703 global_node_page_state(NR_FILE_MAPPED),
4704 global_node_page_state(NR_SHMEM),
4705 global_zone_page_state(NR_PAGETABLE),
4706 global_zone_page_state(NR_BOUNCE),
4707 global_zone_page_state(NR_FREE_PAGES),
4709 global_zone_page_state(NR_FREE_CMA_PAGES));
4711 for_each_online_pgdat(pgdat) {
4712 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4716 " active_anon:%lukB"
4717 " inactive_anon:%lukB"
4718 " active_file:%lukB"
4719 " inactive_file:%lukB"
4720 " unevictable:%lukB"
4721 " isolated(anon):%lukB"
4722 " isolated(file):%lukB"
4727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4729 " shmem_pmdmapped: %lukB"
4732 " writeback_tmp:%lukB"
4734 " all_unreclaimable? %s"
4737 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4738 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4739 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4740 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4741 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4742 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4743 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4744 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4745 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4746 K(node_page_state(pgdat, NR_WRITEBACK)),
4747 K(node_page_state(pgdat, NR_SHMEM)),
4748 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4749 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4750 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4752 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4754 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4755 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4756 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4760 for_each_populated_zone(zone) {
4763 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4767 for_each_online_cpu(cpu)
4768 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4777 " active_anon:%lukB"
4778 " inactive_anon:%lukB"
4779 " active_file:%lukB"
4780 " inactive_file:%lukB"
4781 " unevictable:%lukB"
4782 " writepending:%lukB"
4786 " kernel_stack:%lukB"
4794 K(zone_page_state(zone, NR_FREE_PAGES)),
4795 K(min_wmark_pages(zone)),
4796 K(low_wmark_pages(zone)),
4797 K(high_wmark_pages(zone)),
4798 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4799 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4800 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4801 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4802 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4803 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4804 K(zone->present_pages),
4805 K(zone->managed_pages),
4806 K(zone_page_state(zone, NR_MLOCK)),
4807 zone_page_state(zone, NR_KERNEL_STACK_KB),
4808 K(zone_page_state(zone, NR_PAGETABLE)),
4809 K(zone_page_state(zone, NR_BOUNCE)),
4811 K(this_cpu_read(zone->pageset->pcp.count)),
4812 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4813 printk("lowmem_reserve[]:");
4814 for (i = 0; i < MAX_NR_ZONES; i++)
4815 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4816 printk(KERN_CONT "\n");
4819 for_each_populated_zone(zone) {
4821 unsigned long nr[MAX_ORDER], flags, total = 0;
4822 unsigned char types[MAX_ORDER];
4824 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4827 printk(KERN_CONT "%s: ", zone->name);
4829 spin_lock_irqsave(&zone->lock, flags);
4830 for (order = 0; order < MAX_ORDER; order++) {
4831 struct free_area *area = &zone->free_area[order];
4834 nr[order] = area->nr_free;
4835 total += nr[order] << order;
4838 for (type = 0; type < MIGRATE_TYPES; type++) {
4839 if (!list_empty(&area->free_list[type]))
4840 types[order] |= 1 << type;
4843 spin_unlock_irqrestore(&zone->lock, flags);
4844 for (order = 0; order < MAX_ORDER; order++) {
4845 printk(KERN_CONT "%lu*%lukB ",
4846 nr[order], K(1UL) << order);
4848 show_migration_types(types[order]);
4850 printk(KERN_CONT "= %lukB\n", K(total));
4853 hugetlb_show_meminfo();
4855 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4857 show_swap_cache_info();
4860 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4862 zoneref->zone = zone;
4863 zoneref->zone_idx = zone_idx(zone);
4867 * Builds allocation fallback zone lists.
4869 * Add all populated zones of a node to the zonelist.
4871 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4874 enum zone_type zone_type = MAX_NR_ZONES;
4879 zone = pgdat->node_zones + zone_type;
4880 if (managed_zone(zone)) {
4881 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4882 check_highest_zone(zone_type);
4884 } while (zone_type);
4891 static int __parse_numa_zonelist_order(char *s)
4894 * We used to support different zonlists modes but they turned
4895 * out to be just not useful. Let's keep the warning in place
4896 * if somebody still use the cmd line parameter so that we do
4897 * not fail it silently
4899 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4900 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4906 static __init int setup_numa_zonelist_order(char *s)
4911 return __parse_numa_zonelist_order(s);
4913 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4915 char numa_zonelist_order[] = "Node";
4918 * sysctl handler for numa_zonelist_order
4920 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4921 void __user *buffer, size_t *length,
4928 return proc_dostring(table, write, buffer, length, ppos);
4929 str = memdup_user_nul(buffer, 16);
4931 return PTR_ERR(str);
4933 ret = __parse_numa_zonelist_order(str);
4939 #define MAX_NODE_LOAD (nr_online_nodes)
4940 static int node_load[MAX_NUMNODES];
4943 * find_next_best_node - find the next node that should appear in a given node's fallback list
4944 * @node: node whose fallback list we're appending
4945 * @used_node_mask: nodemask_t of already used nodes
4947 * We use a number of factors to determine which is the next node that should
4948 * appear on a given node's fallback list. The node should not have appeared
4949 * already in @node's fallback list, and it should be the next closest node
4950 * according to the distance array (which contains arbitrary distance values
4951 * from each node to each node in the system), and should also prefer nodes
4952 * with no CPUs, since presumably they'll have very little allocation pressure
4953 * on them otherwise.
4954 * It returns -1 if no node is found.
4956 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4959 int min_val = INT_MAX;
4960 int best_node = NUMA_NO_NODE;
4961 const struct cpumask *tmp = cpumask_of_node(0);
4963 /* Use the local node if we haven't already */
4964 if (!node_isset(node, *used_node_mask)) {
4965 node_set(node, *used_node_mask);
4969 for_each_node_state(n, N_MEMORY) {
4971 /* Don't want a node to appear more than once */
4972 if (node_isset(n, *used_node_mask))
4975 /* Use the distance array to find the distance */
4976 val = node_distance(node, n);
4978 /* Penalize nodes under us ("prefer the next node") */
4981 /* Give preference to headless and unused nodes */
4982 tmp = cpumask_of_node(n);
4983 if (!cpumask_empty(tmp))
4984 val += PENALTY_FOR_NODE_WITH_CPUS;
4986 /* Slight preference for less loaded node */
4987 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4988 val += node_load[n];
4990 if (val < min_val) {
4997 node_set(best_node, *used_node_mask);
5004 * Build zonelists ordered by node and zones within node.
5005 * This results in maximum locality--normal zone overflows into local
5006 * DMA zone, if any--but risks exhausting DMA zone.
5008 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5011 struct zoneref *zonerefs;
5014 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5016 for (i = 0; i < nr_nodes; i++) {
5019 pg_data_t *node = NODE_DATA(node_order[i]);
5021 nr_zones = build_zonerefs_node(node, zonerefs);
5022 zonerefs += nr_zones;
5024 zonerefs->zone = NULL;
5025 zonerefs->zone_idx = 0;
5029 * Build gfp_thisnode zonelists
5031 static void build_thisnode_zonelists(pg_data_t *pgdat)
5033 struct zoneref *zonerefs;
5036 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5037 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5038 zonerefs += nr_zones;
5039 zonerefs->zone = NULL;
5040 zonerefs->zone_idx = 0;
5044 * Build zonelists ordered by zone and nodes within zones.
5045 * This results in conserving DMA zone[s] until all Normal memory is
5046 * exhausted, but results in overflowing to remote node while memory
5047 * may still exist in local DMA zone.
5050 static void build_zonelists(pg_data_t *pgdat)
5052 static int node_order[MAX_NUMNODES];
5053 int node, load, nr_nodes = 0;
5054 nodemask_t used_mask;
5055 int local_node, prev_node;
5057 /* NUMA-aware ordering of nodes */
5058 local_node = pgdat->node_id;
5059 load = nr_online_nodes;
5060 prev_node = local_node;
5061 nodes_clear(used_mask);
5063 memset(node_order, 0, sizeof(node_order));
5064 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5066 * We don't want to pressure a particular node.
5067 * So adding penalty to the first node in same
5068 * distance group to make it round-robin.
5070 if (node_distance(local_node, node) !=
5071 node_distance(local_node, prev_node))
5072 node_load[node] = load;
5074 node_order[nr_nodes++] = node;
5079 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5080 build_thisnode_zonelists(pgdat);
5083 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5085 * Return node id of node used for "local" allocations.
5086 * I.e., first node id of first zone in arg node's generic zonelist.
5087 * Used for initializing percpu 'numa_mem', which is used primarily
5088 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5090 int local_memory_node(int node)
5094 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5095 gfp_zone(GFP_KERNEL),
5097 return z->zone->node;
5101 static void setup_min_unmapped_ratio(void);
5102 static void setup_min_slab_ratio(void);
5103 #else /* CONFIG_NUMA */
5105 static void build_zonelists(pg_data_t *pgdat)
5107 int node, local_node;
5108 struct zoneref *zonerefs;
5111 local_node = pgdat->node_id;
5113 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5114 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5115 zonerefs += nr_zones;
5118 * Now we build the zonelist so that it contains the zones
5119 * of all the other nodes.
5120 * We don't want to pressure a particular node, so when
5121 * building the zones for node N, we make sure that the
5122 * zones coming right after the local ones are those from
5123 * node N+1 (modulo N)
5125 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5126 if (!node_online(node))
5128 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5129 zonerefs += nr_zones;
5131 for (node = 0; node < local_node; node++) {
5132 if (!node_online(node))
5134 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5135 zonerefs += nr_zones;
5138 zonerefs->zone = NULL;
5139 zonerefs->zone_idx = 0;
5142 #endif /* CONFIG_NUMA */
5145 * Boot pageset table. One per cpu which is going to be used for all
5146 * zones and all nodes. The parameters will be set in such a way
5147 * that an item put on a list will immediately be handed over to
5148 * the buddy list. This is safe since pageset manipulation is done
5149 * with interrupts disabled.
5151 * The boot_pagesets must be kept even after bootup is complete for
5152 * unused processors and/or zones. They do play a role for bootstrapping
5153 * hotplugged processors.
5155 * zoneinfo_show() and maybe other functions do
5156 * not check if the processor is online before following the pageset pointer.
5157 * Other parts of the kernel may not check if the zone is available.
5159 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5160 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5161 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5163 static void __build_all_zonelists(void *data)
5166 int __maybe_unused cpu;
5167 pg_data_t *self = data;
5168 static DEFINE_SPINLOCK(lock);
5173 memset(node_load, 0, sizeof(node_load));
5177 * This node is hotadded and no memory is yet present. So just
5178 * building zonelists is fine - no need to touch other nodes.
5180 if (self && !node_online(self->node_id)) {
5181 build_zonelists(self);
5183 for_each_online_node(nid) {
5184 pg_data_t *pgdat = NODE_DATA(nid);
5186 build_zonelists(pgdat);
5189 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5191 * We now know the "local memory node" for each node--
5192 * i.e., the node of the first zone in the generic zonelist.
5193 * Set up numa_mem percpu variable for on-line cpus. During
5194 * boot, only the boot cpu should be on-line; we'll init the
5195 * secondary cpus' numa_mem as they come on-line. During
5196 * node/memory hotplug, we'll fixup all on-line cpus.
5198 for_each_online_cpu(cpu)
5199 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5206 static noinline void __init
5207 build_all_zonelists_init(void)
5211 __build_all_zonelists(NULL);
5214 * Initialize the boot_pagesets that are going to be used
5215 * for bootstrapping processors. The real pagesets for
5216 * each zone will be allocated later when the per cpu
5217 * allocator is available.
5219 * boot_pagesets are used also for bootstrapping offline
5220 * cpus if the system is already booted because the pagesets
5221 * are needed to initialize allocators on a specific cpu too.
5222 * F.e. the percpu allocator needs the page allocator which
5223 * needs the percpu allocator in order to allocate its pagesets
5224 * (a chicken-egg dilemma).
5226 for_each_possible_cpu(cpu)
5227 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5229 mminit_verify_zonelist();
5230 cpuset_init_current_mems_allowed();
5234 * unless system_state == SYSTEM_BOOTING.
5236 * __ref due to call of __init annotated helper build_all_zonelists_init
5237 * [protected by SYSTEM_BOOTING].
5239 void __ref build_all_zonelists(pg_data_t *pgdat)
5241 if (system_state == SYSTEM_BOOTING) {
5242 build_all_zonelists_init();
5244 __build_all_zonelists(pgdat);
5245 /* cpuset refresh routine should be here */
5247 vm_total_pages = nr_free_pagecache_pages();
5249 * Disable grouping by mobility if the number of pages in the
5250 * system is too low to allow the mechanism to work. It would be
5251 * more accurate, but expensive to check per-zone. This check is
5252 * made on memory-hotadd so a system can start with mobility
5253 * disabled and enable it later
5255 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5256 page_group_by_mobility_disabled = 1;
5258 page_group_by_mobility_disabled = 0;
5260 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5262 page_group_by_mobility_disabled ? "off" : "on",
5265 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5270 * Initially all pages are reserved - free ones are freed
5271 * up by free_all_bootmem() once the early boot process is
5272 * done. Non-atomic initialization, single-pass.
5274 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5275 unsigned long start_pfn, enum memmap_context context)
5277 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5278 unsigned long end_pfn = start_pfn + size;
5279 pg_data_t *pgdat = NODE_DATA(nid);
5281 unsigned long nr_initialised = 0;
5282 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5283 struct memblock_region *r = NULL, *tmp;
5286 if (highest_memmap_pfn < end_pfn - 1)
5287 highest_memmap_pfn = end_pfn - 1;
5290 * Honor reservation requested by the driver for this ZONE_DEVICE
5293 if (altmap && start_pfn == altmap->base_pfn)
5294 start_pfn += altmap->reserve;
5296 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5298 * There can be holes in boot-time mem_map[]s handed to this
5299 * function. They do not exist on hotplugged memory.
5301 if (context != MEMMAP_EARLY)
5304 if (!early_pfn_valid(pfn))
5306 if (!early_pfn_in_nid(pfn, nid))
5308 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5311 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5313 * Check given memblock attribute by firmware which can affect
5314 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5315 * mirrored, it's an overlapped memmap init. skip it.
5317 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5318 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5319 for_each_memblock(memory, tmp)
5320 if (pfn < memblock_region_memory_end_pfn(tmp))
5324 if (pfn >= memblock_region_memory_base_pfn(r) &&
5325 memblock_is_mirror(r)) {
5326 /* already initialized as NORMAL */
5327 pfn = memblock_region_memory_end_pfn(r);
5335 * Mark the block movable so that blocks are reserved for
5336 * movable at startup. This will force kernel allocations
5337 * to reserve their blocks rather than leaking throughout
5338 * the address space during boot when many long-lived
5339 * kernel allocations are made.
5341 * bitmap is created for zone's valid pfn range. but memmap
5342 * can be created for invalid pages (for alignment)
5343 * check here not to call set_pageblock_migratetype() against
5346 if (!(pfn & (pageblock_nr_pages - 1))) {
5347 struct page *page = pfn_to_page(pfn);
5349 __init_single_page(page, pfn, zone, nid);
5350 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5353 __init_single_pfn(pfn, zone, nid);
5358 static void __meminit zone_init_free_lists(struct zone *zone)
5360 unsigned int order, t;
5361 for_each_migratetype_order(order, t) {
5362 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5363 zone->free_area[order].nr_free = 0;
5367 #ifndef __HAVE_ARCH_MEMMAP_INIT
5368 #define memmap_init(size, nid, zone, start_pfn) \
5369 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5372 static int zone_batchsize(struct zone *zone)
5378 * The per-cpu-pages pools are set to around 1000th of the
5379 * size of the zone. But no more than 1/2 of a meg.
5381 * OK, so we don't know how big the cache is. So guess.
5383 batch = zone->managed_pages / 1024;
5384 if (batch * PAGE_SIZE > 512 * 1024)
5385 batch = (512 * 1024) / PAGE_SIZE;
5386 batch /= 4; /* We effectively *= 4 below */
5391 * Clamp the batch to a 2^n - 1 value. Having a power
5392 * of 2 value was found to be more likely to have
5393 * suboptimal cache aliasing properties in some cases.
5395 * For example if 2 tasks are alternately allocating
5396 * batches of pages, one task can end up with a lot
5397 * of pages of one half of the possible page colors
5398 * and the other with pages of the other colors.
5400 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5405 /* The deferral and batching of frees should be suppressed under NOMMU
5408 * The problem is that NOMMU needs to be able to allocate large chunks
5409 * of contiguous memory as there's no hardware page translation to
5410 * assemble apparent contiguous memory from discontiguous pages.
5412 * Queueing large contiguous runs of pages for batching, however,
5413 * causes the pages to actually be freed in smaller chunks. As there
5414 * can be a significant delay between the individual batches being
5415 * recycled, this leads to the once large chunks of space being
5416 * fragmented and becoming unavailable for high-order allocations.
5423 * pcp->high and pcp->batch values are related and dependent on one another:
5424 * ->batch must never be higher then ->high.
5425 * The following function updates them in a safe manner without read side
5428 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5429 * those fields changing asynchronously (acording the the above rule).
5431 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5432 * outside of boot time (or some other assurance that no concurrent updaters
5435 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5436 unsigned long batch)
5438 /* start with a fail safe value for batch */
5442 /* Update high, then batch, in order */
5449 /* a companion to pageset_set_high() */
5450 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5452 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5455 static void pageset_init(struct per_cpu_pageset *p)
5457 struct per_cpu_pages *pcp;
5460 memset(p, 0, sizeof(*p));
5464 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5465 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5468 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5471 pageset_set_batch(p, batch);
5475 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5476 * to the value high for the pageset p.
5478 static void pageset_set_high(struct per_cpu_pageset *p,
5481 unsigned long batch = max(1UL, high / 4);
5482 if ((high / 4) > (PAGE_SHIFT * 8))
5483 batch = PAGE_SHIFT * 8;
5485 pageset_update(&p->pcp, high, batch);
5488 static void pageset_set_high_and_batch(struct zone *zone,
5489 struct per_cpu_pageset *pcp)
5491 if (percpu_pagelist_fraction)
5492 pageset_set_high(pcp,
5493 (zone->managed_pages /
5494 percpu_pagelist_fraction));
5496 pageset_set_batch(pcp, zone_batchsize(zone));
5499 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5501 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5504 pageset_set_high_and_batch(zone, pcp);
5507 void __meminit setup_zone_pageset(struct zone *zone)
5510 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5511 for_each_possible_cpu(cpu)
5512 zone_pageset_init(zone, cpu);
5516 * Allocate per cpu pagesets and initialize them.
5517 * Before this call only boot pagesets were available.
5519 void __init setup_per_cpu_pageset(void)
5521 struct pglist_data *pgdat;
5524 for_each_populated_zone(zone)
5525 setup_zone_pageset(zone);
5527 for_each_online_pgdat(pgdat)
5528 pgdat->per_cpu_nodestats =
5529 alloc_percpu(struct per_cpu_nodestat);
5532 static __meminit void zone_pcp_init(struct zone *zone)
5535 * per cpu subsystem is not up at this point. The following code
5536 * relies on the ability of the linker to provide the
5537 * offset of a (static) per cpu variable into the per cpu area.
5539 zone->pageset = &boot_pageset;
5541 if (populated_zone(zone))
5542 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5543 zone->name, zone->present_pages,
5544 zone_batchsize(zone));
5547 void __meminit init_currently_empty_zone(struct zone *zone,
5548 unsigned long zone_start_pfn,
5551 struct pglist_data *pgdat = zone->zone_pgdat;
5553 pgdat->nr_zones = zone_idx(zone) + 1;
5555 zone->zone_start_pfn = zone_start_pfn;
5557 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5558 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5560 (unsigned long)zone_idx(zone),
5561 zone_start_pfn, (zone_start_pfn + size));
5563 zone_init_free_lists(zone);
5564 zone->initialized = 1;
5567 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5568 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5571 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5573 int __meminit __early_pfn_to_nid(unsigned long pfn,
5574 struct mminit_pfnnid_cache *state)
5576 unsigned long start_pfn, end_pfn;
5579 if (state->last_start <= pfn && pfn < state->last_end)
5580 return state->last_nid;
5582 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5584 state->last_start = start_pfn;
5585 state->last_end = end_pfn;
5586 state->last_nid = nid;
5591 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5594 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5595 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5596 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5598 * If an architecture guarantees that all ranges registered contain no holes
5599 * and may be freed, this this function may be used instead of calling
5600 * memblock_free_early_nid() manually.
5602 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5604 unsigned long start_pfn, end_pfn;
5607 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5608 start_pfn = min(start_pfn, max_low_pfn);
5609 end_pfn = min(end_pfn, max_low_pfn);
5611 if (start_pfn < end_pfn)
5612 memblock_free_early_nid(PFN_PHYS(start_pfn),
5613 (end_pfn - start_pfn) << PAGE_SHIFT,
5619 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5620 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5622 * If an architecture guarantees that all ranges registered contain no holes and may
5623 * be freed, this function may be used instead of calling memory_present() manually.
5625 void __init sparse_memory_present_with_active_regions(int nid)
5627 unsigned long start_pfn, end_pfn;
5630 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5631 memory_present(this_nid, start_pfn, end_pfn);
5635 * get_pfn_range_for_nid - Return the start and end page frames for a node
5636 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5637 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5638 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5640 * It returns the start and end page frame of a node based on information
5641 * provided by memblock_set_node(). If called for a node
5642 * with no available memory, a warning is printed and the start and end
5645 void __meminit get_pfn_range_for_nid(unsigned int nid,
5646 unsigned long *start_pfn, unsigned long *end_pfn)
5648 unsigned long this_start_pfn, this_end_pfn;
5654 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5655 *start_pfn = min(*start_pfn, this_start_pfn);
5656 *end_pfn = max(*end_pfn, this_end_pfn);
5659 if (*start_pfn == -1UL)
5664 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5665 * assumption is made that zones within a node are ordered in monotonic
5666 * increasing memory addresses so that the "highest" populated zone is used
5668 static void __init find_usable_zone_for_movable(void)
5671 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5672 if (zone_index == ZONE_MOVABLE)
5675 if (arch_zone_highest_possible_pfn[zone_index] >
5676 arch_zone_lowest_possible_pfn[zone_index])
5680 VM_BUG_ON(zone_index == -1);
5681 movable_zone = zone_index;
5685 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5686 * because it is sized independent of architecture. Unlike the other zones,
5687 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5688 * in each node depending on the size of each node and how evenly kernelcore
5689 * is distributed. This helper function adjusts the zone ranges
5690 * provided by the architecture for a given node by using the end of the
5691 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5692 * zones within a node are in order of monotonic increases memory addresses
5694 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5695 unsigned long zone_type,
5696 unsigned long node_start_pfn,
5697 unsigned long node_end_pfn,
5698 unsigned long *zone_start_pfn,
5699 unsigned long *zone_end_pfn)
5701 /* Only adjust if ZONE_MOVABLE is on this node */
5702 if (zone_movable_pfn[nid]) {
5703 /* Size ZONE_MOVABLE */
5704 if (zone_type == ZONE_MOVABLE) {
5705 *zone_start_pfn = zone_movable_pfn[nid];
5706 *zone_end_pfn = min(node_end_pfn,
5707 arch_zone_highest_possible_pfn[movable_zone]);
5709 /* Adjust for ZONE_MOVABLE starting within this range */
5710 } else if (!mirrored_kernelcore &&
5711 *zone_start_pfn < zone_movable_pfn[nid] &&
5712 *zone_end_pfn > zone_movable_pfn[nid]) {
5713 *zone_end_pfn = zone_movable_pfn[nid];
5715 /* Check if this whole range is within ZONE_MOVABLE */
5716 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5717 *zone_start_pfn = *zone_end_pfn;
5722 * Return the number of pages a zone spans in a node, including holes
5723 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5725 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5726 unsigned long zone_type,
5727 unsigned long node_start_pfn,
5728 unsigned long node_end_pfn,
5729 unsigned long *zone_start_pfn,
5730 unsigned long *zone_end_pfn,
5731 unsigned long *ignored)
5733 /* When hotadd a new node from cpu_up(), the node should be empty */
5734 if (!node_start_pfn && !node_end_pfn)
5737 /* Get the start and end of the zone */
5738 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5739 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5740 adjust_zone_range_for_zone_movable(nid, zone_type,
5741 node_start_pfn, node_end_pfn,
5742 zone_start_pfn, zone_end_pfn);
5744 /* Check that this node has pages within the zone's required range */
5745 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5748 /* Move the zone boundaries inside the node if necessary */
5749 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5750 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5752 /* Return the spanned pages */
5753 return *zone_end_pfn - *zone_start_pfn;
5757 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5758 * then all holes in the requested range will be accounted for.
5760 unsigned long __meminit __absent_pages_in_range(int nid,
5761 unsigned long range_start_pfn,
5762 unsigned long range_end_pfn)
5764 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5765 unsigned long start_pfn, end_pfn;
5768 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5769 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5770 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5771 nr_absent -= end_pfn - start_pfn;
5777 * absent_pages_in_range - Return number of page frames in holes within a range
5778 * @start_pfn: The start PFN to start searching for holes
5779 * @end_pfn: The end PFN to stop searching for holes
5781 * It returns the number of pages frames in memory holes within a range.
5783 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5784 unsigned long end_pfn)
5786 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5789 /* Return the number of page frames in holes in a zone on a node */
5790 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5791 unsigned long zone_type,
5792 unsigned long node_start_pfn,
5793 unsigned long node_end_pfn,
5794 unsigned long *ignored)
5796 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5797 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5798 unsigned long zone_start_pfn, zone_end_pfn;
5799 unsigned long nr_absent;
5801 /* When hotadd a new node from cpu_up(), the node should be empty */
5802 if (!node_start_pfn && !node_end_pfn)
5805 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5806 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5808 adjust_zone_range_for_zone_movable(nid, zone_type,
5809 node_start_pfn, node_end_pfn,
5810 &zone_start_pfn, &zone_end_pfn);
5811 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5814 * ZONE_MOVABLE handling.
5815 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5818 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5819 unsigned long start_pfn, end_pfn;
5820 struct memblock_region *r;
5822 for_each_memblock(memory, r) {
5823 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5824 zone_start_pfn, zone_end_pfn);
5825 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5826 zone_start_pfn, zone_end_pfn);
5828 if (zone_type == ZONE_MOVABLE &&
5829 memblock_is_mirror(r))
5830 nr_absent += end_pfn - start_pfn;
5832 if (zone_type == ZONE_NORMAL &&
5833 !memblock_is_mirror(r))
5834 nr_absent += end_pfn - start_pfn;
5841 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5842 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5843 unsigned long zone_type,
5844 unsigned long node_start_pfn,
5845 unsigned long node_end_pfn,
5846 unsigned long *zone_start_pfn,
5847 unsigned long *zone_end_pfn,
5848 unsigned long *zones_size)
5852 *zone_start_pfn = node_start_pfn;
5853 for (zone = 0; zone < zone_type; zone++)
5854 *zone_start_pfn += zones_size[zone];
5856 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5858 return zones_size[zone_type];
5861 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5862 unsigned long zone_type,
5863 unsigned long node_start_pfn,
5864 unsigned long node_end_pfn,
5865 unsigned long *zholes_size)
5870 return zholes_size[zone_type];
5873 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5875 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5876 unsigned long node_start_pfn,
5877 unsigned long node_end_pfn,
5878 unsigned long *zones_size,
5879 unsigned long *zholes_size)
5881 unsigned long realtotalpages = 0, totalpages = 0;
5884 for (i = 0; i < MAX_NR_ZONES; i++) {
5885 struct zone *zone = pgdat->node_zones + i;
5886 unsigned long zone_start_pfn, zone_end_pfn;
5887 unsigned long size, real_size;
5889 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5895 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5896 node_start_pfn, node_end_pfn,
5899 zone->zone_start_pfn = zone_start_pfn;
5901 zone->zone_start_pfn = 0;
5902 zone->spanned_pages = size;
5903 zone->present_pages = real_size;
5906 realtotalpages += real_size;
5909 pgdat->node_spanned_pages = totalpages;
5910 pgdat->node_present_pages = realtotalpages;
5911 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5915 #ifndef CONFIG_SPARSEMEM
5917 * Calculate the size of the zone->blockflags rounded to an unsigned long
5918 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5919 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5920 * round what is now in bits to nearest long in bits, then return it in
5923 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5925 unsigned long usemapsize;
5927 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5928 usemapsize = roundup(zonesize, pageblock_nr_pages);
5929 usemapsize = usemapsize >> pageblock_order;
5930 usemapsize *= NR_PAGEBLOCK_BITS;
5931 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5933 return usemapsize / 8;
5936 static void __init setup_usemap(struct pglist_data *pgdat,
5938 unsigned long zone_start_pfn,
5939 unsigned long zonesize)
5941 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5942 zone->pageblock_flags = NULL;
5944 zone->pageblock_flags =
5945 memblock_virt_alloc_node_nopanic(usemapsize,
5949 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5950 unsigned long zone_start_pfn, unsigned long zonesize) {}
5951 #endif /* CONFIG_SPARSEMEM */
5953 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5955 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5956 void __paginginit set_pageblock_order(void)
5960 /* Check that pageblock_nr_pages has not already been setup */
5961 if (pageblock_order)
5964 if (HPAGE_SHIFT > PAGE_SHIFT)
5965 order = HUGETLB_PAGE_ORDER;
5967 order = MAX_ORDER - 1;
5970 * Assume the largest contiguous order of interest is a huge page.
5971 * This value may be variable depending on boot parameters on IA64 and
5974 pageblock_order = order;
5976 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5979 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5980 * is unused as pageblock_order is set at compile-time. See
5981 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5984 void __paginginit set_pageblock_order(void)
5988 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5990 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5991 unsigned long present_pages)
5993 unsigned long pages = spanned_pages;
5996 * Provide a more accurate estimation if there are holes within
5997 * the zone and SPARSEMEM is in use. If there are holes within the
5998 * zone, each populated memory region may cost us one or two extra
5999 * memmap pages due to alignment because memmap pages for each
6000 * populated regions may not be naturally aligned on page boundary.
6001 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6003 if (spanned_pages > present_pages + (present_pages >> 4) &&
6004 IS_ENABLED(CONFIG_SPARSEMEM))
6005 pages = present_pages;
6007 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6011 * Set up the zone data structures:
6012 * - mark all pages reserved
6013 * - mark all memory queues empty
6014 * - clear the memory bitmaps
6016 * NOTE: pgdat should get zeroed by caller.
6018 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6021 int nid = pgdat->node_id;
6023 pgdat_resize_init(pgdat);
6024 #ifdef CONFIG_NUMA_BALANCING
6025 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6026 pgdat->numabalancing_migrate_nr_pages = 0;
6027 pgdat->numabalancing_migrate_next_window = jiffies;
6029 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6030 spin_lock_init(&pgdat->split_queue_lock);
6031 INIT_LIST_HEAD(&pgdat->split_queue);
6032 pgdat->split_queue_len = 0;
6034 init_waitqueue_head(&pgdat->kswapd_wait);
6035 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6036 #ifdef CONFIG_COMPACTION
6037 init_waitqueue_head(&pgdat->kcompactd_wait);
6039 pgdat_page_ext_init(pgdat);
6040 spin_lock_init(&pgdat->lru_lock);
6041 lruvec_init(node_lruvec(pgdat));
6043 pgdat->per_cpu_nodestats = &boot_nodestats;
6045 for (j = 0; j < MAX_NR_ZONES; j++) {
6046 struct zone *zone = pgdat->node_zones + j;
6047 unsigned long size, realsize, freesize, memmap_pages;
6048 unsigned long zone_start_pfn = zone->zone_start_pfn;
6050 size = zone->spanned_pages;
6051 realsize = freesize = zone->present_pages;
6054 * Adjust freesize so that it accounts for how much memory
6055 * is used by this zone for memmap. This affects the watermark
6056 * and per-cpu initialisations
6058 memmap_pages = calc_memmap_size(size, realsize);
6059 if (!is_highmem_idx(j)) {
6060 if (freesize >= memmap_pages) {
6061 freesize -= memmap_pages;
6064 " %s zone: %lu pages used for memmap\n",
6065 zone_names[j], memmap_pages);
6067 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6068 zone_names[j], memmap_pages, freesize);
6071 /* Account for reserved pages */
6072 if (j == 0 && freesize > dma_reserve) {
6073 freesize -= dma_reserve;
6074 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6075 zone_names[0], dma_reserve);
6078 if (!is_highmem_idx(j))
6079 nr_kernel_pages += freesize;
6080 /* Charge for highmem memmap if there are enough kernel pages */
6081 else if (nr_kernel_pages > memmap_pages * 2)
6082 nr_kernel_pages -= memmap_pages;
6083 nr_all_pages += freesize;
6086 * Set an approximate value for lowmem here, it will be adjusted
6087 * when the bootmem allocator frees pages into the buddy system.
6088 * And all highmem pages will be managed by the buddy system.
6090 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6094 zone->name = zone_names[j];
6095 zone->zone_pgdat = pgdat;
6096 spin_lock_init(&zone->lock);
6097 zone_seqlock_init(zone);
6098 zone_pcp_init(zone);
6103 set_pageblock_order();
6104 setup_usemap(pgdat, zone, zone_start_pfn, size);
6105 init_currently_empty_zone(zone, zone_start_pfn, size);
6106 memmap_init(size, nid, j, zone_start_pfn);
6110 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6112 unsigned long __maybe_unused start = 0;
6113 unsigned long __maybe_unused offset = 0;
6115 /* Skip empty nodes */
6116 if (!pgdat->node_spanned_pages)
6119 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6120 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6121 offset = pgdat->node_start_pfn - start;
6122 /* ia64 gets its own node_mem_map, before this, without bootmem */
6123 if (!pgdat->node_mem_map) {
6124 unsigned long size, end;
6128 * The zone's endpoints aren't required to be MAX_ORDER
6129 * aligned but the node_mem_map endpoints must be in order
6130 * for the buddy allocator to function correctly.
6132 end = pgdat_end_pfn(pgdat);
6133 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6134 size = (end - start) * sizeof(struct page);
6135 map = alloc_remap(pgdat->node_id, size);
6137 map = memblock_virt_alloc_node_nopanic(size,
6139 pgdat->node_mem_map = map + offset;
6141 #ifndef CONFIG_NEED_MULTIPLE_NODES
6143 * With no DISCONTIG, the global mem_map is just set as node 0's
6145 if (pgdat == NODE_DATA(0)) {
6146 mem_map = NODE_DATA(0)->node_mem_map;
6147 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6148 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6150 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6153 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6156 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6157 unsigned long node_start_pfn, unsigned long *zholes_size)
6159 pg_data_t *pgdat = NODE_DATA(nid);
6160 unsigned long start_pfn = 0;
6161 unsigned long end_pfn = 0;
6163 /* pg_data_t should be reset to zero when it's allocated */
6164 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6166 pgdat->node_id = nid;
6167 pgdat->node_start_pfn = node_start_pfn;
6168 pgdat->per_cpu_nodestats = NULL;
6169 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6170 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6171 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6172 (u64)start_pfn << PAGE_SHIFT,
6173 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6175 start_pfn = node_start_pfn;
6177 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6178 zones_size, zholes_size);
6180 alloc_node_mem_map(pgdat);
6181 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6182 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6183 nid, (unsigned long)pgdat,
6184 (unsigned long)pgdat->node_mem_map);
6187 reset_deferred_meminit(pgdat);
6188 free_area_init_core(pgdat);
6191 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6193 #if MAX_NUMNODES > 1
6195 * Figure out the number of possible node ids.
6197 void __init setup_nr_node_ids(void)
6199 unsigned int highest;
6201 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6202 nr_node_ids = highest + 1;
6207 * node_map_pfn_alignment - determine the maximum internode alignment
6209 * This function should be called after node map is populated and sorted.
6210 * It calculates the maximum power of two alignment which can distinguish
6213 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6214 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6215 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6216 * shifted, 1GiB is enough and this function will indicate so.
6218 * This is used to test whether pfn -> nid mapping of the chosen memory
6219 * model has fine enough granularity to avoid incorrect mapping for the
6220 * populated node map.
6222 * Returns the determined alignment in pfn's. 0 if there is no alignment
6223 * requirement (single node).
6225 unsigned long __init node_map_pfn_alignment(void)
6227 unsigned long accl_mask = 0, last_end = 0;
6228 unsigned long start, end, mask;
6232 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6233 if (!start || last_nid < 0 || last_nid == nid) {
6240 * Start with a mask granular enough to pin-point to the
6241 * start pfn and tick off bits one-by-one until it becomes
6242 * too coarse to separate the current node from the last.
6244 mask = ~((1 << __ffs(start)) - 1);
6245 while (mask && last_end <= (start & (mask << 1)))
6248 /* accumulate all internode masks */
6252 /* convert mask to number of pages */
6253 return ~accl_mask + 1;
6256 /* Find the lowest pfn for a node */
6257 static unsigned long __init find_min_pfn_for_node(int nid)
6259 unsigned long min_pfn = ULONG_MAX;
6260 unsigned long start_pfn;
6263 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6264 min_pfn = min(min_pfn, start_pfn);
6266 if (min_pfn == ULONG_MAX) {
6267 pr_warn("Could not find start_pfn for node %d\n", nid);
6275 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6277 * It returns the minimum PFN based on information provided via
6278 * memblock_set_node().
6280 unsigned long __init find_min_pfn_with_active_regions(void)
6282 return find_min_pfn_for_node(MAX_NUMNODES);
6286 * early_calculate_totalpages()
6287 * Sum pages in active regions for movable zone.
6288 * Populate N_MEMORY for calculating usable_nodes.
6290 static unsigned long __init early_calculate_totalpages(void)
6292 unsigned long totalpages = 0;
6293 unsigned long start_pfn, end_pfn;
6296 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6297 unsigned long pages = end_pfn - start_pfn;
6299 totalpages += pages;
6301 node_set_state(nid, N_MEMORY);
6307 * Find the PFN the Movable zone begins in each node. Kernel memory
6308 * is spread evenly between nodes as long as the nodes have enough
6309 * memory. When they don't, some nodes will have more kernelcore than
6312 static void __init find_zone_movable_pfns_for_nodes(void)
6315 unsigned long usable_startpfn;
6316 unsigned long kernelcore_node, kernelcore_remaining;
6317 /* save the state before borrow the nodemask */
6318 nodemask_t saved_node_state = node_states[N_MEMORY];
6319 unsigned long totalpages = early_calculate_totalpages();
6320 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6321 struct memblock_region *r;
6323 /* Need to find movable_zone earlier when movable_node is specified. */
6324 find_usable_zone_for_movable();
6327 * If movable_node is specified, ignore kernelcore and movablecore
6330 if (movable_node_is_enabled()) {
6331 for_each_memblock(memory, r) {
6332 if (!memblock_is_hotpluggable(r))
6337 usable_startpfn = PFN_DOWN(r->base);
6338 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6339 min(usable_startpfn, zone_movable_pfn[nid]) :
6347 * If kernelcore=mirror is specified, ignore movablecore option
6349 if (mirrored_kernelcore) {
6350 bool mem_below_4gb_not_mirrored = false;
6352 for_each_memblock(memory, r) {
6353 if (memblock_is_mirror(r))
6358 usable_startpfn = memblock_region_memory_base_pfn(r);
6360 if (usable_startpfn < 0x100000) {
6361 mem_below_4gb_not_mirrored = true;
6365 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6366 min(usable_startpfn, zone_movable_pfn[nid]) :
6370 if (mem_below_4gb_not_mirrored)
6371 pr_warn("This configuration results in unmirrored kernel memory.");
6377 * If movablecore=nn[KMG] was specified, calculate what size of
6378 * kernelcore that corresponds so that memory usable for
6379 * any allocation type is evenly spread. If both kernelcore
6380 * and movablecore are specified, then the value of kernelcore
6381 * will be used for required_kernelcore if it's greater than
6382 * what movablecore would have allowed.
6384 if (required_movablecore) {
6385 unsigned long corepages;
6388 * Round-up so that ZONE_MOVABLE is at least as large as what
6389 * was requested by the user
6391 required_movablecore =
6392 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6393 required_movablecore = min(totalpages, required_movablecore);
6394 corepages = totalpages - required_movablecore;
6396 required_kernelcore = max(required_kernelcore, corepages);
6400 * If kernelcore was not specified or kernelcore size is larger
6401 * than totalpages, there is no ZONE_MOVABLE.
6403 if (!required_kernelcore || required_kernelcore >= totalpages)
6406 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6407 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6410 /* Spread kernelcore memory as evenly as possible throughout nodes */
6411 kernelcore_node = required_kernelcore / usable_nodes;
6412 for_each_node_state(nid, N_MEMORY) {
6413 unsigned long start_pfn, end_pfn;
6416 * Recalculate kernelcore_node if the division per node
6417 * now exceeds what is necessary to satisfy the requested
6418 * amount of memory for the kernel
6420 if (required_kernelcore < kernelcore_node)
6421 kernelcore_node = required_kernelcore / usable_nodes;
6424 * As the map is walked, we track how much memory is usable
6425 * by the kernel using kernelcore_remaining. When it is
6426 * 0, the rest of the node is usable by ZONE_MOVABLE
6428 kernelcore_remaining = kernelcore_node;
6430 /* Go through each range of PFNs within this node */
6431 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6432 unsigned long size_pages;
6434 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6435 if (start_pfn >= end_pfn)
6438 /* Account for what is only usable for kernelcore */
6439 if (start_pfn < usable_startpfn) {
6440 unsigned long kernel_pages;
6441 kernel_pages = min(end_pfn, usable_startpfn)
6444 kernelcore_remaining -= min(kernel_pages,
6445 kernelcore_remaining);
6446 required_kernelcore -= min(kernel_pages,
6447 required_kernelcore);
6449 /* Continue if range is now fully accounted */
6450 if (end_pfn <= usable_startpfn) {
6453 * Push zone_movable_pfn to the end so
6454 * that if we have to rebalance
6455 * kernelcore across nodes, we will
6456 * not double account here
6458 zone_movable_pfn[nid] = end_pfn;
6461 start_pfn = usable_startpfn;
6465 * The usable PFN range for ZONE_MOVABLE is from
6466 * start_pfn->end_pfn. Calculate size_pages as the
6467 * number of pages used as kernelcore
6469 size_pages = end_pfn - start_pfn;
6470 if (size_pages > kernelcore_remaining)
6471 size_pages = kernelcore_remaining;
6472 zone_movable_pfn[nid] = start_pfn + size_pages;
6475 * Some kernelcore has been met, update counts and
6476 * break if the kernelcore for this node has been
6479 required_kernelcore -= min(required_kernelcore,
6481 kernelcore_remaining -= size_pages;
6482 if (!kernelcore_remaining)
6488 * If there is still required_kernelcore, we do another pass with one
6489 * less node in the count. This will push zone_movable_pfn[nid] further
6490 * along on the nodes that still have memory until kernelcore is
6494 if (usable_nodes && required_kernelcore > usable_nodes)
6498 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6499 for (nid = 0; nid < MAX_NUMNODES; nid++)
6500 zone_movable_pfn[nid] =
6501 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6504 /* restore the node_state */
6505 node_states[N_MEMORY] = saved_node_state;
6508 /* Any regular or high memory on that node ? */
6509 static void check_for_memory(pg_data_t *pgdat, int nid)
6511 enum zone_type zone_type;
6513 if (N_MEMORY == N_NORMAL_MEMORY)
6516 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6517 struct zone *zone = &pgdat->node_zones[zone_type];
6518 if (populated_zone(zone)) {
6519 node_set_state(nid, N_HIGH_MEMORY);
6520 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6521 zone_type <= ZONE_NORMAL)
6522 node_set_state(nid, N_NORMAL_MEMORY);
6529 * free_area_init_nodes - Initialise all pg_data_t and zone data
6530 * @max_zone_pfn: an array of max PFNs for each zone
6532 * This will call free_area_init_node() for each active node in the system.
6533 * Using the page ranges provided by memblock_set_node(), the size of each
6534 * zone in each node and their holes is calculated. If the maximum PFN
6535 * between two adjacent zones match, it is assumed that the zone is empty.
6536 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6537 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6538 * starts where the previous one ended. For example, ZONE_DMA32 starts
6539 * at arch_max_dma_pfn.
6541 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6543 unsigned long start_pfn, end_pfn;
6546 /* Record where the zone boundaries are */
6547 memset(arch_zone_lowest_possible_pfn, 0,
6548 sizeof(arch_zone_lowest_possible_pfn));
6549 memset(arch_zone_highest_possible_pfn, 0,
6550 sizeof(arch_zone_highest_possible_pfn));
6552 start_pfn = find_min_pfn_with_active_regions();
6554 for (i = 0; i < MAX_NR_ZONES; i++) {
6555 if (i == ZONE_MOVABLE)
6558 end_pfn = max(max_zone_pfn[i], start_pfn);
6559 arch_zone_lowest_possible_pfn[i] = start_pfn;
6560 arch_zone_highest_possible_pfn[i] = end_pfn;
6562 start_pfn = end_pfn;
6565 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6566 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6567 find_zone_movable_pfns_for_nodes();
6569 /* Print out the zone ranges */
6570 pr_info("Zone ranges:\n");
6571 for (i = 0; i < MAX_NR_ZONES; i++) {
6572 if (i == ZONE_MOVABLE)
6574 pr_info(" %-8s ", zone_names[i]);
6575 if (arch_zone_lowest_possible_pfn[i] ==
6576 arch_zone_highest_possible_pfn[i])
6579 pr_cont("[mem %#018Lx-%#018Lx]\n",
6580 (u64)arch_zone_lowest_possible_pfn[i]
6582 ((u64)arch_zone_highest_possible_pfn[i]
6583 << PAGE_SHIFT) - 1);
6586 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6587 pr_info("Movable zone start for each node\n");
6588 for (i = 0; i < MAX_NUMNODES; i++) {
6589 if (zone_movable_pfn[i])
6590 pr_info(" Node %d: %#018Lx\n", i,
6591 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6594 /* Print out the early node map */
6595 pr_info("Early memory node ranges\n");
6596 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6597 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6598 (u64)start_pfn << PAGE_SHIFT,
6599 ((u64)end_pfn << PAGE_SHIFT) - 1);
6601 /* Initialise every node */
6602 mminit_verify_pageflags_layout();
6603 setup_nr_node_ids();
6604 for_each_online_node(nid) {
6605 pg_data_t *pgdat = NODE_DATA(nid);
6606 free_area_init_node(nid, NULL,
6607 find_min_pfn_for_node(nid), NULL);
6609 /* Any memory on that node */
6610 if (pgdat->node_present_pages)
6611 node_set_state(nid, N_MEMORY);
6612 check_for_memory(pgdat, nid);
6616 static int __init cmdline_parse_core(char *p, unsigned long *core)
6618 unsigned long long coremem;
6622 coremem = memparse(p, &p);
6623 *core = coremem >> PAGE_SHIFT;
6625 /* Paranoid check that UL is enough for the coremem value */
6626 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6632 * kernelcore=size sets the amount of memory for use for allocations that
6633 * cannot be reclaimed or migrated.
6635 static int __init cmdline_parse_kernelcore(char *p)
6637 /* parse kernelcore=mirror */
6638 if (parse_option_str(p, "mirror")) {
6639 mirrored_kernelcore = true;
6643 return cmdline_parse_core(p, &required_kernelcore);
6647 * movablecore=size sets the amount of memory for use for allocations that
6648 * can be reclaimed or migrated.
6650 static int __init cmdline_parse_movablecore(char *p)
6652 return cmdline_parse_core(p, &required_movablecore);
6655 early_param("kernelcore", cmdline_parse_kernelcore);
6656 early_param("movablecore", cmdline_parse_movablecore);
6658 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6660 void adjust_managed_page_count(struct page *page, long count)
6662 spin_lock(&managed_page_count_lock);
6663 page_zone(page)->managed_pages += count;
6664 totalram_pages += count;
6665 #ifdef CONFIG_HIGHMEM
6666 if (PageHighMem(page))
6667 totalhigh_pages += count;
6669 spin_unlock(&managed_page_count_lock);
6671 EXPORT_SYMBOL(adjust_managed_page_count);
6673 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6676 unsigned long pages = 0;
6678 start = (void *)PAGE_ALIGN((unsigned long)start);
6679 end = (void *)((unsigned long)end & PAGE_MASK);
6680 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6681 if ((unsigned int)poison <= 0xFF)
6682 memset(pos, poison, PAGE_SIZE);
6683 free_reserved_page(virt_to_page(pos));
6687 pr_info("Freeing %s memory: %ldK\n",
6688 s, pages << (PAGE_SHIFT - 10));
6692 EXPORT_SYMBOL(free_reserved_area);
6694 #ifdef CONFIG_HIGHMEM
6695 void free_highmem_page(struct page *page)
6697 __free_reserved_page(page);
6699 page_zone(page)->managed_pages++;
6705 void __init mem_init_print_info(const char *str)
6707 unsigned long physpages, codesize, datasize, rosize, bss_size;
6708 unsigned long init_code_size, init_data_size;
6710 physpages = get_num_physpages();
6711 codesize = _etext - _stext;
6712 datasize = _edata - _sdata;
6713 rosize = __end_rodata - __start_rodata;
6714 bss_size = __bss_stop - __bss_start;
6715 init_data_size = __init_end - __init_begin;
6716 init_code_size = _einittext - _sinittext;
6719 * Detect special cases and adjust section sizes accordingly:
6720 * 1) .init.* may be embedded into .data sections
6721 * 2) .init.text.* may be out of [__init_begin, __init_end],
6722 * please refer to arch/tile/kernel/vmlinux.lds.S.
6723 * 3) .rodata.* may be embedded into .text or .data sections.
6725 #define adj_init_size(start, end, size, pos, adj) \
6727 if (start <= pos && pos < end && size > adj) \
6731 adj_init_size(__init_begin, __init_end, init_data_size,
6732 _sinittext, init_code_size);
6733 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6734 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6735 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6736 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6738 #undef adj_init_size
6740 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6741 #ifdef CONFIG_HIGHMEM
6745 nr_free_pages() << (PAGE_SHIFT - 10),
6746 physpages << (PAGE_SHIFT - 10),
6747 codesize >> 10, datasize >> 10, rosize >> 10,
6748 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6749 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6750 totalcma_pages << (PAGE_SHIFT - 10),
6751 #ifdef CONFIG_HIGHMEM
6752 totalhigh_pages << (PAGE_SHIFT - 10),
6754 str ? ", " : "", str ? str : "");
6758 * set_dma_reserve - set the specified number of pages reserved in the first zone
6759 * @new_dma_reserve: The number of pages to mark reserved
6761 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6762 * In the DMA zone, a significant percentage may be consumed by kernel image
6763 * and other unfreeable allocations which can skew the watermarks badly. This
6764 * function may optionally be used to account for unfreeable pages in the
6765 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6766 * smaller per-cpu batchsize.
6768 void __init set_dma_reserve(unsigned long new_dma_reserve)
6770 dma_reserve = new_dma_reserve;
6773 void __init free_area_init(unsigned long *zones_size)
6775 free_area_init_node(0, zones_size,
6776 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6779 static int page_alloc_cpu_dead(unsigned int cpu)
6782 lru_add_drain_cpu(cpu);
6786 * Spill the event counters of the dead processor
6787 * into the current processors event counters.
6788 * This artificially elevates the count of the current
6791 vm_events_fold_cpu(cpu);
6794 * Zero the differential counters of the dead processor
6795 * so that the vm statistics are consistent.
6797 * This is only okay since the processor is dead and cannot
6798 * race with what we are doing.
6800 cpu_vm_stats_fold(cpu);
6804 void __init page_alloc_init(void)
6808 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6809 "mm/page_alloc:dead", NULL,
6810 page_alloc_cpu_dead);
6815 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6816 * or min_free_kbytes changes.
6818 static void calculate_totalreserve_pages(void)
6820 struct pglist_data *pgdat;
6821 unsigned long reserve_pages = 0;
6822 enum zone_type i, j;
6824 for_each_online_pgdat(pgdat) {
6826 pgdat->totalreserve_pages = 0;
6828 for (i = 0; i < MAX_NR_ZONES; i++) {
6829 struct zone *zone = pgdat->node_zones + i;
6832 /* Find valid and maximum lowmem_reserve in the zone */
6833 for (j = i; j < MAX_NR_ZONES; j++) {
6834 if (zone->lowmem_reserve[j] > max)
6835 max = zone->lowmem_reserve[j];
6838 /* we treat the high watermark as reserved pages. */
6839 max += high_wmark_pages(zone);
6841 if (max > zone->managed_pages)
6842 max = zone->managed_pages;
6844 pgdat->totalreserve_pages += max;
6846 reserve_pages += max;
6849 totalreserve_pages = reserve_pages;
6853 * setup_per_zone_lowmem_reserve - called whenever
6854 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6855 * has a correct pages reserved value, so an adequate number of
6856 * pages are left in the zone after a successful __alloc_pages().
6858 static void setup_per_zone_lowmem_reserve(void)
6860 struct pglist_data *pgdat;
6861 enum zone_type j, idx;
6863 for_each_online_pgdat(pgdat) {
6864 for (j = 0; j < MAX_NR_ZONES; j++) {
6865 struct zone *zone = pgdat->node_zones + j;
6866 unsigned long managed_pages = zone->managed_pages;
6868 zone->lowmem_reserve[j] = 0;
6872 struct zone *lower_zone;
6876 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6877 sysctl_lowmem_reserve_ratio[idx] = 1;
6879 lower_zone = pgdat->node_zones + idx;
6880 lower_zone->lowmem_reserve[j] = managed_pages /
6881 sysctl_lowmem_reserve_ratio[idx];
6882 managed_pages += lower_zone->managed_pages;
6887 /* update totalreserve_pages */
6888 calculate_totalreserve_pages();
6891 static void __setup_per_zone_wmarks(void)
6893 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6894 unsigned long lowmem_pages = 0;
6896 unsigned long flags;
6898 /* Calculate total number of !ZONE_HIGHMEM pages */
6899 for_each_zone(zone) {
6900 if (!is_highmem(zone))
6901 lowmem_pages += zone->managed_pages;
6904 for_each_zone(zone) {
6907 spin_lock_irqsave(&zone->lock, flags);
6908 tmp = (u64)pages_min * zone->managed_pages;
6909 do_div(tmp, lowmem_pages);
6910 if (is_highmem(zone)) {
6912 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6913 * need highmem pages, so cap pages_min to a small
6916 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6917 * deltas control asynch page reclaim, and so should
6918 * not be capped for highmem.
6920 unsigned long min_pages;
6922 min_pages = zone->managed_pages / 1024;
6923 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6924 zone->watermark[WMARK_MIN] = min_pages;
6927 * If it's a lowmem zone, reserve a number of pages
6928 * proportionate to the zone's size.
6930 zone->watermark[WMARK_MIN] = tmp;
6934 * Set the kswapd watermarks distance according to the
6935 * scale factor in proportion to available memory, but
6936 * ensure a minimum size on small systems.
6938 tmp = max_t(u64, tmp >> 2,
6939 mult_frac(zone->managed_pages,
6940 watermark_scale_factor, 10000));
6942 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6943 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6945 spin_unlock_irqrestore(&zone->lock, flags);
6948 /* update totalreserve_pages */
6949 calculate_totalreserve_pages();
6953 * setup_per_zone_wmarks - called when min_free_kbytes changes
6954 * or when memory is hot-{added|removed}
6956 * Ensures that the watermark[min,low,high] values for each zone are set
6957 * correctly with respect to min_free_kbytes.
6959 void setup_per_zone_wmarks(void)
6961 static DEFINE_SPINLOCK(lock);
6964 __setup_per_zone_wmarks();
6969 * Initialise min_free_kbytes.
6971 * For small machines we want it small (128k min). For large machines
6972 * we want it large (64MB max). But it is not linear, because network
6973 * bandwidth does not increase linearly with machine size. We use
6975 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6976 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6992 int __meminit init_per_zone_wmark_min(void)
6994 unsigned long lowmem_kbytes;
6995 int new_min_free_kbytes;
6997 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6998 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7000 if (new_min_free_kbytes > user_min_free_kbytes) {
7001 min_free_kbytes = new_min_free_kbytes;
7002 if (min_free_kbytes < 128)
7003 min_free_kbytes = 128;
7004 if (min_free_kbytes > 65536)
7005 min_free_kbytes = 65536;
7007 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7008 new_min_free_kbytes, user_min_free_kbytes);
7010 setup_per_zone_wmarks();
7011 refresh_zone_stat_thresholds();
7012 setup_per_zone_lowmem_reserve();
7015 setup_min_unmapped_ratio();
7016 setup_min_slab_ratio();
7021 core_initcall(init_per_zone_wmark_min)
7024 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7025 * that we can call two helper functions whenever min_free_kbytes
7028 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7029 void __user *buffer, size_t *length, loff_t *ppos)
7033 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7038 user_min_free_kbytes = min_free_kbytes;
7039 setup_per_zone_wmarks();
7044 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7045 void __user *buffer, size_t *length, loff_t *ppos)
7049 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7054 setup_per_zone_wmarks();
7060 static void setup_min_unmapped_ratio(void)
7065 for_each_online_pgdat(pgdat)
7066 pgdat->min_unmapped_pages = 0;
7069 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7070 sysctl_min_unmapped_ratio) / 100;
7074 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7075 void __user *buffer, size_t *length, loff_t *ppos)
7079 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7083 setup_min_unmapped_ratio();
7088 static void setup_min_slab_ratio(void)
7093 for_each_online_pgdat(pgdat)
7094 pgdat->min_slab_pages = 0;
7097 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7098 sysctl_min_slab_ratio) / 100;
7101 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7102 void __user *buffer, size_t *length, loff_t *ppos)
7106 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7110 setup_min_slab_ratio();
7117 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7118 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7119 * whenever sysctl_lowmem_reserve_ratio changes.
7121 * The reserve ratio obviously has absolutely no relation with the
7122 * minimum watermarks. The lowmem reserve ratio can only make sense
7123 * if in function of the boot time zone sizes.
7125 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7126 void __user *buffer, size_t *length, loff_t *ppos)
7128 proc_dointvec_minmax(table, write, buffer, length, ppos);
7129 setup_per_zone_lowmem_reserve();
7134 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7135 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7136 * pagelist can have before it gets flushed back to buddy allocator.
7138 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7139 void __user *buffer, size_t *length, loff_t *ppos)
7142 int old_percpu_pagelist_fraction;
7145 mutex_lock(&pcp_batch_high_lock);
7146 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7148 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7149 if (!write || ret < 0)
7152 /* Sanity checking to avoid pcp imbalance */
7153 if (percpu_pagelist_fraction &&
7154 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7155 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7161 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7164 for_each_populated_zone(zone) {
7167 for_each_possible_cpu(cpu)
7168 pageset_set_high_and_batch(zone,
7169 per_cpu_ptr(zone->pageset, cpu));
7172 mutex_unlock(&pcp_batch_high_lock);
7177 int hashdist = HASHDIST_DEFAULT;
7179 static int __init set_hashdist(char *str)
7183 hashdist = simple_strtoul(str, &str, 0);
7186 __setup("hashdist=", set_hashdist);
7189 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7191 * Returns the number of pages that arch has reserved but
7192 * is not known to alloc_large_system_hash().
7194 static unsigned long __init arch_reserved_kernel_pages(void)
7201 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7202 * machines. As memory size is increased the scale is also increased but at
7203 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7204 * quadruples the scale is increased by one, which means the size of hash table
7205 * only doubles, instead of quadrupling as well.
7206 * Because 32-bit systems cannot have large physical memory, where this scaling
7207 * makes sense, it is disabled on such platforms.
7209 #if __BITS_PER_LONG > 32
7210 #define ADAPT_SCALE_BASE (64ul << 30)
7211 #define ADAPT_SCALE_SHIFT 2
7212 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7216 * allocate a large system hash table from bootmem
7217 * - it is assumed that the hash table must contain an exact power-of-2
7218 * quantity of entries
7219 * - limit is the number of hash buckets, not the total allocation size
7221 void *__init alloc_large_system_hash(const char *tablename,
7222 unsigned long bucketsize,
7223 unsigned long numentries,
7226 unsigned int *_hash_shift,
7227 unsigned int *_hash_mask,
7228 unsigned long low_limit,
7229 unsigned long high_limit)
7231 unsigned long long max = high_limit;
7232 unsigned long log2qty, size;
7236 /* allow the kernel cmdline to have a say */
7238 /* round applicable memory size up to nearest megabyte */
7239 numentries = nr_kernel_pages;
7240 numentries -= arch_reserved_kernel_pages();
7242 /* It isn't necessary when PAGE_SIZE >= 1MB */
7243 if (PAGE_SHIFT < 20)
7244 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7246 #if __BITS_PER_LONG > 32
7248 unsigned long adapt;
7250 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7251 adapt <<= ADAPT_SCALE_SHIFT)
7256 /* limit to 1 bucket per 2^scale bytes of low memory */
7257 if (scale > PAGE_SHIFT)
7258 numentries >>= (scale - PAGE_SHIFT);
7260 numentries <<= (PAGE_SHIFT - scale);
7262 /* Make sure we've got at least a 0-order allocation.. */
7263 if (unlikely(flags & HASH_SMALL)) {
7264 /* Makes no sense without HASH_EARLY */
7265 WARN_ON(!(flags & HASH_EARLY));
7266 if (!(numentries >> *_hash_shift)) {
7267 numentries = 1UL << *_hash_shift;
7268 BUG_ON(!numentries);
7270 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7271 numentries = PAGE_SIZE / bucketsize;
7273 numentries = roundup_pow_of_two(numentries);
7275 /* limit allocation size to 1/16 total memory by default */
7277 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7278 do_div(max, bucketsize);
7280 max = min(max, 0x80000000ULL);
7282 if (numentries < low_limit)
7283 numentries = low_limit;
7284 if (numentries > max)
7287 log2qty = ilog2(numentries);
7290 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7291 * currently not used when HASH_EARLY is specified.
7293 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7295 size = bucketsize << log2qty;
7296 if (flags & HASH_EARLY)
7297 table = memblock_virt_alloc_nopanic(size, 0);
7299 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7302 * If bucketsize is not a power-of-two, we may free
7303 * some pages at the end of hash table which
7304 * alloc_pages_exact() automatically does
7306 if (get_order(size) < MAX_ORDER) {
7307 table = alloc_pages_exact(size, gfp_flags);
7308 kmemleak_alloc(table, size, 1, gfp_flags);
7311 } while (!table && size > PAGE_SIZE && --log2qty);
7314 panic("Failed to allocate %s hash table\n", tablename);
7316 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7317 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7320 *_hash_shift = log2qty;
7322 *_hash_mask = (1 << log2qty) - 1;
7328 * This function checks whether pageblock includes unmovable pages or not.
7329 * If @count is not zero, it is okay to include less @count unmovable pages
7331 * PageLRU check without isolation or lru_lock could race so that
7332 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7333 * check without lock_page also may miss some movable non-lru pages at
7334 * race condition. So you can't expect this function should be exact.
7336 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7337 bool skip_hwpoisoned_pages)
7339 unsigned long pfn, iter, found;
7343 * For avoiding noise data, lru_add_drain_all() should be called
7344 * If ZONE_MOVABLE, the zone never contains unmovable pages
7346 if (zone_idx(zone) == ZONE_MOVABLE)
7348 mt = get_pageblock_migratetype(page);
7349 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7352 pfn = page_to_pfn(page);
7353 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7354 unsigned long check = pfn + iter;
7356 if (!pfn_valid_within(check))
7359 page = pfn_to_page(check);
7362 * Hugepages are not in LRU lists, but they're movable.
7363 * We need not scan over tail pages bacause we don't
7364 * handle each tail page individually in migration.
7366 if (PageHuge(page)) {
7367 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7372 * We can't use page_count without pin a page
7373 * because another CPU can free compound page.
7374 * This check already skips compound tails of THP
7375 * because their page->_refcount is zero at all time.
7377 if (!page_ref_count(page)) {
7378 if (PageBuddy(page))
7379 iter += (1 << page_order(page)) - 1;
7384 * The HWPoisoned page may be not in buddy system, and
7385 * page_count() is not 0.
7387 if (skip_hwpoisoned_pages && PageHWPoison(page))
7390 if (__PageMovable(page))
7396 * If there are RECLAIMABLE pages, we need to check
7397 * it. But now, memory offline itself doesn't call
7398 * shrink_node_slabs() and it still to be fixed.
7401 * If the page is not RAM, page_count()should be 0.
7402 * we don't need more check. This is an _used_ not-movable page.
7404 * The problematic thing here is PG_reserved pages. PG_reserved
7405 * is set to both of a memory hole page and a _used_ kernel
7414 bool is_pageblock_removable_nolock(struct page *page)
7420 * We have to be careful here because we are iterating over memory
7421 * sections which are not zone aware so we might end up outside of
7422 * the zone but still within the section.
7423 * We have to take care about the node as well. If the node is offline
7424 * its NODE_DATA will be NULL - see page_zone.
7426 if (!node_online(page_to_nid(page)))
7429 zone = page_zone(page);
7430 pfn = page_to_pfn(page);
7431 if (!zone_spans_pfn(zone, pfn))
7434 return !has_unmovable_pages(zone, page, 0, true);
7437 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7439 static unsigned long pfn_max_align_down(unsigned long pfn)
7441 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7442 pageblock_nr_pages) - 1);
7445 static unsigned long pfn_max_align_up(unsigned long pfn)
7447 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7448 pageblock_nr_pages));
7451 /* [start, end) must belong to a single zone. */
7452 static int __alloc_contig_migrate_range(struct compact_control *cc,
7453 unsigned long start, unsigned long end)
7455 /* This function is based on compact_zone() from compaction.c. */
7456 unsigned long nr_reclaimed;
7457 unsigned long pfn = start;
7458 unsigned int tries = 0;
7463 while (pfn < end || !list_empty(&cc->migratepages)) {
7464 if (fatal_signal_pending(current)) {
7469 if (list_empty(&cc->migratepages)) {
7470 cc->nr_migratepages = 0;
7471 pfn = isolate_migratepages_range(cc, pfn, end);
7477 } else if (++tries == 5) {
7478 ret = ret < 0 ? ret : -EBUSY;
7482 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7484 cc->nr_migratepages -= nr_reclaimed;
7486 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7487 NULL, 0, cc->mode, MR_CMA);
7490 putback_movable_pages(&cc->migratepages);
7497 * alloc_contig_range() -- tries to allocate given range of pages
7498 * @start: start PFN to allocate
7499 * @end: one-past-the-last PFN to allocate
7500 * @migratetype: migratetype of the underlaying pageblocks (either
7501 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7502 * in range must have the same migratetype and it must
7503 * be either of the two.
7504 * @gfp_mask: GFP mask to use during compaction
7506 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7507 * aligned, however it's the caller's responsibility to guarantee that
7508 * we are the only thread that changes migrate type of pageblocks the
7511 * The PFN range must belong to a single zone.
7513 * Returns zero on success or negative error code. On success all
7514 * pages which PFN is in [start, end) are allocated for the caller and
7515 * need to be freed with free_contig_range().
7517 int alloc_contig_range(unsigned long start, unsigned long end,
7518 unsigned migratetype, gfp_t gfp_mask)
7520 unsigned long outer_start, outer_end;
7524 struct compact_control cc = {
7525 .nr_migratepages = 0,
7527 .zone = page_zone(pfn_to_page(start)),
7528 .mode = MIGRATE_SYNC,
7529 .ignore_skip_hint = true,
7530 .gfp_mask = current_gfp_context(gfp_mask),
7532 INIT_LIST_HEAD(&cc.migratepages);
7535 * What we do here is we mark all pageblocks in range as
7536 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7537 * have different sizes, and due to the way page allocator
7538 * work, we align the range to biggest of the two pages so
7539 * that page allocator won't try to merge buddies from
7540 * different pageblocks and change MIGRATE_ISOLATE to some
7541 * other migration type.
7543 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7544 * migrate the pages from an unaligned range (ie. pages that
7545 * we are interested in). This will put all the pages in
7546 * range back to page allocator as MIGRATE_ISOLATE.
7548 * When this is done, we take the pages in range from page
7549 * allocator removing them from the buddy system. This way
7550 * page allocator will never consider using them.
7552 * This lets us mark the pageblocks back as
7553 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7554 * aligned range but not in the unaligned, original range are
7555 * put back to page allocator so that buddy can use them.
7558 ret = start_isolate_page_range(pfn_max_align_down(start),
7559 pfn_max_align_up(end), migratetype,
7565 * In case of -EBUSY, we'd like to know which page causes problem.
7566 * So, just fall through. test_pages_isolated() has a tracepoint
7567 * which will report the busy page.
7569 * It is possible that busy pages could become available before
7570 * the call to test_pages_isolated, and the range will actually be
7571 * allocated. So, if we fall through be sure to clear ret so that
7572 * -EBUSY is not accidentally used or returned to caller.
7574 ret = __alloc_contig_migrate_range(&cc, start, end);
7575 if (ret && ret != -EBUSY)
7580 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7581 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7582 * more, all pages in [start, end) are free in page allocator.
7583 * What we are going to do is to allocate all pages from
7584 * [start, end) (that is remove them from page allocator).
7586 * The only problem is that pages at the beginning and at the
7587 * end of interesting range may be not aligned with pages that
7588 * page allocator holds, ie. they can be part of higher order
7589 * pages. Because of this, we reserve the bigger range and
7590 * once this is done free the pages we are not interested in.
7592 * We don't have to hold zone->lock here because the pages are
7593 * isolated thus they won't get removed from buddy.
7596 lru_add_drain_all();
7597 drain_all_pages(cc.zone);
7600 outer_start = start;
7601 while (!PageBuddy(pfn_to_page(outer_start))) {
7602 if (++order >= MAX_ORDER) {
7603 outer_start = start;
7606 outer_start &= ~0UL << order;
7609 if (outer_start != start) {
7610 order = page_order(pfn_to_page(outer_start));
7613 * outer_start page could be small order buddy page and
7614 * it doesn't include start page. Adjust outer_start
7615 * in this case to report failed page properly
7616 * on tracepoint in test_pages_isolated()
7618 if (outer_start + (1UL << order) <= start)
7619 outer_start = start;
7622 /* Make sure the range is really isolated. */
7623 if (test_pages_isolated(outer_start, end, false)) {
7624 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7625 __func__, outer_start, end);
7630 /* Grab isolated pages from freelists. */
7631 outer_end = isolate_freepages_range(&cc, outer_start, end);
7637 /* Free head and tail (if any) */
7638 if (start != outer_start)
7639 free_contig_range(outer_start, start - outer_start);
7640 if (end != outer_end)
7641 free_contig_range(end, outer_end - end);
7644 undo_isolate_page_range(pfn_max_align_down(start),
7645 pfn_max_align_up(end), migratetype);
7649 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7651 unsigned int count = 0;
7653 for (; nr_pages--; pfn++) {
7654 struct page *page = pfn_to_page(pfn);
7656 count += page_count(page) != 1;
7659 WARN(count != 0, "%d pages are still in use!\n", count);
7663 #ifdef CONFIG_MEMORY_HOTPLUG
7665 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7666 * page high values need to be recalulated.
7668 void __meminit zone_pcp_update(struct zone *zone)
7671 mutex_lock(&pcp_batch_high_lock);
7672 for_each_possible_cpu(cpu)
7673 pageset_set_high_and_batch(zone,
7674 per_cpu_ptr(zone->pageset, cpu));
7675 mutex_unlock(&pcp_batch_high_lock);
7679 void zone_pcp_reset(struct zone *zone)
7681 unsigned long flags;
7683 struct per_cpu_pageset *pset;
7685 /* avoid races with drain_pages() */
7686 local_irq_save(flags);
7687 if (zone->pageset != &boot_pageset) {
7688 for_each_online_cpu(cpu) {
7689 pset = per_cpu_ptr(zone->pageset, cpu);
7690 drain_zonestat(zone, pset);
7692 free_percpu(zone->pageset);
7693 zone->pageset = &boot_pageset;
7695 local_irq_restore(flags);
7698 #ifdef CONFIG_MEMORY_HOTREMOVE
7700 * All pages in the range must be in a single zone and isolated
7701 * before calling this.
7704 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7708 unsigned int order, i;
7710 unsigned long flags;
7711 /* find the first valid pfn */
7712 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7717 offline_mem_sections(pfn, end_pfn);
7718 zone = page_zone(pfn_to_page(pfn));
7719 spin_lock_irqsave(&zone->lock, flags);
7721 while (pfn < end_pfn) {
7722 if (!pfn_valid(pfn)) {
7726 page = pfn_to_page(pfn);
7728 * The HWPoisoned page may be not in buddy system, and
7729 * page_count() is not 0.
7731 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7733 SetPageReserved(page);
7737 BUG_ON(page_count(page));
7738 BUG_ON(!PageBuddy(page));
7739 order = page_order(page);
7740 #ifdef CONFIG_DEBUG_VM
7741 pr_info("remove from free list %lx %d %lx\n",
7742 pfn, 1 << order, end_pfn);
7744 list_del(&page->lru);
7745 rmv_page_order(page);
7746 zone->free_area[order].nr_free--;
7747 for (i = 0; i < (1 << order); i++)
7748 SetPageReserved((page+i));
7749 pfn += (1 << order);
7751 spin_unlock_irqrestore(&zone->lock, flags);
7755 bool is_free_buddy_page(struct page *page)
7757 struct zone *zone = page_zone(page);
7758 unsigned long pfn = page_to_pfn(page);
7759 unsigned long flags;
7762 spin_lock_irqsave(&zone->lock, flags);
7763 for (order = 0; order < MAX_ORDER; order++) {
7764 struct page *page_head = page - (pfn & ((1 << order) - 1));
7766 if (PageBuddy(page_head) && page_order(page_head) >= order)
7769 spin_unlock_irqrestore(&zone->lock, flags);
7771 return order < MAX_ORDER;