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>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node);
83 EXPORT_PER_CPU_SYMBOL(numa_node);
86 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
88 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
90 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
91 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
92 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
93 * defined in <linux/topology.h>.
95 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
96 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
97 int _node_numa_mem_[MAX_NUMNODES];
100 /* work_structs for global per-cpu drains */
101 DEFINE_MUTEX(pcpu_drain_mutex);
102 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
104 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
105 volatile unsigned long latent_entropy __latent_entropy;
106 EXPORT_SYMBOL(latent_entropy);
110 * Array of node states.
112 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
113 [N_POSSIBLE] = NODE_MASK_ALL,
114 [N_ONLINE] = { { [0] = 1UL } },
116 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
117 #ifdef CONFIG_HIGHMEM
118 [N_HIGH_MEMORY] = { { [0] = 1UL } },
120 [N_MEMORY] = { { [0] = 1UL } },
121 [N_CPU] = { { [0] = 1UL } },
124 EXPORT_SYMBOL(node_states);
126 /* Protect totalram_pages and zone->managed_pages */
127 static DEFINE_SPINLOCK(managed_page_count_lock);
129 unsigned long totalram_pages __read_mostly;
130 unsigned long totalreserve_pages __read_mostly;
131 unsigned long totalcma_pages __read_mostly;
133 int percpu_pagelist_fraction;
134 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
137 * A cached value of the page's pageblock's migratetype, used when the page is
138 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
139 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
140 * Also the migratetype set in the page does not necessarily match the pcplist
141 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
142 * other index - this ensures that it will be put on the correct CMA freelist.
144 static inline int get_pcppage_migratetype(struct page *page)
149 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
151 page->index = migratetype;
154 #ifdef CONFIG_PM_SLEEP
156 * The following functions are used by the suspend/hibernate code to temporarily
157 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
158 * while devices are suspended. To avoid races with the suspend/hibernate code,
159 * they should always be called with pm_mutex held (gfp_allowed_mask also should
160 * only be modified with pm_mutex held, unless the suspend/hibernate code is
161 * guaranteed not to run in parallel with that modification).
164 static gfp_t saved_gfp_mask;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&pm_mutex));
169 if (saved_gfp_mask) {
170 gfp_allowed_mask = saved_gfp_mask;
175 void pm_restrict_gfp_mask(void)
177 WARN_ON(!mutex_is_locked(&pm_mutex));
178 WARN_ON(saved_gfp_mask);
179 saved_gfp_mask = gfp_allowed_mask;
180 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
183 bool pm_suspended_storage(void)
185 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
189 #endif /* CONFIG_PM_SLEEP */
191 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
192 unsigned int pageblock_order __read_mostly;
195 static void __free_pages_ok(struct page *page, unsigned int order);
198 * results with 256, 32 in the lowmem_reserve sysctl:
199 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
200 * 1G machine -> (16M dma, 784M normal, 224M high)
201 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
202 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
203 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
205 * TBD: should special case ZONE_DMA32 machines here - in those we normally
206 * don't need any ZONE_NORMAL reservation
208 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
215 #ifdef CONFIG_HIGHMEM
221 EXPORT_SYMBOL(totalram_pages);
223 static char * const zone_names[MAX_NR_ZONES] = {
224 #ifdef CONFIG_ZONE_DMA
227 #ifdef CONFIG_ZONE_DMA32
231 #ifdef CONFIG_HIGHMEM
235 #ifdef CONFIG_ZONE_DEVICE
240 char * const migratetype_names[MIGRATE_TYPES] = {
248 #ifdef CONFIG_MEMORY_ISOLATION
253 compound_page_dtor * const compound_page_dtors[] = {
256 #ifdef CONFIG_HUGETLB_PAGE
259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
264 int min_free_kbytes = 1024;
265 int user_min_free_kbytes = -1;
266 int watermark_scale_factor = 10;
268 static unsigned long nr_kernel_pages __meminitdata;
269 static unsigned long nr_all_pages __meminitdata;
270 static unsigned long dma_reserve __meminitdata;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275 static unsigned long required_kernelcore __initdata;
276 static unsigned long required_kernelcore_percent __initdata;
277 static unsigned long required_movablecore __initdata;
278 static unsigned long required_movablecore_percent __initdata;
279 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
280 static bool mirrored_kernelcore __meminitdata;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly = MAX_NUMNODES;
289 int nr_online_nodes __read_mostly = 1;
290 EXPORT_SYMBOL(nr_node_ids);
291 EXPORT_SYMBOL(nr_online_nodes);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297 /* Returns true if the struct page for the pfn is uninitialised */
298 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
300 int nid = early_pfn_to_nid(pfn);
302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
309 * Returns false when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
312 static inline bool update_defer_init(pg_data_t *pgdat,
313 unsigned long pfn, unsigned long zone_end,
314 unsigned long *nr_initialised)
316 /* Always populate low zones for address-constrained allocations */
317 if (zone_end < pgdat_end_pfn(pgdat))
319 /* Xen PV domains need page structures early */
323 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
324 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
325 pgdat->first_deferred_pfn = pfn;
332 static inline bool early_page_uninitialised(unsigned long pfn)
337 static inline bool update_defer_init(pg_data_t *pgdat,
338 unsigned long pfn, unsigned long zone_end,
339 unsigned long *nr_initialised)
345 /* Return a pointer to the bitmap storing bits affecting a block of pages */
346 static inline unsigned long *get_pageblock_bitmap(struct page *page,
349 #ifdef CONFIG_SPARSEMEM
350 return __pfn_to_section(pfn)->pageblock_flags;
352 return page_zone(page)->pageblock_flags;
353 #endif /* CONFIG_SPARSEMEM */
356 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
358 #ifdef CONFIG_SPARSEMEM
359 pfn &= (PAGES_PER_SECTION-1);
360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
363 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
364 #endif /* CONFIG_SPARSEMEM */
368 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
369 * @page: The page within the block of interest
370 * @pfn: The target page frame number
371 * @end_bitidx: The last bit of interest to retrieve
372 * @mask: mask of bits that the caller is interested in
374 * Return: pageblock_bits flags
376 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
378 unsigned long end_bitidx,
381 unsigned long *bitmap;
382 unsigned long bitidx, word_bitidx;
385 bitmap = get_pageblock_bitmap(page, pfn);
386 bitidx = pfn_to_bitidx(page, pfn);
387 word_bitidx = bitidx / BITS_PER_LONG;
388 bitidx &= (BITS_PER_LONG-1);
390 word = bitmap[word_bitidx];
391 bitidx += end_bitidx;
392 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
395 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
396 unsigned long end_bitidx,
399 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
402 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
404 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
408 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
409 * @page: The page within the block of interest
410 * @flags: The flags to set
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest
413 * @mask: mask of bits that the caller is interested in
415 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
417 unsigned long end_bitidx,
420 unsigned long *bitmap;
421 unsigned long bitidx, word_bitidx;
422 unsigned long old_word, word;
424 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
426 bitmap = get_pageblock_bitmap(page, pfn);
427 bitidx = pfn_to_bitidx(page, pfn);
428 word_bitidx = bitidx / BITS_PER_LONG;
429 bitidx &= (BITS_PER_LONG-1);
431 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
433 bitidx += end_bitidx;
434 mask <<= (BITS_PER_LONG - bitidx - 1);
435 flags <<= (BITS_PER_LONG - bitidx - 1);
437 word = READ_ONCE(bitmap[word_bitidx]);
439 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
440 if (word == old_word)
446 void set_pageblock_migratetype(struct page *page, int migratetype)
448 if (unlikely(page_group_by_mobility_disabled &&
449 migratetype < MIGRATE_PCPTYPES))
450 migratetype = MIGRATE_UNMOVABLE;
452 set_pageblock_flags_group(page, (unsigned long)migratetype,
453 PB_migrate, PB_migrate_end);
456 #ifdef CONFIG_DEBUG_VM
457 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
461 unsigned long pfn = page_to_pfn(page);
462 unsigned long sp, start_pfn;
465 seq = zone_span_seqbegin(zone);
466 start_pfn = zone->zone_start_pfn;
467 sp = zone->spanned_pages;
468 if (!zone_spans_pfn(zone, pfn))
470 } while (zone_span_seqretry(zone, seq));
473 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
474 pfn, zone_to_nid(zone), zone->name,
475 start_pfn, start_pfn + sp);
480 static int page_is_consistent(struct zone *zone, struct page *page)
482 if (!pfn_valid_within(page_to_pfn(page)))
484 if (zone != page_zone(page))
490 * Temporary debugging check for pages not lying within a given zone.
492 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
494 if (page_outside_zone_boundaries(zone, page))
496 if (!page_is_consistent(zone, page))
502 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
508 static void bad_page(struct page *page, const char *reason,
509 unsigned long bad_flags)
511 static unsigned long resume;
512 static unsigned long nr_shown;
513 static unsigned long nr_unshown;
516 * Allow a burst of 60 reports, then keep quiet for that minute;
517 * or allow a steady drip of one report per second.
519 if (nr_shown == 60) {
520 if (time_before(jiffies, resume)) {
526 "BUG: Bad page state: %lu messages suppressed\n",
533 resume = jiffies + 60 * HZ;
535 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
536 current->comm, page_to_pfn(page));
537 __dump_page(page, reason);
538 bad_flags &= page->flags;
540 pr_alert("bad because of flags: %#lx(%pGp)\n",
541 bad_flags, &bad_flags);
542 dump_page_owner(page);
547 /* Leave bad fields for debug, except PageBuddy could make trouble */
548 page_mapcount_reset(page); /* remove PageBuddy */
549 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
553 * Higher-order pages are called "compound pages". They are structured thusly:
555 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
557 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
558 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
560 * The first tail page's ->compound_dtor holds the offset in array of compound
561 * page destructors. See compound_page_dtors.
563 * The first tail page's ->compound_order holds the order of allocation.
564 * This usage means that zero-order pages may not be compound.
567 void free_compound_page(struct page *page)
569 __free_pages_ok(page, compound_order(page));
572 void prep_compound_page(struct page *page, unsigned int order)
575 int nr_pages = 1 << order;
577 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
578 set_compound_order(page, order);
580 for (i = 1; i < nr_pages; i++) {
581 struct page *p = page + i;
582 set_page_count(p, 0);
583 p->mapping = TAIL_MAPPING;
584 set_compound_head(p, page);
586 atomic_set(compound_mapcount_ptr(page), -1);
589 #ifdef CONFIG_DEBUG_PAGEALLOC
590 unsigned int _debug_guardpage_minorder;
591 bool _debug_pagealloc_enabled __read_mostly
592 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
593 EXPORT_SYMBOL(_debug_pagealloc_enabled);
594 bool _debug_guardpage_enabled __read_mostly;
596 static int __init early_debug_pagealloc(char *buf)
600 return kstrtobool(buf, &_debug_pagealloc_enabled);
602 early_param("debug_pagealloc", early_debug_pagealloc);
604 static bool need_debug_guardpage(void)
606 /* If we don't use debug_pagealloc, we don't need guard page */
607 if (!debug_pagealloc_enabled())
610 if (!debug_guardpage_minorder())
616 static void init_debug_guardpage(void)
618 if (!debug_pagealloc_enabled())
621 if (!debug_guardpage_minorder())
624 _debug_guardpage_enabled = true;
627 struct page_ext_operations debug_guardpage_ops = {
628 .need = need_debug_guardpage,
629 .init = init_debug_guardpage,
632 static int __init debug_guardpage_minorder_setup(char *buf)
636 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
637 pr_err("Bad debug_guardpage_minorder value\n");
640 _debug_guardpage_minorder = res;
641 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
644 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
646 static inline bool set_page_guard(struct zone *zone, struct page *page,
647 unsigned int order, int migratetype)
649 struct page_ext *page_ext;
651 if (!debug_guardpage_enabled())
654 if (order >= debug_guardpage_minorder())
657 page_ext = lookup_page_ext(page);
658 if (unlikely(!page_ext))
661 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
663 INIT_LIST_HEAD(&page->lru);
664 set_page_private(page, order);
665 /* Guard pages are not available for any usage */
666 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
671 static inline void clear_page_guard(struct zone *zone, struct page *page,
672 unsigned int order, int migratetype)
674 struct page_ext *page_ext;
676 if (!debug_guardpage_enabled())
679 page_ext = lookup_page_ext(page);
680 if (unlikely(!page_ext))
683 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
685 set_page_private(page, 0);
686 if (!is_migrate_isolate(migratetype))
687 __mod_zone_freepage_state(zone, (1 << order), migratetype);
690 struct page_ext_operations debug_guardpage_ops;
691 static inline bool set_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype) { return false; }
693 static inline void clear_page_guard(struct zone *zone, struct page *page,
694 unsigned int order, int migratetype) {}
697 static inline void set_page_order(struct page *page, unsigned int order)
699 set_page_private(page, order);
700 __SetPageBuddy(page);
703 static inline void rmv_page_order(struct page *page)
705 __ClearPageBuddy(page);
706 set_page_private(page, 0);
710 * This function checks whether a page is free && is the buddy
711 * we can do coalesce a page and its buddy if
712 * (a) the buddy is not in a hole (check before calling!) &&
713 * (b) the buddy is in the buddy system &&
714 * (c) a page and its buddy have the same order &&
715 * (d) a page and its buddy are in the same zone.
717 * For recording whether a page is in the buddy system, we set ->_mapcount
718 * PAGE_BUDDY_MAPCOUNT_VALUE.
719 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
720 * serialized by zone->lock.
722 * For recording page's order, we use page_private(page).
724 static inline int page_is_buddy(struct page *page, struct page *buddy,
727 if (page_is_guard(buddy) && page_order(buddy) == order) {
728 if (page_zone_id(page) != page_zone_id(buddy))
731 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
736 if (PageBuddy(buddy) && page_order(buddy) == order) {
738 * zone check is done late to avoid uselessly
739 * calculating zone/node ids for pages that could
742 if (page_zone_id(page) != page_zone_id(buddy))
745 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
753 * Freeing function for a buddy system allocator.
755 * The concept of a buddy system is to maintain direct-mapped table
756 * (containing bit values) for memory blocks of various "orders".
757 * The bottom level table contains the map for the smallest allocatable
758 * units of memory (here, pages), and each level above it describes
759 * pairs of units from the levels below, hence, "buddies".
760 * At a high level, all that happens here is marking the table entry
761 * at the bottom level available, and propagating the changes upward
762 * as necessary, plus some accounting needed to play nicely with other
763 * parts of the VM system.
764 * At each level, we keep a list of pages, which are heads of continuous
765 * free pages of length of (1 << order) and marked with _mapcount
766 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
768 * So when we are allocating or freeing one, we can derive the state of the
769 * other. That is, if we allocate a small block, and both were
770 * free, the remainder of the region must be split into blocks.
771 * If a block is freed, and its buddy is also free, then this
772 * triggers coalescing into a block of larger size.
777 static inline void __free_one_page(struct page *page,
779 struct zone *zone, unsigned int order,
782 unsigned long combined_pfn;
783 unsigned long uninitialized_var(buddy_pfn);
785 unsigned int max_order;
787 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
789 VM_BUG_ON(!zone_is_initialized(zone));
790 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
792 VM_BUG_ON(migratetype == -1);
793 if (likely(!is_migrate_isolate(migratetype)))
794 __mod_zone_freepage_state(zone, 1 << order, migratetype);
796 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
797 VM_BUG_ON_PAGE(bad_range(zone, page), page);
800 while (order < max_order - 1) {
801 buddy_pfn = __find_buddy_pfn(pfn, order);
802 buddy = page + (buddy_pfn - pfn);
804 if (!pfn_valid_within(buddy_pfn))
806 if (!page_is_buddy(page, buddy, order))
809 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
810 * merge with it and move up one order.
812 if (page_is_guard(buddy)) {
813 clear_page_guard(zone, buddy, order, migratetype);
815 list_del(&buddy->lru);
816 zone->free_area[order].nr_free--;
817 rmv_page_order(buddy);
819 combined_pfn = buddy_pfn & pfn;
820 page = page + (combined_pfn - pfn);
824 if (max_order < MAX_ORDER) {
825 /* If we are here, it means order is >= pageblock_order.
826 * We want to prevent merge between freepages on isolate
827 * pageblock and normal pageblock. Without this, pageblock
828 * isolation could cause incorrect freepage or CMA accounting.
830 * We don't want to hit this code for the more frequent
833 if (unlikely(has_isolate_pageblock(zone))) {
836 buddy_pfn = __find_buddy_pfn(pfn, order);
837 buddy = page + (buddy_pfn - pfn);
838 buddy_mt = get_pageblock_migratetype(buddy);
840 if (migratetype != buddy_mt
841 && (is_migrate_isolate(migratetype) ||
842 is_migrate_isolate(buddy_mt)))
846 goto continue_merging;
850 set_page_order(page, order);
853 * If this is not the largest possible page, check if the buddy
854 * of the next-highest order is free. If it is, it's possible
855 * that pages are being freed that will coalesce soon. In case,
856 * that is happening, add the free page to the tail of the list
857 * so it's less likely to be used soon and more likely to be merged
858 * as a higher order page
860 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
861 struct page *higher_page, *higher_buddy;
862 combined_pfn = buddy_pfn & pfn;
863 higher_page = page + (combined_pfn - pfn);
864 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
865 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
866 if (pfn_valid_within(buddy_pfn) &&
867 page_is_buddy(higher_page, higher_buddy, order + 1)) {
868 list_add_tail(&page->lru,
869 &zone->free_area[order].free_list[migratetype]);
874 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
876 zone->free_area[order].nr_free++;
880 * A bad page could be due to a number of fields. Instead of multiple branches,
881 * try and check multiple fields with one check. The caller must do a detailed
882 * check if necessary.
884 static inline bool page_expected_state(struct page *page,
885 unsigned long check_flags)
887 if (unlikely(atomic_read(&page->_mapcount) != -1))
890 if (unlikely((unsigned long)page->mapping |
891 page_ref_count(page) |
893 (unsigned long)page->mem_cgroup |
895 (page->flags & check_flags)))
901 static void free_pages_check_bad(struct page *page)
903 const char *bad_reason;
904 unsigned long bad_flags;
909 if (unlikely(atomic_read(&page->_mapcount) != -1))
910 bad_reason = "nonzero mapcount";
911 if (unlikely(page->mapping != NULL))
912 bad_reason = "non-NULL mapping";
913 if (unlikely(page_ref_count(page) != 0))
914 bad_reason = "nonzero _refcount";
915 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
916 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
917 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
920 if (unlikely(page->mem_cgroup))
921 bad_reason = "page still charged to cgroup";
923 bad_page(page, bad_reason, bad_flags);
926 static inline int free_pages_check(struct page *page)
928 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
931 /* Something has gone sideways, find it */
932 free_pages_check_bad(page);
936 static int free_tail_pages_check(struct page *head_page, struct page *page)
941 * We rely page->lru.next never has bit 0 set, unless the page
942 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
944 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
946 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
950 switch (page - head_page) {
952 /* the first tail page: ->mapping is compound_mapcount() */
953 if (unlikely(compound_mapcount(page))) {
954 bad_page(page, "nonzero compound_mapcount", 0);
960 * the second tail page: ->mapping is
961 * page_deferred_list().next -- ignore value.
965 if (page->mapping != TAIL_MAPPING) {
966 bad_page(page, "corrupted mapping in tail page", 0);
971 if (unlikely(!PageTail(page))) {
972 bad_page(page, "PageTail not set", 0);
975 if (unlikely(compound_head(page) != head_page)) {
976 bad_page(page, "compound_head not consistent", 0);
981 page->mapping = NULL;
982 clear_compound_head(page);
986 static __always_inline bool free_pages_prepare(struct page *page,
987 unsigned int order, bool check_free)
991 VM_BUG_ON_PAGE(PageTail(page), page);
993 trace_mm_page_free(page, order);
996 * Check tail pages before head page information is cleared to
997 * avoid checking PageCompound for order-0 pages.
999 if (unlikely(order)) {
1000 bool compound = PageCompound(page);
1003 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1006 ClearPageDoubleMap(page);
1007 for (i = 1; i < (1 << order); i++) {
1009 bad += free_tail_pages_check(page, page + i);
1010 if (unlikely(free_pages_check(page + i))) {
1014 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1017 if (PageMappingFlags(page))
1018 page->mapping = NULL;
1019 if (memcg_kmem_enabled() && PageKmemcg(page))
1020 memcg_kmem_uncharge(page, order);
1022 bad += free_pages_check(page);
1026 page_cpupid_reset_last(page);
1027 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1028 reset_page_owner(page, order);
1030 if (!PageHighMem(page)) {
1031 debug_check_no_locks_freed(page_address(page),
1032 PAGE_SIZE << order);
1033 debug_check_no_obj_freed(page_address(page),
1034 PAGE_SIZE << order);
1036 arch_free_page(page, order);
1037 kernel_poison_pages(page, 1 << order, 0);
1038 kernel_map_pages(page, 1 << order, 0);
1039 kasan_free_pages(page, order);
1044 #ifdef CONFIG_DEBUG_VM
1045 static inline bool free_pcp_prepare(struct page *page)
1047 return free_pages_prepare(page, 0, true);
1050 static inline bool bulkfree_pcp_prepare(struct page *page)
1055 static bool free_pcp_prepare(struct page *page)
1057 return free_pages_prepare(page, 0, false);
1060 static bool bulkfree_pcp_prepare(struct page *page)
1062 return free_pages_check(page);
1064 #endif /* CONFIG_DEBUG_VM */
1067 * Frees a number of pages from the PCP lists
1068 * Assumes all pages on list are in same zone, and of same order.
1069 * count is the number of pages to free.
1071 * If the zone was previously in an "all pages pinned" state then look to
1072 * see if this freeing clears that state.
1074 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1075 * pinned" detection logic.
1077 static void free_pcppages_bulk(struct zone *zone, int count,
1078 struct per_cpu_pages *pcp)
1080 int migratetype = 0;
1082 bool isolated_pageblocks;
1084 spin_lock(&zone->lock);
1085 isolated_pageblocks = has_isolate_pageblock(zone);
1089 struct list_head *list;
1092 * Remove pages from lists in a round-robin fashion. A
1093 * batch_free count is maintained that is incremented when an
1094 * empty list is encountered. This is so more pages are freed
1095 * off fuller lists instead of spinning excessively around empty
1100 if (++migratetype == MIGRATE_PCPTYPES)
1102 list = &pcp->lists[migratetype];
1103 } while (list_empty(list));
1105 /* This is the only non-empty list. Free them all. */
1106 if (batch_free == MIGRATE_PCPTYPES)
1110 int mt; /* migratetype of the to-be-freed page */
1112 page = list_last_entry(list, struct page, lru);
1113 /* must delete as __free_one_page list manipulates */
1114 list_del(&page->lru);
1117 mt = get_pcppage_migratetype(page);
1118 /* MIGRATE_ISOLATE page should not go to pcplists */
1119 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1120 /* Pageblock could have been isolated meanwhile */
1121 if (unlikely(isolated_pageblocks))
1122 mt = get_pageblock_migratetype(page);
1124 if (bulkfree_pcp_prepare(page))
1127 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1128 trace_mm_page_pcpu_drain(page, 0, mt);
1129 } while (--count && --batch_free && !list_empty(list));
1131 spin_unlock(&zone->lock);
1134 static void free_one_page(struct zone *zone,
1135 struct page *page, unsigned long pfn,
1139 spin_lock(&zone->lock);
1140 if (unlikely(has_isolate_pageblock(zone) ||
1141 is_migrate_isolate(migratetype))) {
1142 migratetype = get_pfnblock_migratetype(page, pfn);
1144 __free_one_page(page, pfn, zone, order, migratetype);
1145 spin_unlock(&zone->lock);
1148 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1149 unsigned long zone, int nid)
1151 mm_zero_struct_page(page);
1152 set_page_links(page, zone, nid, pfn);
1153 init_page_count(page);
1154 page_mapcount_reset(page);
1155 page_cpupid_reset_last(page);
1157 INIT_LIST_HEAD(&page->lru);
1158 #ifdef WANT_PAGE_VIRTUAL
1159 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1160 if (!is_highmem_idx(zone))
1161 set_page_address(page, __va(pfn << PAGE_SHIFT));
1165 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1166 static void __meminit init_reserved_page(unsigned long pfn)
1171 if (!early_page_uninitialised(pfn))
1174 nid = early_pfn_to_nid(pfn);
1175 pgdat = NODE_DATA(nid);
1177 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1178 struct zone *zone = &pgdat->node_zones[zid];
1180 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1183 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1186 static inline void init_reserved_page(unsigned long pfn)
1189 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1192 * Initialised pages do not have PageReserved set. This function is
1193 * called for each range allocated by the bootmem allocator and
1194 * marks the pages PageReserved. The remaining valid pages are later
1195 * sent to the buddy page allocator.
1197 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1199 unsigned long start_pfn = PFN_DOWN(start);
1200 unsigned long end_pfn = PFN_UP(end);
1202 for (; start_pfn < end_pfn; start_pfn++) {
1203 if (pfn_valid(start_pfn)) {
1204 struct page *page = pfn_to_page(start_pfn);
1206 init_reserved_page(start_pfn);
1208 /* Avoid false-positive PageTail() */
1209 INIT_LIST_HEAD(&page->lru);
1211 SetPageReserved(page);
1216 static void __free_pages_ok(struct page *page, unsigned int order)
1218 unsigned long flags;
1220 unsigned long pfn = page_to_pfn(page);
1222 if (!free_pages_prepare(page, order, true))
1225 migratetype = get_pfnblock_migratetype(page, pfn);
1226 local_irq_save(flags);
1227 __count_vm_events(PGFREE, 1 << order);
1228 free_one_page(page_zone(page), page, pfn, order, migratetype);
1229 local_irq_restore(flags);
1232 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1234 unsigned int nr_pages = 1 << order;
1235 struct page *p = page;
1239 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1241 __ClearPageReserved(p);
1242 set_page_count(p, 0);
1244 __ClearPageReserved(p);
1245 set_page_count(p, 0);
1247 page_zone(page)->managed_pages += nr_pages;
1248 set_page_refcounted(page);
1249 __free_pages(page, order);
1252 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1253 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1255 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1257 int __meminit early_pfn_to_nid(unsigned long pfn)
1259 static DEFINE_SPINLOCK(early_pfn_lock);
1262 spin_lock(&early_pfn_lock);
1263 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1265 nid = first_online_node;
1266 spin_unlock(&early_pfn_lock);
1272 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1273 static inline bool __meminit __maybe_unused
1274 meminit_pfn_in_nid(unsigned long pfn, int node,
1275 struct mminit_pfnnid_cache *state)
1279 nid = __early_pfn_to_nid(pfn, state);
1280 if (nid >= 0 && nid != node)
1285 /* Only safe to use early in boot when initialisation is single-threaded */
1286 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1288 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1293 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1297 static inline bool __meminit __maybe_unused
1298 meminit_pfn_in_nid(unsigned long pfn, int node,
1299 struct mminit_pfnnid_cache *state)
1306 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1309 if (early_page_uninitialised(pfn))
1311 return __free_pages_boot_core(page, order);
1315 * Check that the whole (or subset of) a pageblock given by the interval of
1316 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1317 * with the migration of free compaction scanner. The scanners then need to
1318 * use only pfn_valid_within() check for arches that allow holes within
1321 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1323 * It's possible on some configurations to have a setup like node0 node1 node0
1324 * i.e. it's possible that all pages within a zones range of pages do not
1325 * belong to a single zone. We assume that a border between node0 and node1
1326 * can occur within a single pageblock, but not a node0 node1 node0
1327 * interleaving within a single pageblock. It is therefore sufficient to check
1328 * the first and last page of a pageblock and avoid checking each individual
1329 * page in a pageblock.
1331 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1332 unsigned long end_pfn, struct zone *zone)
1334 struct page *start_page;
1335 struct page *end_page;
1337 /* end_pfn is one past the range we are checking */
1340 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1343 start_page = pfn_to_online_page(start_pfn);
1347 if (page_zone(start_page) != zone)
1350 end_page = pfn_to_page(end_pfn);
1352 /* This gives a shorter code than deriving page_zone(end_page) */
1353 if (page_zone_id(start_page) != page_zone_id(end_page))
1359 void set_zone_contiguous(struct zone *zone)
1361 unsigned long block_start_pfn = zone->zone_start_pfn;
1362 unsigned long block_end_pfn;
1364 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1365 for (; block_start_pfn < zone_end_pfn(zone);
1366 block_start_pfn = block_end_pfn,
1367 block_end_pfn += pageblock_nr_pages) {
1369 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1371 if (!__pageblock_pfn_to_page(block_start_pfn,
1372 block_end_pfn, zone))
1376 /* We confirm that there is no hole */
1377 zone->contiguous = true;
1380 void clear_zone_contiguous(struct zone *zone)
1382 zone->contiguous = false;
1385 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1386 static void __init deferred_free_range(unsigned long pfn,
1387 unsigned long nr_pages)
1395 page = pfn_to_page(pfn);
1397 /* Free a large naturally-aligned chunk if possible */
1398 if (nr_pages == pageblock_nr_pages &&
1399 (pfn & (pageblock_nr_pages - 1)) == 0) {
1400 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1401 __free_pages_boot_core(page, pageblock_order);
1405 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1406 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1407 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1408 __free_pages_boot_core(page, 0);
1412 /* Completion tracking for deferred_init_memmap() threads */
1413 static atomic_t pgdat_init_n_undone __initdata;
1414 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1416 static inline void __init pgdat_init_report_one_done(void)
1418 if (atomic_dec_and_test(&pgdat_init_n_undone))
1419 complete(&pgdat_init_all_done_comp);
1423 * Returns true if page needs to be initialized or freed to buddy allocator.
1425 * First we check if pfn is valid on architectures where it is possible to have
1426 * holes within pageblock_nr_pages. On systems where it is not possible, this
1427 * function is optimized out.
1429 * Then, we check if a current large page is valid by only checking the validity
1432 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1433 * within a node: a pfn is between start and end of a node, but does not belong
1434 * to this memory node.
1436 static inline bool __init
1437 deferred_pfn_valid(int nid, unsigned long pfn,
1438 struct mminit_pfnnid_cache *nid_init_state)
1440 if (!pfn_valid_within(pfn))
1442 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1444 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1450 * Free pages to buddy allocator. Try to free aligned pages in
1451 * pageblock_nr_pages sizes.
1453 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1454 unsigned long end_pfn)
1456 struct mminit_pfnnid_cache nid_init_state = { };
1457 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1458 unsigned long nr_free = 0;
1460 for (; pfn < end_pfn; pfn++) {
1461 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1462 deferred_free_range(pfn - nr_free, nr_free);
1464 } else if (!(pfn & nr_pgmask)) {
1465 deferred_free_range(pfn - nr_free, nr_free);
1467 touch_nmi_watchdog();
1472 /* Free the last block of pages to allocator */
1473 deferred_free_range(pfn - nr_free, nr_free);
1477 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1478 * by performing it only once every pageblock_nr_pages.
1479 * Return number of pages initialized.
1481 static unsigned long __init deferred_init_pages(int nid, int zid,
1483 unsigned long end_pfn)
1485 struct mminit_pfnnid_cache nid_init_state = { };
1486 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1487 unsigned long nr_pages = 0;
1488 struct page *page = NULL;
1490 for (; pfn < end_pfn; pfn++) {
1491 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1494 } else if (!page || !(pfn & nr_pgmask)) {
1495 page = pfn_to_page(pfn);
1496 touch_nmi_watchdog();
1500 __init_single_page(page, pfn, zid, nid);
1506 /* Initialise remaining memory on a node */
1507 static int __init deferred_init_memmap(void *data)
1509 pg_data_t *pgdat = data;
1510 int nid = pgdat->node_id;
1511 unsigned long start = jiffies;
1512 unsigned long nr_pages = 0;
1513 unsigned long spfn, epfn, first_init_pfn, flags;
1514 phys_addr_t spa, epa;
1517 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1520 /* Bind memory initialisation thread to a local node if possible */
1521 if (!cpumask_empty(cpumask))
1522 set_cpus_allowed_ptr(current, cpumask);
1524 pgdat_resize_lock(pgdat, &flags);
1525 first_init_pfn = pgdat->first_deferred_pfn;
1526 if (first_init_pfn == ULONG_MAX) {
1527 pgdat_resize_unlock(pgdat, &flags);
1528 pgdat_init_report_one_done();
1532 /* Sanity check boundaries */
1533 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1534 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1535 pgdat->first_deferred_pfn = ULONG_MAX;
1537 /* Only the highest zone is deferred so find it */
1538 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1539 zone = pgdat->node_zones + zid;
1540 if (first_init_pfn < zone_end_pfn(zone))
1543 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1546 * Initialize and free pages. We do it in two loops: first we initialize
1547 * struct page, than free to buddy allocator, because while we are
1548 * freeing pages we can access pages that are ahead (computing buddy
1549 * page in __free_one_page()).
1551 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1552 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1553 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1554 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1556 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1557 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1558 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1559 deferred_free_pages(nid, zid, spfn, epfn);
1561 pgdat_resize_unlock(pgdat, &flags);
1563 /* Sanity check that the next zone really is unpopulated */
1564 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1566 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1567 jiffies_to_msecs(jiffies - start));
1569 pgdat_init_report_one_done();
1574 * During boot we initialize deferred pages on-demand, as needed, but once
1575 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1576 * and we can permanently disable that path.
1578 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1581 * If this zone has deferred pages, try to grow it by initializing enough
1582 * deferred pages to satisfy the allocation specified by order, rounded up to
1583 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1584 * of SECTION_SIZE bytes by initializing struct pages in increments of
1585 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1587 * Return true when zone was grown, otherwise return false. We return true even
1588 * when we grow less than requested, to let the caller decide if there are
1589 * enough pages to satisfy the allocation.
1591 * Note: We use noinline because this function is needed only during boot, and
1592 * it is called from a __ref function _deferred_grow_zone. This way we are
1593 * making sure that it is not inlined into permanent text section.
1595 static noinline bool __init
1596 deferred_grow_zone(struct zone *zone, unsigned int order)
1598 int zid = zone_idx(zone);
1599 int nid = zone_to_nid(zone);
1600 pg_data_t *pgdat = NODE_DATA(nid);
1601 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1602 unsigned long nr_pages = 0;
1603 unsigned long first_init_pfn, spfn, epfn, t, flags;
1604 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1605 phys_addr_t spa, epa;
1608 /* Only the last zone may have deferred pages */
1609 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1612 pgdat_resize_lock(pgdat, &flags);
1615 * If deferred pages have been initialized while we were waiting for
1616 * the lock, return true, as the zone was grown. The caller will retry
1617 * this zone. We won't return to this function since the caller also
1618 * has this static branch.
1620 if (!static_branch_unlikely(&deferred_pages)) {
1621 pgdat_resize_unlock(pgdat, &flags);
1626 * If someone grew this zone while we were waiting for spinlock, return
1627 * true, as there might be enough pages already.
1629 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1630 pgdat_resize_unlock(pgdat, &flags);
1634 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1636 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1637 pgdat_resize_unlock(pgdat, &flags);
1641 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1642 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1643 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1645 while (spfn < epfn && nr_pages < nr_pages_needed) {
1646 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1647 first_deferred_pfn = min(t, epfn);
1648 nr_pages += deferred_init_pages(nid, zid, spfn,
1649 first_deferred_pfn);
1650 spfn = first_deferred_pfn;
1653 if (nr_pages >= nr_pages_needed)
1657 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1658 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1659 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1660 deferred_free_pages(nid, zid, spfn, epfn);
1662 if (first_deferred_pfn == epfn)
1665 pgdat->first_deferred_pfn = first_deferred_pfn;
1666 pgdat_resize_unlock(pgdat, &flags);
1668 return nr_pages > 0;
1672 * deferred_grow_zone() is __init, but it is called from
1673 * get_page_from_freelist() during early boot until deferred_pages permanently
1674 * disables this call. This is why we have refdata wrapper to avoid warning,
1675 * and to ensure that the function body gets unloaded.
1678 _deferred_grow_zone(struct zone *zone, unsigned int order)
1680 return deferred_grow_zone(zone, order);
1683 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1685 void __init page_alloc_init_late(void)
1689 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1692 /* There will be num_node_state(N_MEMORY) threads */
1693 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1694 for_each_node_state(nid, N_MEMORY) {
1695 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1698 /* Block until all are initialised */
1699 wait_for_completion(&pgdat_init_all_done_comp);
1702 * We initialized the rest of the deferred pages. Permanently disable
1703 * on-demand struct page initialization.
1705 static_branch_disable(&deferred_pages);
1707 /* Reinit limits that are based on free pages after the kernel is up */
1708 files_maxfiles_init();
1710 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1711 /* Discard memblock private memory */
1715 for_each_populated_zone(zone)
1716 set_zone_contiguous(zone);
1720 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1721 void __init init_cma_reserved_pageblock(struct page *page)
1723 unsigned i = pageblock_nr_pages;
1724 struct page *p = page;
1727 __ClearPageReserved(p);
1728 set_page_count(p, 0);
1731 set_pageblock_migratetype(page, MIGRATE_CMA);
1733 if (pageblock_order >= MAX_ORDER) {
1734 i = pageblock_nr_pages;
1737 set_page_refcounted(p);
1738 __free_pages(p, MAX_ORDER - 1);
1739 p += MAX_ORDER_NR_PAGES;
1740 } while (i -= MAX_ORDER_NR_PAGES);
1742 set_page_refcounted(page);
1743 __free_pages(page, pageblock_order);
1746 adjust_managed_page_count(page, pageblock_nr_pages);
1751 * The order of subdivision here is critical for the IO subsystem.
1752 * Please do not alter this order without good reasons and regression
1753 * testing. Specifically, as large blocks of memory are subdivided,
1754 * the order in which smaller blocks are delivered depends on the order
1755 * they're subdivided in this function. This is the primary factor
1756 * influencing the order in which pages are delivered to the IO
1757 * subsystem according to empirical testing, and this is also justified
1758 * by considering the behavior of a buddy system containing a single
1759 * large block of memory acted on by a series of small allocations.
1760 * This behavior is a critical factor in sglist merging's success.
1764 static inline void expand(struct zone *zone, struct page *page,
1765 int low, int high, struct free_area *area,
1768 unsigned long size = 1 << high;
1770 while (high > low) {
1774 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1777 * Mark as guard pages (or page), that will allow to
1778 * merge back to allocator when buddy will be freed.
1779 * Corresponding page table entries will not be touched,
1780 * pages will stay not present in virtual address space
1782 if (set_page_guard(zone, &page[size], high, migratetype))
1785 list_add(&page[size].lru, &area->free_list[migratetype]);
1787 set_page_order(&page[size], high);
1791 static void check_new_page_bad(struct page *page)
1793 const char *bad_reason = NULL;
1794 unsigned long bad_flags = 0;
1796 if (unlikely(atomic_read(&page->_mapcount) != -1))
1797 bad_reason = "nonzero mapcount";
1798 if (unlikely(page->mapping != NULL))
1799 bad_reason = "non-NULL mapping";
1800 if (unlikely(page_ref_count(page) != 0))
1801 bad_reason = "nonzero _count";
1802 if (unlikely(page->flags & __PG_HWPOISON)) {
1803 bad_reason = "HWPoisoned (hardware-corrupted)";
1804 bad_flags = __PG_HWPOISON;
1805 /* Don't complain about hwpoisoned pages */
1806 page_mapcount_reset(page); /* remove PageBuddy */
1809 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1810 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1811 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1814 if (unlikely(page->mem_cgroup))
1815 bad_reason = "page still charged to cgroup";
1817 bad_page(page, bad_reason, bad_flags);
1821 * This page is about to be returned from the page allocator
1823 static inline int check_new_page(struct page *page)
1825 if (likely(page_expected_state(page,
1826 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1829 check_new_page_bad(page);
1833 static inline bool free_pages_prezeroed(void)
1835 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1836 page_poisoning_enabled();
1839 #ifdef CONFIG_DEBUG_VM
1840 static bool check_pcp_refill(struct page *page)
1845 static bool check_new_pcp(struct page *page)
1847 return check_new_page(page);
1850 static bool check_pcp_refill(struct page *page)
1852 return check_new_page(page);
1854 static bool check_new_pcp(struct page *page)
1858 #endif /* CONFIG_DEBUG_VM */
1860 static bool check_new_pages(struct page *page, unsigned int order)
1863 for (i = 0; i < (1 << order); i++) {
1864 struct page *p = page + i;
1866 if (unlikely(check_new_page(p)))
1873 inline void post_alloc_hook(struct page *page, unsigned int order,
1876 set_page_private(page, 0);
1877 set_page_refcounted(page);
1879 arch_alloc_page(page, order);
1880 kernel_map_pages(page, 1 << order, 1);
1881 kernel_poison_pages(page, 1 << order, 1);
1882 kasan_alloc_pages(page, order);
1883 set_page_owner(page, order, gfp_flags);
1886 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1887 unsigned int alloc_flags)
1891 post_alloc_hook(page, order, gfp_flags);
1893 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1894 for (i = 0; i < (1 << order); i++)
1895 clear_highpage(page + i);
1897 if (order && (gfp_flags & __GFP_COMP))
1898 prep_compound_page(page, order);
1901 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1902 * allocate the page. The expectation is that the caller is taking
1903 * steps that will free more memory. The caller should avoid the page
1904 * being used for !PFMEMALLOC purposes.
1906 if (alloc_flags & ALLOC_NO_WATERMARKS)
1907 set_page_pfmemalloc(page);
1909 clear_page_pfmemalloc(page);
1913 * Go through the free lists for the given migratetype and remove
1914 * the smallest available page from the freelists
1916 static __always_inline
1917 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1920 unsigned int current_order;
1921 struct free_area *area;
1924 /* Find a page of the appropriate size in the preferred list */
1925 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1926 area = &(zone->free_area[current_order]);
1927 page = list_first_entry_or_null(&area->free_list[migratetype],
1931 list_del(&page->lru);
1932 rmv_page_order(page);
1934 expand(zone, page, order, current_order, area, migratetype);
1935 set_pcppage_migratetype(page, migratetype);
1944 * This array describes the order lists are fallen back to when
1945 * the free lists for the desirable migrate type are depleted
1947 static int fallbacks[MIGRATE_TYPES][4] = {
1948 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1949 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1950 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1952 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1954 #ifdef CONFIG_MEMORY_ISOLATION
1955 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1960 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1963 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1966 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1967 unsigned int order) { return NULL; }
1971 * Move the free pages in a range to the free lists of the requested type.
1972 * Note that start_page and end_pages are not aligned on a pageblock
1973 * boundary. If alignment is required, use move_freepages_block()
1975 static int move_freepages(struct zone *zone,
1976 struct page *start_page, struct page *end_page,
1977 int migratetype, int *num_movable)
1981 int pages_moved = 0;
1983 #ifndef CONFIG_HOLES_IN_ZONE
1985 * page_zone is not safe to call in this context when
1986 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1987 * anyway as we check zone boundaries in move_freepages_block().
1988 * Remove at a later date when no bug reports exist related to
1989 * grouping pages by mobility
1991 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
1992 pfn_valid(page_to_pfn(end_page)) &&
1993 page_zone(start_page) != page_zone(end_page));
1999 for (page = start_page; page <= end_page;) {
2000 if (!pfn_valid_within(page_to_pfn(page))) {
2005 /* Make sure we are not inadvertently changing nodes */
2006 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2008 if (!PageBuddy(page)) {
2010 * We assume that pages that could be isolated for
2011 * migration are movable. But we don't actually try
2012 * isolating, as that would be expensive.
2015 (PageLRU(page) || __PageMovable(page)))
2022 order = page_order(page);
2023 list_move(&page->lru,
2024 &zone->free_area[order].free_list[migratetype]);
2026 pages_moved += 1 << order;
2032 int move_freepages_block(struct zone *zone, struct page *page,
2033 int migratetype, int *num_movable)
2035 unsigned long start_pfn, end_pfn;
2036 struct page *start_page, *end_page;
2038 start_pfn = page_to_pfn(page);
2039 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2040 start_page = pfn_to_page(start_pfn);
2041 end_page = start_page + pageblock_nr_pages - 1;
2042 end_pfn = start_pfn + pageblock_nr_pages - 1;
2044 /* Do not cross zone boundaries */
2045 if (!zone_spans_pfn(zone, start_pfn))
2047 if (!zone_spans_pfn(zone, end_pfn))
2050 return move_freepages(zone, start_page, end_page, migratetype,
2054 static void change_pageblock_range(struct page *pageblock_page,
2055 int start_order, int migratetype)
2057 int nr_pageblocks = 1 << (start_order - pageblock_order);
2059 while (nr_pageblocks--) {
2060 set_pageblock_migratetype(pageblock_page, migratetype);
2061 pageblock_page += pageblock_nr_pages;
2066 * When we are falling back to another migratetype during allocation, try to
2067 * steal extra free pages from the same pageblocks to satisfy further
2068 * allocations, instead of polluting multiple pageblocks.
2070 * If we are stealing a relatively large buddy page, it is likely there will
2071 * be more free pages in the pageblock, so try to steal them all. For
2072 * reclaimable and unmovable allocations, we steal regardless of page size,
2073 * as fragmentation caused by those allocations polluting movable pageblocks
2074 * is worse than movable allocations stealing from unmovable and reclaimable
2077 static bool can_steal_fallback(unsigned int order, int start_mt)
2080 * Leaving this order check is intended, although there is
2081 * relaxed order check in next check. The reason is that
2082 * we can actually steal whole pageblock if this condition met,
2083 * but, below check doesn't guarantee it and that is just heuristic
2084 * so could be changed anytime.
2086 if (order >= pageblock_order)
2089 if (order >= pageblock_order / 2 ||
2090 start_mt == MIGRATE_RECLAIMABLE ||
2091 start_mt == MIGRATE_UNMOVABLE ||
2092 page_group_by_mobility_disabled)
2099 * This function implements actual steal behaviour. If order is large enough,
2100 * we can steal whole pageblock. If not, we first move freepages in this
2101 * pageblock to our migratetype and determine how many already-allocated pages
2102 * are there in the pageblock with a compatible migratetype. If at least half
2103 * of pages are free or compatible, we can change migratetype of the pageblock
2104 * itself, so pages freed in the future will be put on the correct free list.
2106 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2107 int start_type, bool whole_block)
2109 unsigned int current_order = page_order(page);
2110 struct free_area *area;
2111 int free_pages, movable_pages, alike_pages;
2114 old_block_type = get_pageblock_migratetype(page);
2117 * This can happen due to races and we want to prevent broken
2118 * highatomic accounting.
2120 if (is_migrate_highatomic(old_block_type))
2123 /* Take ownership for orders >= pageblock_order */
2124 if (current_order >= pageblock_order) {
2125 change_pageblock_range(page, current_order, start_type);
2129 /* We are not allowed to try stealing from the whole block */
2133 free_pages = move_freepages_block(zone, page, start_type,
2136 * Determine how many pages are compatible with our allocation.
2137 * For movable allocation, it's the number of movable pages which
2138 * we just obtained. For other types it's a bit more tricky.
2140 if (start_type == MIGRATE_MOVABLE) {
2141 alike_pages = movable_pages;
2144 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2145 * to MOVABLE pageblock, consider all non-movable pages as
2146 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2147 * vice versa, be conservative since we can't distinguish the
2148 * exact migratetype of non-movable pages.
2150 if (old_block_type == MIGRATE_MOVABLE)
2151 alike_pages = pageblock_nr_pages
2152 - (free_pages + movable_pages);
2157 /* moving whole block can fail due to zone boundary conditions */
2162 * If a sufficient number of pages in the block are either free or of
2163 * comparable migratability as our allocation, claim the whole block.
2165 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2166 page_group_by_mobility_disabled)
2167 set_pageblock_migratetype(page, start_type);
2172 area = &zone->free_area[current_order];
2173 list_move(&page->lru, &area->free_list[start_type]);
2177 * Check whether there is a suitable fallback freepage with requested order.
2178 * If only_stealable is true, this function returns fallback_mt only if
2179 * we can steal other freepages all together. This would help to reduce
2180 * fragmentation due to mixed migratetype pages in one pageblock.
2182 int find_suitable_fallback(struct free_area *area, unsigned int order,
2183 int migratetype, bool only_stealable, bool *can_steal)
2188 if (area->nr_free == 0)
2193 fallback_mt = fallbacks[migratetype][i];
2194 if (fallback_mt == MIGRATE_TYPES)
2197 if (list_empty(&area->free_list[fallback_mt]))
2200 if (can_steal_fallback(order, migratetype))
2203 if (!only_stealable)
2214 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2215 * there are no empty page blocks that contain a page with a suitable order
2217 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2218 unsigned int alloc_order)
2221 unsigned long max_managed, flags;
2224 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2225 * Check is race-prone but harmless.
2227 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2228 if (zone->nr_reserved_highatomic >= max_managed)
2231 spin_lock_irqsave(&zone->lock, flags);
2233 /* Recheck the nr_reserved_highatomic limit under the lock */
2234 if (zone->nr_reserved_highatomic >= max_managed)
2238 mt = get_pageblock_migratetype(page);
2239 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2240 && !is_migrate_cma(mt)) {
2241 zone->nr_reserved_highatomic += pageblock_nr_pages;
2242 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2243 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2247 spin_unlock_irqrestore(&zone->lock, flags);
2251 * Used when an allocation is about to fail under memory pressure. This
2252 * potentially hurts the reliability of high-order allocations when under
2253 * intense memory pressure but failed atomic allocations should be easier
2254 * to recover from than an OOM.
2256 * If @force is true, try to unreserve a pageblock even though highatomic
2257 * pageblock is exhausted.
2259 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2262 struct zonelist *zonelist = ac->zonelist;
2263 unsigned long flags;
2270 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2273 * Preserve at least one pageblock unless memory pressure
2276 if (!force && zone->nr_reserved_highatomic <=
2280 spin_lock_irqsave(&zone->lock, flags);
2281 for (order = 0; order < MAX_ORDER; order++) {
2282 struct free_area *area = &(zone->free_area[order]);
2284 page = list_first_entry_or_null(
2285 &area->free_list[MIGRATE_HIGHATOMIC],
2291 * In page freeing path, migratetype change is racy so
2292 * we can counter several free pages in a pageblock
2293 * in this loop althoug we changed the pageblock type
2294 * from highatomic to ac->migratetype. So we should
2295 * adjust the count once.
2297 if (is_migrate_highatomic_page(page)) {
2299 * It should never happen but changes to
2300 * locking could inadvertently allow a per-cpu
2301 * drain to add pages to MIGRATE_HIGHATOMIC
2302 * while unreserving so be safe and watch for
2305 zone->nr_reserved_highatomic -= min(
2307 zone->nr_reserved_highatomic);
2311 * Convert to ac->migratetype and avoid the normal
2312 * pageblock stealing heuristics. Minimally, the caller
2313 * is doing the work and needs the pages. More
2314 * importantly, if the block was always converted to
2315 * MIGRATE_UNMOVABLE or another type then the number
2316 * of pageblocks that cannot be completely freed
2319 set_pageblock_migratetype(page, ac->migratetype);
2320 ret = move_freepages_block(zone, page, ac->migratetype,
2323 spin_unlock_irqrestore(&zone->lock, flags);
2327 spin_unlock_irqrestore(&zone->lock, flags);
2334 * Try finding a free buddy page on the fallback list and put it on the free
2335 * list of requested migratetype, possibly along with other pages from the same
2336 * block, depending on fragmentation avoidance heuristics. Returns true if
2337 * fallback was found so that __rmqueue_smallest() can grab it.
2339 * The use of signed ints for order and current_order is a deliberate
2340 * deviation from the rest of this file, to make the for loop
2341 * condition simpler.
2343 static __always_inline bool
2344 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2346 struct free_area *area;
2353 * Find the largest available free page in the other list. This roughly
2354 * approximates finding the pageblock with the most free pages, which
2355 * would be too costly to do exactly.
2357 for (current_order = MAX_ORDER - 1; current_order >= order;
2359 area = &(zone->free_area[current_order]);
2360 fallback_mt = find_suitable_fallback(area, current_order,
2361 start_migratetype, false, &can_steal);
2362 if (fallback_mt == -1)
2366 * We cannot steal all free pages from the pageblock and the
2367 * requested migratetype is movable. In that case it's better to
2368 * steal and split the smallest available page instead of the
2369 * largest available page, because even if the next movable
2370 * allocation falls back into a different pageblock than this
2371 * one, it won't cause permanent fragmentation.
2373 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2374 && current_order > order)
2383 for (current_order = order; current_order < MAX_ORDER;
2385 area = &(zone->free_area[current_order]);
2386 fallback_mt = find_suitable_fallback(area, current_order,
2387 start_migratetype, false, &can_steal);
2388 if (fallback_mt != -1)
2393 * This should not happen - we already found a suitable fallback
2394 * when looking for the largest page.
2396 VM_BUG_ON(current_order == MAX_ORDER);
2399 page = list_first_entry(&area->free_list[fallback_mt],
2402 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2404 trace_mm_page_alloc_extfrag(page, order, current_order,
2405 start_migratetype, fallback_mt);
2412 * Do the hard work of removing an element from the buddy allocator.
2413 * Call me with the zone->lock already held.
2415 static __always_inline struct page *
2416 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2421 page = __rmqueue_smallest(zone, order, migratetype);
2422 if (unlikely(!page)) {
2423 if (migratetype == MIGRATE_MOVABLE)
2424 page = __rmqueue_cma_fallback(zone, order);
2426 if (!page && __rmqueue_fallback(zone, order, migratetype))
2430 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2435 * Obtain a specified number of elements from the buddy allocator, all under
2436 * a single hold of the lock, for efficiency. Add them to the supplied list.
2437 * Returns the number of new pages which were placed at *list.
2439 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2440 unsigned long count, struct list_head *list,
2445 spin_lock(&zone->lock);
2446 for (i = 0; i < count; ++i) {
2447 struct page *page = __rmqueue(zone, order, migratetype);
2448 if (unlikely(page == NULL))
2451 if (unlikely(check_pcp_refill(page)))
2455 * Split buddy pages returned by expand() are received here in
2456 * physical page order. The page is added to the tail of
2457 * caller's list. From the callers perspective, the linked list
2458 * is ordered by page number under some conditions. This is
2459 * useful for IO devices that can forward direction from the
2460 * head, thus also in the physical page order. This is useful
2461 * for IO devices that can merge IO requests if the physical
2462 * pages are ordered properly.
2464 list_add_tail(&page->lru, list);
2466 if (is_migrate_cma(get_pcppage_migratetype(page)))
2467 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2472 * i pages were removed from the buddy list even if some leak due
2473 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2474 * on i. Do not confuse with 'alloced' which is the number of
2475 * pages added to the pcp list.
2477 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2478 spin_unlock(&zone->lock);
2484 * Called from the vmstat counter updater to drain pagesets of this
2485 * currently executing processor on remote nodes after they have
2488 * Note that this function must be called with the thread pinned to
2489 * a single processor.
2491 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2493 unsigned long flags;
2494 int to_drain, batch;
2496 local_irq_save(flags);
2497 batch = READ_ONCE(pcp->batch);
2498 to_drain = min(pcp->count, batch);
2500 free_pcppages_bulk(zone, to_drain, pcp);
2501 local_irq_restore(flags);
2506 * Drain pcplists of the indicated processor and zone.
2508 * The processor must either be the current processor and the
2509 * thread pinned to the current processor or a processor that
2512 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2514 unsigned long flags;
2515 struct per_cpu_pageset *pset;
2516 struct per_cpu_pages *pcp;
2518 local_irq_save(flags);
2519 pset = per_cpu_ptr(zone->pageset, cpu);
2523 free_pcppages_bulk(zone, pcp->count, pcp);
2524 local_irq_restore(flags);
2528 * Drain pcplists of all zones on the indicated processor.
2530 * The processor must either be the current processor and the
2531 * thread pinned to the current processor or a processor that
2534 static void drain_pages(unsigned int cpu)
2538 for_each_populated_zone(zone) {
2539 drain_pages_zone(cpu, zone);
2544 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2546 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2547 * the single zone's pages.
2549 void drain_local_pages(struct zone *zone)
2551 int cpu = smp_processor_id();
2554 drain_pages_zone(cpu, zone);
2559 static void drain_local_pages_wq(struct work_struct *work)
2562 * drain_all_pages doesn't use proper cpu hotplug protection so
2563 * we can race with cpu offline when the WQ can move this from
2564 * a cpu pinned worker to an unbound one. We can operate on a different
2565 * cpu which is allright but we also have to make sure to not move to
2569 drain_local_pages(NULL);
2574 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2576 * When zone parameter is non-NULL, spill just the single zone's pages.
2578 * Note that this can be extremely slow as the draining happens in a workqueue.
2580 void drain_all_pages(struct zone *zone)
2585 * Allocate in the BSS so we wont require allocation in
2586 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2588 static cpumask_t cpus_with_pcps;
2591 * Make sure nobody triggers this path before mm_percpu_wq is fully
2594 if (WARN_ON_ONCE(!mm_percpu_wq))
2598 * Do not drain if one is already in progress unless it's specific to
2599 * a zone. Such callers are primarily CMA and memory hotplug and need
2600 * the drain to be complete when the call returns.
2602 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2605 mutex_lock(&pcpu_drain_mutex);
2609 * We don't care about racing with CPU hotplug event
2610 * as offline notification will cause the notified
2611 * cpu to drain that CPU pcps and on_each_cpu_mask
2612 * disables preemption as part of its processing
2614 for_each_online_cpu(cpu) {
2615 struct per_cpu_pageset *pcp;
2617 bool has_pcps = false;
2620 pcp = per_cpu_ptr(zone->pageset, cpu);
2624 for_each_populated_zone(z) {
2625 pcp = per_cpu_ptr(z->pageset, cpu);
2626 if (pcp->pcp.count) {
2634 cpumask_set_cpu(cpu, &cpus_with_pcps);
2636 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2639 for_each_cpu(cpu, &cpus_with_pcps) {
2640 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2641 INIT_WORK(work, drain_local_pages_wq);
2642 queue_work_on(cpu, mm_percpu_wq, work);
2644 for_each_cpu(cpu, &cpus_with_pcps)
2645 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2647 mutex_unlock(&pcpu_drain_mutex);
2650 #ifdef CONFIG_HIBERNATION
2653 * Touch the watchdog for every WD_PAGE_COUNT pages.
2655 #define WD_PAGE_COUNT (128*1024)
2657 void mark_free_pages(struct zone *zone)
2659 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2660 unsigned long flags;
2661 unsigned int order, t;
2664 if (zone_is_empty(zone))
2667 spin_lock_irqsave(&zone->lock, flags);
2669 max_zone_pfn = zone_end_pfn(zone);
2670 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2671 if (pfn_valid(pfn)) {
2672 page = pfn_to_page(pfn);
2674 if (!--page_count) {
2675 touch_nmi_watchdog();
2676 page_count = WD_PAGE_COUNT;
2679 if (page_zone(page) != zone)
2682 if (!swsusp_page_is_forbidden(page))
2683 swsusp_unset_page_free(page);
2686 for_each_migratetype_order(order, t) {
2687 list_for_each_entry(page,
2688 &zone->free_area[order].free_list[t], lru) {
2691 pfn = page_to_pfn(page);
2692 for (i = 0; i < (1UL << order); i++) {
2693 if (!--page_count) {
2694 touch_nmi_watchdog();
2695 page_count = WD_PAGE_COUNT;
2697 swsusp_set_page_free(pfn_to_page(pfn + i));
2701 spin_unlock_irqrestore(&zone->lock, flags);
2703 #endif /* CONFIG_PM */
2705 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2709 if (!free_pcp_prepare(page))
2712 migratetype = get_pfnblock_migratetype(page, pfn);
2713 set_pcppage_migratetype(page, migratetype);
2717 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2719 struct zone *zone = page_zone(page);
2720 struct per_cpu_pages *pcp;
2723 migratetype = get_pcppage_migratetype(page);
2724 __count_vm_event(PGFREE);
2727 * We only track unmovable, reclaimable and movable on pcp lists.
2728 * Free ISOLATE pages back to the allocator because they are being
2729 * offlined but treat HIGHATOMIC as movable pages so we can get those
2730 * areas back if necessary. Otherwise, we may have to free
2731 * excessively into the page allocator
2733 if (migratetype >= MIGRATE_PCPTYPES) {
2734 if (unlikely(is_migrate_isolate(migratetype))) {
2735 free_one_page(zone, page, pfn, 0, migratetype);
2738 migratetype = MIGRATE_MOVABLE;
2741 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2742 list_add(&page->lru, &pcp->lists[migratetype]);
2744 if (pcp->count >= pcp->high) {
2745 unsigned long batch = READ_ONCE(pcp->batch);
2746 free_pcppages_bulk(zone, batch, pcp);
2751 * Free a 0-order page
2753 void free_unref_page(struct page *page)
2755 unsigned long flags;
2756 unsigned long pfn = page_to_pfn(page);
2758 if (!free_unref_page_prepare(page, pfn))
2761 local_irq_save(flags);
2762 free_unref_page_commit(page, pfn);
2763 local_irq_restore(flags);
2767 * Free a list of 0-order pages
2769 void free_unref_page_list(struct list_head *list)
2771 struct page *page, *next;
2772 unsigned long flags, pfn;
2773 int batch_count = 0;
2775 /* Prepare pages for freeing */
2776 list_for_each_entry_safe(page, next, list, lru) {
2777 pfn = page_to_pfn(page);
2778 if (!free_unref_page_prepare(page, pfn))
2779 list_del(&page->lru);
2780 set_page_private(page, pfn);
2783 local_irq_save(flags);
2784 list_for_each_entry_safe(page, next, list, lru) {
2785 unsigned long pfn = page_private(page);
2787 set_page_private(page, 0);
2788 trace_mm_page_free_batched(page);
2789 free_unref_page_commit(page, pfn);
2792 * Guard against excessive IRQ disabled times when we get
2793 * a large list of pages to free.
2795 if (++batch_count == SWAP_CLUSTER_MAX) {
2796 local_irq_restore(flags);
2798 local_irq_save(flags);
2801 local_irq_restore(flags);
2805 * split_page takes a non-compound higher-order page, and splits it into
2806 * n (1<<order) sub-pages: page[0..n]
2807 * Each sub-page must be freed individually.
2809 * Note: this is probably too low level an operation for use in drivers.
2810 * Please consult with lkml before using this in your driver.
2812 void split_page(struct page *page, unsigned int order)
2816 VM_BUG_ON_PAGE(PageCompound(page), page);
2817 VM_BUG_ON_PAGE(!page_count(page), page);
2819 for (i = 1; i < (1 << order); i++)
2820 set_page_refcounted(page + i);
2821 split_page_owner(page, order);
2823 EXPORT_SYMBOL_GPL(split_page);
2825 int __isolate_free_page(struct page *page, unsigned int order)
2827 unsigned long watermark;
2831 BUG_ON(!PageBuddy(page));
2833 zone = page_zone(page);
2834 mt = get_pageblock_migratetype(page);
2836 if (!is_migrate_isolate(mt)) {
2838 * Obey watermarks as if the page was being allocated. We can
2839 * emulate a high-order watermark check with a raised order-0
2840 * watermark, because we already know our high-order page
2843 watermark = min_wmark_pages(zone) + (1UL << order);
2844 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2847 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2850 /* Remove page from free list */
2851 list_del(&page->lru);
2852 zone->free_area[order].nr_free--;
2853 rmv_page_order(page);
2856 * Set the pageblock if the isolated page is at least half of a
2859 if (order >= pageblock_order - 1) {
2860 struct page *endpage = page + (1 << order) - 1;
2861 for (; page < endpage; page += pageblock_nr_pages) {
2862 int mt = get_pageblock_migratetype(page);
2863 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2864 && !is_migrate_highatomic(mt))
2865 set_pageblock_migratetype(page,
2871 return 1UL << order;
2875 * Update NUMA hit/miss statistics
2877 * Must be called with interrupts disabled.
2879 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2882 enum numa_stat_item local_stat = NUMA_LOCAL;
2884 /* skip numa counters update if numa stats is disabled */
2885 if (!static_branch_likely(&vm_numa_stat_key))
2888 if (z->node != numa_node_id())
2889 local_stat = NUMA_OTHER;
2891 if (z->node == preferred_zone->node)
2892 __inc_numa_state(z, NUMA_HIT);
2894 __inc_numa_state(z, NUMA_MISS);
2895 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2897 __inc_numa_state(z, local_stat);
2901 /* Remove page from the per-cpu list, caller must protect the list */
2902 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2903 struct per_cpu_pages *pcp,
2904 struct list_head *list)
2909 if (list_empty(list)) {
2910 pcp->count += rmqueue_bulk(zone, 0,
2913 if (unlikely(list_empty(list)))
2917 page = list_first_entry(list, struct page, lru);
2918 list_del(&page->lru);
2920 } while (check_new_pcp(page));
2925 /* Lock and remove page from the per-cpu list */
2926 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2927 struct zone *zone, unsigned int order,
2928 gfp_t gfp_flags, int migratetype)
2930 struct per_cpu_pages *pcp;
2931 struct list_head *list;
2933 unsigned long flags;
2935 local_irq_save(flags);
2936 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2937 list = &pcp->lists[migratetype];
2938 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2940 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2941 zone_statistics(preferred_zone, zone);
2943 local_irq_restore(flags);
2948 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2951 struct page *rmqueue(struct zone *preferred_zone,
2952 struct zone *zone, unsigned int order,
2953 gfp_t gfp_flags, unsigned int alloc_flags,
2956 unsigned long flags;
2959 if (likely(order == 0)) {
2960 page = rmqueue_pcplist(preferred_zone, zone, order,
2961 gfp_flags, migratetype);
2966 * We most definitely don't want callers attempting to
2967 * allocate greater than order-1 page units with __GFP_NOFAIL.
2969 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2970 spin_lock_irqsave(&zone->lock, flags);
2974 if (alloc_flags & ALLOC_HARDER) {
2975 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2977 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2980 page = __rmqueue(zone, order, migratetype);
2981 } while (page && check_new_pages(page, order));
2982 spin_unlock(&zone->lock);
2985 __mod_zone_freepage_state(zone, -(1 << order),
2986 get_pcppage_migratetype(page));
2988 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2989 zone_statistics(preferred_zone, zone);
2990 local_irq_restore(flags);
2993 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2997 local_irq_restore(flags);
3001 #ifdef CONFIG_FAIL_PAGE_ALLOC
3004 struct fault_attr attr;
3006 bool ignore_gfp_highmem;
3007 bool ignore_gfp_reclaim;
3009 } fail_page_alloc = {
3010 .attr = FAULT_ATTR_INITIALIZER,
3011 .ignore_gfp_reclaim = true,
3012 .ignore_gfp_highmem = true,
3016 static int __init setup_fail_page_alloc(char *str)
3018 return setup_fault_attr(&fail_page_alloc.attr, str);
3020 __setup("fail_page_alloc=", setup_fail_page_alloc);
3022 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3024 if (order < fail_page_alloc.min_order)
3026 if (gfp_mask & __GFP_NOFAIL)
3028 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3030 if (fail_page_alloc.ignore_gfp_reclaim &&
3031 (gfp_mask & __GFP_DIRECT_RECLAIM))
3034 return should_fail(&fail_page_alloc.attr, 1 << order);
3037 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3039 static int __init fail_page_alloc_debugfs(void)
3041 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3044 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3045 &fail_page_alloc.attr);
3047 return PTR_ERR(dir);
3049 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3050 &fail_page_alloc.ignore_gfp_reclaim))
3052 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3053 &fail_page_alloc.ignore_gfp_highmem))
3055 if (!debugfs_create_u32("min-order", mode, dir,
3056 &fail_page_alloc.min_order))
3061 debugfs_remove_recursive(dir);
3066 late_initcall(fail_page_alloc_debugfs);
3068 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3070 #else /* CONFIG_FAIL_PAGE_ALLOC */
3072 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3077 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3080 * Return true if free base pages are above 'mark'. For high-order checks it
3081 * will return true of the order-0 watermark is reached and there is at least
3082 * one free page of a suitable size. Checking now avoids taking the zone lock
3083 * to check in the allocation paths if no pages are free.
3085 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3086 int classzone_idx, unsigned int alloc_flags,
3091 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3093 /* free_pages may go negative - that's OK */
3094 free_pages -= (1 << order) - 1;
3096 if (alloc_flags & ALLOC_HIGH)
3100 * If the caller does not have rights to ALLOC_HARDER then subtract
3101 * the high-atomic reserves. This will over-estimate the size of the
3102 * atomic reserve but it avoids a search.
3104 if (likely(!alloc_harder)) {
3105 free_pages -= z->nr_reserved_highatomic;
3108 * OOM victims can try even harder than normal ALLOC_HARDER
3109 * users on the grounds that it's definitely going to be in
3110 * the exit path shortly and free memory. Any allocation it
3111 * makes during the free path will be small and short-lived.
3113 if (alloc_flags & ALLOC_OOM)
3121 /* If allocation can't use CMA areas don't use free CMA pages */
3122 if (!(alloc_flags & ALLOC_CMA))
3123 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3127 * Check watermarks for an order-0 allocation request. If these
3128 * are not met, then a high-order request also cannot go ahead
3129 * even if a suitable page happened to be free.
3131 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3134 /* If this is an order-0 request then the watermark is fine */
3138 /* For a high-order request, check at least one suitable page is free */
3139 for (o = order; o < MAX_ORDER; o++) {
3140 struct free_area *area = &z->free_area[o];
3146 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3147 if (!list_empty(&area->free_list[mt]))
3152 if ((alloc_flags & ALLOC_CMA) &&
3153 !list_empty(&area->free_list[MIGRATE_CMA])) {
3158 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3164 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3165 int classzone_idx, unsigned int alloc_flags)
3167 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3168 zone_page_state(z, NR_FREE_PAGES));
3171 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3172 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3174 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3178 /* If allocation can't use CMA areas don't use free CMA pages */
3179 if (!(alloc_flags & ALLOC_CMA))
3180 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3184 * Fast check for order-0 only. If this fails then the reserves
3185 * need to be calculated. There is a corner case where the check
3186 * passes but only the high-order atomic reserve are free. If
3187 * the caller is !atomic then it'll uselessly search the free
3188 * list. That corner case is then slower but it is harmless.
3190 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3193 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3197 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3198 unsigned long mark, int classzone_idx)
3200 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3202 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3203 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3205 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3210 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3212 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3215 #else /* CONFIG_NUMA */
3216 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3220 #endif /* CONFIG_NUMA */
3223 * get_page_from_freelist goes through the zonelist trying to allocate
3226 static struct page *
3227 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3228 const struct alloc_context *ac)
3230 struct zoneref *z = ac->preferred_zoneref;
3232 struct pglist_data *last_pgdat_dirty_limit = NULL;
3235 * Scan zonelist, looking for a zone with enough free.
3236 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3238 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3243 if (cpusets_enabled() &&
3244 (alloc_flags & ALLOC_CPUSET) &&
3245 !__cpuset_zone_allowed(zone, gfp_mask))
3248 * When allocating a page cache page for writing, we
3249 * want to get it from a node that is within its dirty
3250 * limit, such that no single node holds more than its
3251 * proportional share of globally allowed dirty pages.
3252 * The dirty limits take into account the node's
3253 * lowmem reserves and high watermark so that kswapd
3254 * should be able to balance it without having to
3255 * write pages from its LRU list.
3257 * XXX: For now, allow allocations to potentially
3258 * exceed the per-node dirty limit in the slowpath
3259 * (spread_dirty_pages unset) before going into reclaim,
3260 * which is important when on a NUMA setup the allowed
3261 * nodes are together not big enough to reach the
3262 * global limit. The proper fix for these situations
3263 * will require awareness of nodes in the
3264 * dirty-throttling and the flusher threads.
3266 if (ac->spread_dirty_pages) {
3267 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3270 if (!node_dirty_ok(zone->zone_pgdat)) {
3271 last_pgdat_dirty_limit = zone->zone_pgdat;
3276 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3277 if (!zone_watermark_fast(zone, order, mark,
3278 ac_classzone_idx(ac), alloc_flags)) {
3281 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3283 * Watermark failed for this zone, but see if we can
3284 * grow this zone if it contains deferred pages.
3286 if (static_branch_unlikely(&deferred_pages)) {
3287 if (_deferred_grow_zone(zone, order))
3291 /* Checked here to keep the fast path fast */
3292 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3293 if (alloc_flags & ALLOC_NO_WATERMARKS)
3296 if (node_reclaim_mode == 0 ||
3297 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3300 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3302 case NODE_RECLAIM_NOSCAN:
3305 case NODE_RECLAIM_FULL:
3306 /* scanned but unreclaimable */
3309 /* did we reclaim enough */
3310 if (zone_watermark_ok(zone, order, mark,
3311 ac_classzone_idx(ac), alloc_flags))
3319 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3320 gfp_mask, alloc_flags, ac->migratetype);
3322 prep_new_page(page, order, gfp_mask, alloc_flags);
3325 * If this is a high-order atomic allocation then check
3326 * if the pageblock should be reserved for the future
3328 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3329 reserve_highatomic_pageblock(page, zone, order);
3333 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3334 /* Try again if zone has deferred pages */
3335 if (static_branch_unlikely(&deferred_pages)) {
3336 if (_deferred_grow_zone(zone, order))
3347 * Large machines with many possible nodes should not always dump per-node
3348 * meminfo in irq context.
3350 static inline bool should_suppress_show_mem(void)
3355 ret = in_interrupt();
3360 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3362 unsigned int filter = SHOW_MEM_FILTER_NODES;
3363 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3365 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3369 * This documents exceptions given to allocations in certain
3370 * contexts that are allowed to allocate outside current's set
3373 if (!(gfp_mask & __GFP_NOMEMALLOC))
3374 if (tsk_is_oom_victim(current) ||
3375 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3376 filter &= ~SHOW_MEM_FILTER_NODES;
3377 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3378 filter &= ~SHOW_MEM_FILTER_NODES;
3380 show_mem(filter, nodemask);
3383 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3385 struct va_format vaf;
3387 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3388 DEFAULT_RATELIMIT_BURST);
3390 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3393 va_start(args, fmt);
3396 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3397 current->comm, &vaf, gfp_mask, &gfp_mask,
3398 nodemask_pr_args(nodemask));
3401 cpuset_print_current_mems_allowed();
3404 warn_alloc_show_mem(gfp_mask, nodemask);
3407 static inline struct page *
3408 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3409 unsigned int alloc_flags,
3410 const struct alloc_context *ac)
3414 page = get_page_from_freelist(gfp_mask, order,
3415 alloc_flags|ALLOC_CPUSET, ac);
3417 * fallback to ignore cpuset restriction if our nodes
3421 page = get_page_from_freelist(gfp_mask, order,
3427 static inline struct page *
3428 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3429 const struct alloc_context *ac, unsigned long *did_some_progress)
3431 struct oom_control oc = {
3432 .zonelist = ac->zonelist,
3433 .nodemask = ac->nodemask,
3435 .gfp_mask = gfp_mask,
3440 *did_some_progress = 0;
3443 * Acquire the oom lock. If that fails, somebody else is
3444 * making progress for us.
3446 if (!mutex_trylock(&oom_lock)) {
3447 *did_some_progress = 1;
3448 schedule_timeout_uninterruptible(1);
3453 * Go through the zonelist yet one more time, keep very high watermark
3454 * here, this is only to catch a parallel oom killing, we must fail if
3455 * we're still under heavy pressure. But make sure that this reclaim
3456 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3457 * allocation which will never fail due to oom_lock already held.
3459 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3460 ~__GFP_DIRECT_RECLAIM, order,
3461 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3465 /* Coredumps can quickly deplete all memory reserves */
3466 if (current->flags & PF_DUMPCORE)
3468 /* The OOM killer will not help higher order allocs */
3469 if (order > PAGE_ALLOC_COSTLY_ORDER)
3472 * We have already exhausted all our reclaim opportunities without any
3473 * success so it is time to admit defeat. We will skip the OOM killer
3474 * because it is very likely that the caller has a more reasonable
3475 * fallback than shooting a random task.
3477 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3479 /* The OOM killer does not needlessly kill tasks for lowmem */
3480 if (ac->high_zoneidx < ZONE_NORMAL)
3482 if (pm_suspended_storage())
3485 * XXX: GFP_NOFS allocations should rather fail than rely on
3486 * other request to make a forward progress.
3487 * We are in an unfortunate situation where out_of_memory cannot
3488 * do much for this context but let's try it to at least get
3489 * access to memory reserved if the current task is killed (see
3490 * out_of_memory). Once filesystems are ready to handle allocation
3491 * failures more gracefully we should just bail out here.
3494 /* The OOM killer may not free memory on a specific node */
3495 if (gfp_mask & __GFP_THISNODE)
3498 /* Exhausted what can be done so it's blame time */
3499 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3500 *did_some_progress = 1;
3503 * Help non-failing allocations by giving them access to memory
3506 if (gfp_mask & __GFP_NOFAIL)
3507 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3508 ALLOC_NO_WATERMARKS, ac);
3511 mutex_unlock(&oom_lock);
3516 * Maximum number of compaction retries wit a progress before OOM
3517 * killer is consider as the only way to move forward.
3519 #define MAX_COMPACT_RETRIES 16
3521 #ifdef CONFIG_COMPACTION
3522 /* Try memory compaction for high-order allocations before reclaim */
3523 static struct page *
3524 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3525 unsigned int alloc_flags, const struct alloc_context *ac,
3526 enum compact_priority prio, enum compact_result *compact_result)
3529 unsigned int noreclaim_flag;
3534 noreclaim_flag = memalloc_noreclaim_save();
3535 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3537 memalloc_noreclaim_restore(noreclaim_flag);
3539 if (*compact_result <= COMPACT_INACTIVE)
3543 * At least in one zone compaction wasn't deferred or skipped, so let's
3544 * count a compaction stall
3546 count_vm_event(COMPACTSTALL);
3548 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3551 struct zone *zone = page_zone(page);
3553 zone->compact_blockskip_flush = false;
3554 compaction_defer_reset(zone, order, true);
3555 count_vm_event(COMPACTSUCCESS);
3560 * It's bad if compaction run occurs and fails. The most likely reason
3561 * is that pages exist, but not enough to satisfy watermarks.
3563 count_vm_event(COMPACTFAIL);
3571 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3572 enum compact_result compact_result,
3573 enum compact_priority *compact_priority,
3574 int *compaction_retries)
3576 int max_retries = MAX_COMPACT_RETRIES;
3579 int retries = *compaction_retries;
3580 enum compact_priority priority = *compact_priority;
3585 if (compaction_made_progress(compact_result))
3586 (*compaction_retries)++;
3589 * compaction considers all the zone as desperately out of memory
3590 * so it doesn't really make much sense to retry except when the
3591 * failure could be caused by insufficient priority
3593 if (compaction_failed(compact_result))
3594 goto check_priority;
3597 * make sure the compaction wasn't deferred or didn't bail out early
3598 * due to locks contention before we declare that we should give up.
3599 * But do not retry if the given zonelist is not suitable for
3602 if (compaction_withdrawn(compact_result)) {
3603 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3608 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3609 * costly ones because they are de facto nofail and invoke OOM
3610 * killer to move on while costly can fail and users are ready
3611 * to cope with that. 1/4 retries is rather arbitrary but we
3612 * would need much more detailed feedback from compaction to
3613 * make a better decision.
3615 if (order > PAGE_ALLOC_COSTLY_ORDER)
3617 if (*compaction_retries <= max_retries) {
3623 * Make sure there are attempts at the highest priority if we exhausted
3624 * all retries or failed at the lower priorities.
3627 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3628 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3630 if (*compact_priority > min_priority) {
3631 (*compact_priority)--;
3632 *compaction_retries = 0;
3636 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3640 static inline struct page *
3641 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3642 unsigned int alloc_flags, const struct alloc_context *ac,
3643 enum compact_priority prio, enum compact_result *compact_result)
3645 *compact_result = COMPACT_SKIPPED;
3650 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3651 enum compact_result compact_result,
3652 enum compact_priority *compact_priority,
3653 int *compaction_retries)
3658 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3662 * There are setups with compaction disabled which would prefer to loop
3663 * inside the allocator rather than hit the oom killer prematurely.
3664 * Let's give them a good hope and keep retrying while the order-0
3665 * watermarks are OK.
3667 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3669 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3670 ac_classzone_idx(ac), alloc_flags))
3675 #endif /* CONFIG_COMPACTION */
3677 #ifdef CONFIG_LOCKDEP
3678 struct lockdep_map __fs_reclaim_map =
3679 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3681 static bool __need_fs_reclaim(gfp_t gfp_mask)
3683 gfp_mask = current_gfp_context(gfp_mask);
3685 /* no reclaim without waiting on it */
3686 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3689 /* this guy won't enter reclaim */
3690 if (current->flags & PF_MEMALLOC)
3693 /* We're only interested __GFP_FS allocations for now */
3694 if (!(gfp_mask & __GFP_FS))
3697 if (gfp_mask & __GFP_NOLOCKDEP)
3703 void fs_reclaim_acquire(gfp_t gfp_mask)
3705 if (__need_fs_reclaim(gfp_mask))
3706 lock_map_acquire(&__fs_reclaim_map);
3708 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3710 void fs_reclaim_release(gfp_t gfp_mask)
3712 if (__need_fs_reclaim(gfp_mask))
3713 lock_map_release(&__fs_reclaim_map);
3715 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3718 /* Perform direct synchronous page reclaim */
3720 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3721 const struct alloc_context *ac)
3723 struct reclaim_state reclaim_state;
3725 unsigned int noreclaim_flag;
3729 /* We now go into synchronous reclaim */
3730 cpuset_memory_pressure_bump();
3731 noreclaim_flag = memalloc_noreclaim_save();
3732 fs_reclaim_acquire(gfp_mask);
3733 reclaim_state.reclaimed_slab = 0;
3734 current->reclaim_state = &reclaim_state;
3736 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3739 current->reclaim_state = NULL;
3740 fs_reclaim_release(gfp_mask);
3741 memalloc_noreclaim_restore(noreclaim_flag);
3748 /* The really slow allocator path where we enter direct reclaim */
3749 static inline struct page *
3750 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3751 unsigned int alloc_flags, const struct alloc_context *ac,
3752 unsigned long *did_some_progress)
3754 struct page *page = NULL;
3755 bool drained = false;
3757 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3758 if (unlikely(!(*did_some_progress)))
3762 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3765 * If an allocation failed after direct reclaim, it could be because
3766 * pages are pinned on the per-cpu lists or in high alloc reserves.
3767 * Shrink them them and try again
3769 if (!page && !drained) {
3770 unreserve_highatomic_pageblock(ac, false);
3771 drain_all_pages(NULL);
3779 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3783 pg_data_t *last_pgdat = NULL;
3785 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3786 ac->high_zoneidx, ac->nodemask) {
3787 if (last_pgdat != zone->zone_pgdat)
3788 wakeup_kswapd(zone, order, ac->high_zoneidx);
3789 last_pgdat = zone->zone_pgdat;
3793 static inline unsigned int
3794 gfp_to_alloc_flags(gfp_t gfp_mask)
3796 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3798 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3799 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3802 * The caller may dip into page reserves a bit more if the caller
3803 * cannot run direct reclaim, or if the caller has realtime scheduling
3804 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3805 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3807 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3809 if (gfp_mask & __GFP_ATOMIC) {
3811 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3812 * if it can't schedule.
3814 if (!(gfp_mask & __GFP_NOMEMALLOC))
3815 alloc_flags |= ALLOC_HARDER;
3817 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3818 * comment for __cpuset_node_allowed().
3820 alloc_flags &= ~ALLOC_CPUSET;
3821 } else if (unlikely(rt_task(current)) && !in_interrupt())
3822 alloc_flags |= ALLOC_HARDER;
3825 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3826 alloc_flags |= ALLOC_CMA;
3831 static bool oom_reserves_allowed(struct task_struct *tsk)
3833 if (!tsk_is_oom_victim(tsk))
3837 * !MMU doesn't have oom reaper so give access to memory reserves
3838 * only to the thread with TIF_MEMDIE set
3840 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3847 * Distinguish requests which really need access to full memory
3848 * reserves from oom victims which can live with a portion of it
3850 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3852 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3854 if (gfp_mask & __GFP_MEMALLOC)
3855 return ALLOC_NO_WATERMARKS;
3856 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3857 return ALLOC_NO_WATERMARKS;
3858 if (!in_interrupt()) {
3859 if (current->flags & PF_MEMALLOC)
3860 return ALLOC_NO_WATERMARKS;
3861 else if (oom_reserves_allowed(current))
3868 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3870 return !!__gfp_pfmemalloc_flags(gfp_mask);
3874 * Checks whether it makes sense to retry the reclaim to make a forward progress
3875 * for the given allocation request.
3877 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3878 * without success, or when we couldn't even meet the watermark if we
3879 * reclaimed all remaining pages on the LRU lists.
3881 * Returns true if a retry is viable or false to enter the oom path.
3884 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3885 struct alloc_context *ac, int alloc_flags,
3886 bool did_some_progress, int *no_progress_loops)
3892 * Costly allocations might have made a progress but this doesn't mean
3893 * their order will become available due to high fragmentation so
3894 * always increment the no progress counter for them
3896 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3897 *no_progress_loops = 0;
3899 (*no_progress_loops)++;
3902 * Make sure we converge to OOM if we cannot make any progress
3903 * several times in the row.
3905 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3906 /* Before OOM, exhaust highatomic_reserve */
3907 return unreserve_highatomic_pageblock(ac, true);
3911 * Keep reclaiming pages while there is a chance this will lead
3912 * somewhere. If none of the target zones can satisfy our allocation
3913 * request even if all reclaimable pages are considered then we are
3914 * screwed and have to go OOM.
3916 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3918 unsigned long available;
3919 unsigned long reclaimable;
3920 unsigned long min_wmark = min_wmark_pages(zone);
3923 available = reclaimable = zone_reclaimable_pages(zone);
3924 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3927 * Would the allocation succeed if we reclaimed all
3928 * reclaimable pages?
3930 wmark = __zone_watermark_ok(zone, order, min_wmark,
3931 ac_classzone_idx(ac), alloc_flags, available);
3932 trace_reclaim_retry_zone(z, order, reclaimable,
3933 available, min_wmark, *no_progress_loops, wmark);
3936 * If we didn't make any progress and have a lot of
3937 * dirty + writeback pages then we should wait for
3938 * an IO to complete to slow down the reclaim and
3939 * prevent from pre mature OOM
3941 if (!did_some_progress) {
3942 unsigned long write_pending;
3944 write_pending = zone_page_state_snapshot(zone,
3945 NR_ZONE_WRITE_PENDING);
3947 if (2 * write_pending > reclaimable) {
3948 congestion_wait(BLK_RW_ASYNC, HZ/10);
3954 * Memory allocation/reclaim might be called from a WQ
3955 * context and the current implementation of the WQ
3956 * concurrency control doesn't recognize that
3957 * a particular WQ is congested if the worker thread is
3958 * looping without ever sleeping. Therefore we have to
3959 * do a short sleep here rather than calling
3962 if (current->flags & PF_WQ_WORKER)
3963 schedule_timeout_uninterruptible(1);
3975 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3978 * It's possible that cpuset's mems_allowed and the nodemask from
3979 * mempolicy don't intersect. This should be normally dealt with by
3980 * policy_nodemask(), but it's possible to race with cpuset update in
3981 * such a way the check therein was true, and then it became false
3982 * before we got our cpuset_mems_cookie here.
3983 * This assumes that for all allocations, ac->nodemask can come only
3984 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3985 * when it does not intersect with the cpuset restrictions) or the
3986 * caller can deal with a violated nodemask.
3988 if (cpusets_enabled() && ac->nodemask &&
3989 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3990 ac->nodemask = NULL;
3995 * When updating a task's mems_allowed or mempolicy nodemask, it is
3996 * possible to race with parallel threads in such a way that our
3997 * allocation can fail while the mask is being updated. If we are about
3998 * to fail, check if the cpuset changed during allocation and if so,
4001 if (read_mems_allowed_retry(cpuset_mems_cookie))
4007 static inline struct page *
4008 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4009 struct alloc_context *ac)
4011 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4012 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4013 struct page *page = NULL;
4014 unsigned int alloc_flags;
4015 unsigned long did_some_progress;
4016 enum compact_priority compact_priority;
4017 enum compact_result compact_result;
4018 int compaction_retries;
4019 int no_progress_loops;
4020 unsigned int cpuset_mems_cookie;
4024 * In the slowpath, we sanity check order to avoid ever trying to
4025 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4026 * be using allocators in order of preference for an area that is
4029 if (order >= MAX_ORDER) {
4030 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4035 * We also sanity check to catch abuse of atomic reserves being used by
4036 * callers that are not in atomic context.
4038 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4039 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4040 gfp_mask &= ~__GFP_ATOMIC;
4043 compaction_retries = 0;
4044 no_progress_loops = 0;
4045 compact_priority = DEF_COMPACT_PRIORITY;
4046 cpuset_mems_cookie = read_mems_allowed_begin();
4049 * The fast path uses conservative alloc_flags to succeed only until
4050 * kswapd needs to be woken up, and to avoid the cost of setting up
4051 * alloc_flags precisely. So we do that now.
4053 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4056 * We need to recalculate the starting point for the zonelist iterator
4057 * because we might have used different nodemask in the fast path, or
4058 * there was a cpuset modification and we are retrying - otherwise we
4059 * could end up iterating over non-eligible zones endlessly.
4061 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4062 ac->high_zoneidx, ac->nodemask);
4063 if (!ac->preferred_zoneref->zone)
4066 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4067 wake_all_kswapds(order, ac);
4070 * The adjusted alloc_flags might result in immediate success, so try
4073 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4078 * For costly allocations, try direct compaction first, as it's likely
4079 * that we have enough base pages and don't need to reclaim. For non-
4080 * movable high-order allocations, do that as well, as compaction will
4081 * try prevent permanent fragmentation by migrating from blocks of the
4083 * Don't try this for allocations that are allowed to ignore
4084 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4086 if (can_direct_reclaim &&
4088 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4089 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4090 page = __alloc_pages_direct_compact(gfp_mask, order,
4092 INIT_COMPACT_PRIORITY,
4098 * Checks for costly allocations with __GFP_NORETRY, which
4099 * includes THP page fault allocations
4101 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4103 * If compaction is deferred for high-order allocations,
4104 * it is because sync compaction recently failed. If
4105 * this is the case and the caller requested a THP
4106 * allocation, we do not want to heavily disrupt the
4107 * system, so we fail the allocation instead of entering
4110 if (compact_result == COMPACT_DEFERRED)
4114 * Looks like reclaim/compaction is worth trying, but
4115 * sync compaction could be very expensive, so keep
4116 * using async compaction.
4118 compact_priority = INIT_COMPACT_PRIORITY;
4123 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4124 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4125 wake_all_kswapds(order, ac);
4127 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4129 alloc_flags = reserve_flags;
4132 * Reset the zonelist iterators if memory policies can be ignored.
4133 * These allocations are high priority and system rather than user
4136 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4137 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4138 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4139 ac->high_zoneidx, ac->nodemask);
4142 /* Attempt with potentially adjusted zonelist and alloc_flags */
4143 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4147 /* Caller is not willing to reclaim, we can't balance anything */
4148 if (!can_direct_reclaim)
4151 /* Avoid recursion of direct reclaim */
4152 if (current->flags & PF_MEMALLOC)
4155 /* Try direct reclaim and then allocating */
4156 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4157 &did_some_progress);
4161 /* Try direct compaction and then allocating */
4162 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4163 compact_priority, &compact_result);
4167 /* Do not loop if specifically requested */
4168 if (gfp_mask & __GFP_NORETRY)
4172 * Do not retry costly high order allocations unless they are
4173 * __GFP_RETRY_MAYFAIL
4175 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4178 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4179 did_some_progress > 0, &no_progress_loops))
4183 * It doesn't make any sense to retry for the compaction if the order-0
4184 * reclaim is not able to make any progress because the current
4185 * implementation of the compaction depends on the sufficient amount
4186 * of free memory (see __compaction_suitable)
4188 if (did_some_progress > 0 &&
4189 should_compact_retry(ac, order, alloc_flags,
4190 compact_result, &compact_priority,
4191 &compaction_retries))
4195 /* Deal with possible cpuset update races before we start OOM killing */
4196 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4199 /* Reclaim has failed us, start killing things */
4200 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4204 /* Avoid allocations with no watermarks from looping endlessly */
4205 if (tsk_is_oom_victim(current) &&
4206 (alloc_flags == ALLOC_OOM ||
4207 (gfp_mask & __GFP_NOMEMALLOC)))
4210 /* Retry as long as the OOM killer is making progress */
4211 if (did_some_progress) {
4212 no_progress_loops = 0;
4217 /* Deal with possible cpuset update races before we fail */
4218 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4222 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4225 if (gfp_mask & __GFP_NOFAIL) {
4227 * All existing users of the __GFP_NOFAIL are blockable, so warn
4228 * of any new users that actually require GFP_NOWAIT
4230 if (WARN_ON_ONCE(!can_direct_reclaim))
4234 * PF_MEMALLOC request from this context is rather bizarre
4235 * because we cannot reclaim anything and only can loop waiting
4236 * for somebody to do a work for us
4238 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4241 * non failing costly orders are a hard requirement which we
4242 * are not prepared for much so let's warn about these users
4243 * so that we can identify them and convert them to something
4246 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4249 * Help non-failing allocations by giving them access to memory
4250 * reserves but do not use ALLOC_NO_WATERMARKS because this
4251 * could deplete whole memory reserves which would just make
4252 * the situation worse
4254 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4262 warn_alloc(gfp_mask, ac->nodemask,
4263 "page allocation failure: order:%u", order);
4268 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4269 int preferred_nid, nodemask_t *nodemask,
4270 struct alloc_context *ac, gfp_t *alloc_mask,
4271 unsigned int *alloc_flags)
4273 ac->high_zoneidx = gfp_zone(gfp_mask);
4274 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4275 ac->nodemask = nodemask;
4276 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4278 if (cpusets_enabled()) {
4279 *alloc_mask |= __GFP_HARDWALL;
4281 ac->nodemask = &cpuset_current_mems_allowed;
4283 *alloc_flags |= ALLOC_CPUSET;
4286 fs_reclaim_acquire(gfp_mask);
4287 fs_reclaim_release(gfp_mask);
4289 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4291 if (should_fail_alloc_page(gfp_mask, order))
4294 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4295 *alloc_flags |= ALLOC_CMA;
4300 /* Determine whether to spread dirty pages and what the first usable zone */
4301 static inline void finalise_ac(gfp_t gfp_mask,
4302 unsigned int order, struct alloc_context *ac)
4304 /* Dirty zone balancing only done in the fast path */
4305 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4308 * The preferred zone is used for statistics but crucially it is
4309 * also used as the starting point for the zonelist iterator. It
4310 * may get reset for allocations that ignore memory policies.
4312 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4313 ac->high_zoneidx, ac->nodemask);
4317 * This is the 'heart' of the zoned buddy allocator.
4320 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4321 nodemask_t *nodemask)
4324 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4325 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4326 struct alloc_context ac = { };
4328 gfp_mask &= gfp_allowed_mask;
4329 alloc_mask = gfp_mask;
4330 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4333 finalise_ac(gfp_mask, order, &ac);
4335 /* First allocation attempt */
4336 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4341 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4342 * resp. GFP_NOIO which has to be inherited for all allocation requests
4343 * from a particular context which has been marked by
4344 * memalloc_no{fs,io}_{save,restore}.
4346 alloc_mask = current_gfp_context(gfp_mask);
4347 ac.spread_dirty_pages = false;
4350 * Restore the original nodemask if it was potentially replaced with
4351 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4353 if (unlikely(ac.nodemask != nodemask))
4354 ac.nodemask = nodemask;
4356 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4359 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4360 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4361 __free_pages(page, order);
4365 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4369 EXPORT_SYMBOL(__alloc_pages_nodemask);
4372 * Common helper functions.
4374 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4379 * __get_free_pages() returns a virtual address, which cannot represent
4382 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4384 page = alloc_pages(gfp_mask, order);
4387 return (unsigned long) page_address(page);
4389 EXPORT_SYMBOL(__get_free_pages);
4391 unsigned long get_zeroed_page(gfp_t gfp_mask)
4393 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4395 EXPORT_SYMBOL(get_zeroed_page);
4397 void __free_pages(struct page *page, unsigned int order)
4399 if (put_page_testzero(page)) {
4401 free_unref_page(page);
4403 __free_pages_ok(page, order);
4407 EXPORT_SYMBOL(__free_pages);
4409 void free_pages(unsigned long addr, unsigned int order)
4412 VM_BUG_ON(!virt_addr_valid((void *)addr));
4413 __free_pages(virt_to_page((void *)addr), order);
4417 EXPORT_SYMBOL(free_pages);
4421 * An arbitrary-length arbitrary-offset area of memory which resides
4422 * within a 0 or higher order page. Multiple fragments within that page
4423 * are individually refcounted, in the page's reference counter.
4425 * The page_frag functions below provide a simple allocation framework for
4426 * page fragments. This is used by the network stack and network device
4427 * drivers to provide a backing region of memory for use as either an
4428 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4430 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4433 struct page *page = NULL;
4434 gfp_t gfp = gfp_mask;
4436 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4437 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4439 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4440 PAGE_FRAG_CACHE_MAX_ORDER);
4441 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4443 if (unlikely(!page))
4444 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4446 nc->va = page ? page_address(page) : NULL;
4451 void __page_frag_cache_drain(struct page *page, unsigned int count)
4453 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4455 if (page_ref_sub_and_test(page, count)) {
4456 unsigned int order = compound_order(page);
4459 free_unref_page(page);
4461 __free_pages_ok(page, order);
4464 EXPORT_SYMBOL(__page_frag_cache_drain);
4466 void *page_frag_alloc(struct page_frag_cache *nc,
4467 unsigned int fragsz, gfp_t gfp_mask)
4469 unsigned int size = PAGE_SIZE;
4473 if (unlikely(!nc->va)) {
4475 page = __page_frag_cache_refill(nc, gfp_mask);
4479 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4480 /* if size can vary use size else just use PAGE_SIZE */
4483 /* Even if we own the page, we do not use atomic_set().
4484 * This would break get_page_unless_zero() users.
4486 page_ref_add(page, size - 1);
4488 /* reset page count bias and offset to start of new frag */
4489 nc->pfmemalloc = page_is_pfmemalloc(page);
4490 nc->pagecnt_bias = size;
4494 offset = nc->offset - fragsz;
4495 if (unlikely(offset < 0)) {
4496 page = virt_to_page(nc->va);
4498 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4501 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4502 /* if size can vary use size else just use PAGE_SIZE */
4505 /* OK, page count is 0, we can safely set it */
4506 set_page_count(page, size);
4508 /* reset page count bias and offset to start of new frag */
4509 nc->pagecnt_bias = size;
4510 offset = size - fragsz;
4514 nc->offset = offset;
4516 return nc->va + offset;
4518 EXPORT_SYMBOL(page_frag_alloc);
4521 * Frees a page fragment allocated out of either a compound or order 0 page.
4523 void page_frag_free(void *addr)
4525 struct page *page = virt_to_head_page(addr);
4527 if (unlikely(put_page_testzero(page)))
4528 __free_pages_ok(page, compound_order(page));
4530 EXPORT_SYMBOL(page_frag_free);
4532 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4536 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4537 unsigned long used = addr + PAGE_ALIGN(size);
4539 split_page(virt_to_page((void *)addr), order);
4540 while (used < alloc_end) {
4545 return (void *)addr;
4549 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4550 * @size: the number of bytes to allocate
4551 * @gfp_mask: GFP flags for the allocation
4553 * This function is similar to alloc_pages(), except that it allocates the
4554 * minimum number of pages to satisfy the request. alloc_pages() can only
4555 * allocate memory in power-of-two pages.
4557 * This function is also limited by MAX_ORDER.
4559 * Memory allocated by this function must be released by free_pages_exact().
4561 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4563 unsigned int order = get_order(size);
4566 addr = __get_free_pages(gfp_mask, order);
4567 return make_alloc_exact(addr, order, size);
4569 EXPORT_SYMBOL(alloc_pages_exact);
4572 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4574 * @nid: the preferred node ID where memory should be allocated
4575 * @size: the number of bytes to allocate
4576 * @gfp_mask: GFP flags for the allocation
4578 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4581 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4583 unsigned int order = get_order(size);
4584 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4587 return make_alloc_exact((unsigned long)page_address(p), order, size);
4591 * free_pages_exact - release memory allocated via alloc_pages_exact()
4592 * @virt: the value returned by alloc_pages_exact.
4593 * @size: size of allocation, same value as passed to alloc_pages_exact().
4595 * Release the memory allocated by a previous call to alloc_pages_exact.
4597 void free_pages_exact(void *virt, size_t size)
4599 unsigned long addr = (unsigned long)virt;
4600 unsigned long end = addr + PAGE_ALIGN(size);
4602 while (addr < end) {
4607 EXPORT_SYMBOL(free_pages_exact);
4610 * nr_free_zone_pages - count number of pages beyond high watermark
4611 * @offset: The zone index of the highest zone
4613 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4614 * high watermark within all zones at or below a given zone index. For each
4615 * zone, the number of pages is calculated as:
4617 * nr_free_zone_pages = managed_pages - high_pages
4619 static unsigned long nr_free_zone_pages(int offset)
4624 /* Just pick one node, since fallback list is circular */
4625 unsigned long sum = 0;
4627 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4629 for_each_zone_zonelist(zone, z, zonelist, offset) {
4630 unsigned long size = zone->managed_pages;
4631 unsigned long high = high_wmark_pages(zone);
4640 * nr_free_buffer_pages - count number of pages beyond high watermark
4642 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4643 * watermark within ZONE_DMA and ZONE_NORMAL.
4645 unsigned long nr_free_buffer_pages(void)
4647 return nr_free_zone_pages(gfp_zone(GFP_USER));
4649 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4652 * nr_free_pagecache_pages - count number of pages beyond high watermark
4654 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4655 * high watermark within all zones.
4657 unsigned long nr_free_pagecache_pages(void)
4659 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4662 static inline void show_node(struct zone *zone)
4664 if (IS_ENABLED(CONFIG_NUMA))
4665 printk("Node %d ", zone_to_nid(zone));
4668 long si_mem_available(void)
4671 unsigned long pagecache;
4672 unsigned long wmark_low = 0;
4673 unsigned long pages[NR_LRU_LISTS];
4677 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4678 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4681 wmark_low += zone->watermark[WMARK_LOW];
4684 * Estimate the amount of memory available for userspace allocations,
4685 * without causing swapping.
4687 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4690 * Not all the page cache can be freed, otherwise the system will
4691 * start swapping. Assume at least half of the page cache, or the
4692 * low watermark worth of cache, needs to stay.
4694 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4695 pagecache -= min(pagecache / 2, wmark_low);
4696 available += pagecache;
4699 * Part of the reclaimable slab consists of items that are in use,
4700 * and cannot be freed. Cap this estimate at the low watermark.
4702 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4703 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4710 EXPORT_SYMBOL_GPL(si_mem_available);
4712 void si_meminfo(struct sysinfo *val)
4714 val->totalram = totalram_pages;
4715 val->sharedram = global_node_page_state(NR_SHMEM);
4716 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4717 val->bufferram = nr_blockdev_pages();
4718 val->totalhigh = totalhigh_pages;
4719 val->freehigh = nr_free_highpages();
4720 val->mem_unit = PAGE_SIZE;
4723 EXPORT_SYMBOL(si_meminfo);
4726 void si_meminfo_node(struct sysinfo *val, int nid)
4728 int zone_type; /* needs to be signed */
4729 unsigned long managed_pages = 0;
4730 unsigned long managed_highpages = 0;
4731 unsigned long free_highpages = 0;
4732 pg_data_t *pgdat = NODE_DATA(nid);
4734 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4735 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4736 val->totalram = managed_pages;
4737 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4738 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4739 #ifdef CONFIG_HIGHMEM
4740 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4741 struct zone *zone = &pgdat->node_zones[zone_type];
4743 if (is_highmem(zone)) {
4744 managed_highpages += zone->managed_pages;
4745 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4748 val->totalhigh = managed_highpages;
4749 val->freehigh = free_highpages;
4751 val->totalhigh = managed_highpages;
4752 val->freehigh = free_highpages;
4754 val->mem_unit = PAGE_SIZE;
4759 * Determine whether the node should be displayed or not, depending on whether
4760 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4762 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4764 if (!(flags & SHOW_MEM_FILTER_NODES))
4768 * no node mask - aka implicit memory numa policy. Do not bother with
4769 * the synchronization - read_mems_allowed_begin - because we do not
4770 * have to be precise here.
4773 nodemask = &cpuset_current_mems_allowed;
4775 return !node_isset(nid, *nodemask);
4778 #define K(x) ((x) << (PAGE_SHIFT-10))
4780 static void show_migration_types(unsigned char type)
4782 static const char types[MIGRATE_TYPES] = {
4783 [MIGRATE_UNMOVABLE] = 'U',
4784 [MIGRATE_MOVABLE] = 'M',
4785 [MIGRATE_RECLAIMABLE] = 'E',
4786 [MIGRATE_HIGHATOMIC] = 'H',
4788 [MIGRATE_CMA] = 'C',
4790 #ifdef CONFIG_MEMORY_ISOLATION
4791 [MIGRATE_ISOLATE] = 'I',
4794 char tmp[MIGRATE_TYPES + 1];
4798 for (i = 0; i < MIGRATE_TYPES; i++) {
4799 if (type & (1 << i))
4804 printk(KERN_CONT "(%s) ", tmp);
4808 * Show free area list (used inside shift_scroll-lock stuff)
4809 * We also calculate the percentage fragmentation. We do this by counting the
4810 * memory on each free list with the exception of the first item on the list.
4813 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4816 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4818 unsigned long free_pcp = 0;
4823 for_each_populated_zone(zone) {
4824 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4827 for_each_online_cpu(cpu)
4828 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4831 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4832 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4833 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4834 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4835 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4836 " free:%lu free_pcp:%lu free_cma:%lu\n",
4837 global_node_page_state(NR_ACTIVE_ANON),
4838 global_node_page_state(NR_INACTIVE_ANON),
4839 global_node_page_state(NR_ISOLATED_ANON),
4840 global_node_page_state(NR_ACTIVE_FILE),
4841 global_node_page_state(NR_INACTIVE_FILE),
4842 global_node_page_state(NR_ISOLATED_FILE),
4843 global_node_page_state(NR_UNEVICTABLE),
4844 global_node_page_state(NR_FILE_DIRTY),
4845 global_node_page_state(NR_WRITEBACK),
4846 global_node_page_state(NR_UNSTABLE_NFS),
4847 global_node_page_state(NR_SLAB_RECLAIMABLE),
4848 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4849 global_node_page_state(NR_FILE_MAPPED),
4850 global_node_page_state(NR_SHMEM),
4851 global_zone_page_state(NR_PAGETABLE),
4852 global_zone_page_state(NR_BOUNCE),
4853 global_zone_page_state(NR_FREE_PAGES),
4855 global_zone_page_state(NR_FREE_CMA_PAGES));
4857 for_each_online_pgdat(pgdat) {
4858 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4862 " active_anon:%lukB"
4863 " inactive_anon:%lukB"
4864 " active_file:%lukB"
4865 " inactive_file:%lukB"
4866 " unevictable:%lukB"
4867 " isolated(anon):%lukB"
4868 " isolated(file):%lukB"
4873 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4875 " shmem_pmdmapped: %lukB"
4878 " writeback_tmp:%lukB"
4880 " all_unreclaimable? %s"
4883 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4884 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4885 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4886 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4887 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4888 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4889 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4890 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4891 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4892 K(node_page_state(pgdat, NR_WRITEBACK)),
4893 K(node_page_state(pgdat, NR_SHMEM)),
4894 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4895 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4896 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4898 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4900 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4901 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4902 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4906 for_each_populated_zone(zone) {
4909 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4913 for_each_online_cpu(cpu)
4914 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4923 " active_anon:%lukB"
4924 " inactive_anon:%lukB"
4925 " active_file:%lukB"
4926 " inactive_file:%lukB"
4927 " unevictable:%lukB"
4928 " writepending:%lukB"
4932 " kernel_stack:%lukB"
4940 K(zone_page_state(zone, NR_FREE_PAGES)),
4941 K(min_wmark_pages(zone)),
4942 K(low_wmark_pages(zone)),
4943 K(high_wmark_pages(zone)),
4944 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4945 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4946 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4947 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4948 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4949 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4950 K(zone->present_pages),
4951 K(zone->managed_pages),
4952 K(zone_page_state(zone, NR_MLOCK)),
4953 zone_page_state(zone, NR_KERNEL_STACK_KB),
4954 K(zone_page_state(zone, NR_PAGETABLE)),
4955 K(zone_page_state(zone, NR_BOUNCE)),
4957 K(this_cpu_read(zone->pageset->pcp.count)),
4958 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4959 printk("lowmem_reserve[]:");
4960 for (i = 0; i < MAX_NR_ZONES; i++)
4961 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4962 printk(KERN_CONT "\n");
4965 for_each_populated_zone(zone) {
4967 unsigned long nr[MAX_ORDER], flags, total = 0;
4968 unsigned char types[MAX_ORDER];
4970 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4973 printk(KERN_CONT "%s: ", zone->name);
4975 spin_lock_irqsave(&zone->lock, flags);
4976 for (order = 0; order < MAX_ORDER; order++) {
4977 struct free_area *area = &zone->free_area[order];
4980 nr[order] = area->nr_free;
4981 total += nr[order] << order;
4984 for (type = 0; type < MIGRATE_TYPES; type++) {
4985 if (!list_empty(&area->free_list[type]))
4986 types[order] |= 1 << type;
4989 spin_unlock_irqrestore(&zone->lock, flags);
4990 for (order = 0; order < MAX_ORDER; order++) {
4991 printk(KERN_CONT "%lu*%lukB ",
4992 nr[order], K(1UL) << order);
4994 show_migration_types(types[order]);
4996 printk(KERN_CONT "= %lukB\n", K(total));
4999 hugetlb_show_meminfo();
5001 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5003 show_swap_cache_info();
5006 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5008 zoneref->zone = zone;
5009 zoneref->zone_idx = zone_idx(zone);
5013 * Builds allocation fallback zone lists.
5015 * Add all populated zones of a node to the zonelist.
5017 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5020 enum zone_type zone_type = MAX_NR_ZONES;
5025 zone = pgdat->node_zones + zone_type;
5026 if (managed_zone(zone)) {
5027 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5028 check_highest_zone(zone_type);
5030 } while (zone_type);
5037 static int __parse_numa_zonelist_order(char *s)
5040 * We used to support different zonlists modes but they turned
5041 * out to be just not useful. Let's keep the warning in place
5042 * if somebody still use the cmd line parameter so that we do
5043 * not fail it silently
5045 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5046 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5052 static __init int setup_numa_zonelist_order(char *s)
5057 return __parse_numa_zonelist_order(s);
5059 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5061 char numa_zonelist_order[] = "Node";
5064 * sysctl handler for numa_zonelist_order
5066 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5067 void __user *buffer, size_t *length,
5074 return proc_dostring(table, write, buffer, length, ppos);
5075 str = memdup_user_nul(buffer, 16);
5077 return PTR_ERR(str);
5079 ret = __parse_numa_zonelist_order(str);
5085 #define MAX_NODE_LOAD (nr_online_nodes)
5086 static int node_load[MAX_NUMNODES];
5089 * find_next_best_node - find the next node that should appear in a given node's fallback list
5090 * @node: node whose fallback list we're appending
5091 * @used_node_mask: nodemask_t of already used nodes
5093 * We use a number of factors to determine which is the next node that should
5094 * appear on a given node's fallback list. The node should not have appeared
5095 * already in @node's fallback list, and it should be the next closest node
5096 * according to the distance array (which contains arbitrary distance values
5097 * from each node to each node in the system), and should also prefer nodes
5098 * with no CPUs, since presumably they'll have very little allocation pressure
5099 * on them otherwise.
5100 * It returns -1 if no node is found.
5102 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5105 int min_val = INT_MAX;
5106 int best_node = NUMA_NO_NODE;
5107 const struct cpumask *tmp = cpumask_of_node(0);
5109 /* Use the local node if we haven't already */
5110 if (!node_isset(node, *used_node_mask)) {
5111 node_set(node, *used_node_mask);
5115 for_each_node_state(n, N_MEMORY) {
5117 /* Don't want a node to appear more than once */
5118 if (node_isset(n, *used_node_mask))
5121 /* Use the distance array to find the distance */
5122 val = node_distance(node, n);
5124 /* Penalize nodes under us ("prefer the next node") */
5127 /* Give preference to headless and unused nodes */
5128 tmp = cpumask_of_node(n);
5129 if (!cpumask_empty(tmp))
5130 val += PENALTY_FOR_NODE_WITH_CPUS;
5132 /* Slight preference for less loaded node */
5133 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5134 val += node_load[n];
5136 if (val < min_val) {
5143 node_set(best_node, *used_node_mask);
5150 * Build zonelists ordered by node and zones within node.
5151 * This results in maximum locality--normal zone overflows into local
5152 * DMA zone, if any--but risks exhausting DMA zone.
5154 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5157 struct zoneref *zonerefs;
5160 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5162 for (i = 0; i < nr_nodes; i++) {
5165 pg_data_t *node = NODE_DATA(node_order[i]);
5167 nr_zones = build_zonerefs_node(node, zonerefs);
5168 zonerefs += nr_zones;
5170 zonerefs->zone = NULL;
5171 zonerefs->zone_idx = 0;
5175 * Build gfp_thisnode zonelists
5177 static void build_thisnode_zonelists(pg_data_t *pgdat)
5179 struct zoneref *zonerefs;
5182 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5183 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5184 zonerefs += nr_zones;
5185 zonerefs->zone = NULL;
5186 zonerefs->zone_idx = 0;
5190 * Build zonelists ordered by zone and nodes within zones.
5191 * This results in conserving DMA zone[s] until all Normal memory is
5192 * exhausted, but results in overflowing to remote node while memory
5193 * may still exist in local DMA zone.
5196 static void build_zonelists(pg_data_t *pgdat)
5198 static int node_order[MAX_NUMNODES];
5199 int node, load, nr_nodes = 0;
5200 nodemask_t used_mask;
5201 int local_node, prev_node;
5203 /* NUMA-aware ordering of nodes */
5204 local_node = pgdat->node_id;
5205 load = nr_online_nodes;
5206 prev_node = local_node;
5207 nodes_clear(used_mask);
5209 memset(node_order, 0, sizeof(node_order));
5210 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5212 * We don't want to pressure a particular node.
5213 * So adding penalty to the first node in same
5214 * distance group to make it round-robin.
5216 if (node_distance(local_node, node) !=
5217 node_distance(local_node, prev_node))
5218 node_load[node] = load;
5220 node_order[nr_nodes++] = node;
5225 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5226 build_thisnode_zonelists(pgdat);
5229 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5231 * Return node id of node used for "local" allocations.
5232 * I.e., first node id of first zone in arg node's generic zonelist.
5233 * Used for initializing percpu 'numa_mem', which is used primarily
5234 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5236 int local_memory_node(int node)
5240 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5241 gfp_zone(GFP_KERNEL),
5243 return z->zone->node;
5247 static void setup_min_unmapped_ratio(void);
5248 static void setup_min_slab_ratio(void);
5249 #else /* CONFIG_NUMA */
5251 static void build_zonelists(pg_data_t *pgdat)
5253 int node, local_node;
5254 struct zoneref *zonerefs;
5257 local_node = pgdat->node_id;
5259 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5260 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5261 zonerefs += nr_zones;
5264 * Now we build the zonelist so that it contains the zones
5265 * of all the other nodes.
5266 * We don't want to pressure a particular node, so when
5267 * building the zones for node N, we make sure that the
5268 * zones coming right after the local ones are those from
5269 * node N+1 (modulo N)
5271 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5272 if (!node_online(node))
5274 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5275 zonerefs += nr_zones;
5277 for (node = 0; node < local_node; node++) {
5278 if (!node_online(node))
5280 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5281 zonerefs += nr_zones;
5284 zonerefs->zone = NULL;
5285 zonerefs->zone_idx = 0;
5288 #endif /* CONFIG_NUMA */
5291 * Boot pageset table. One per cpu which is going to be used for all
5292 * zones and all nodes. The parameters will be set in such a way
5293 * that an item put on a list will immediately be handed over to
5294 * the buddy list. This is safe since pageset manipulation is done
5295 * with interrupts disabled.
5297 * The boot_pagesets must be kept even after bootup is complete for
5298 * unused processors and/or zones. They do play a role for bootstrapping
5299 * hotplugged processors.
5301 * zoneinfo_show() and maybe other functions do
5302 * not check if the processor is online before following the pageset pointer.
5303 * Other parts of the kernel may not check if the zone is available.
5305 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5306 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5307 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5309 static void __build_all_zonelists(void *data)
5312 int __maybe_unused cpu;
5313 pg_data_t *self = data;
5314 static DEFINE_SPINLOCK(lock);
5319 memset(node_load, 0, sizeof(node_load));
5323 * This node is hotadded and no memory is yet present. So just
5324 * building zonelists is fine - no need to touch other nodes.
5326 if (self && !node_online(self->node_id)) {
5327 build_zonelists(self);
5329 for_each_online_node(nid) {
5330 pg_data_t *pgdat = NODE_DATA(nid);
5332 build_zonelists(pgdat);
5335 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5337 * We now know the "local memory node" for each node--
5338 * i.e., the node of the first zone in the generic zonelist.
5339 * Set up numa_mem percpu variable for on-line cpus. During
5340 * boot, only the boot cpu should be on-line; we'll init the
5341 * secondary cpus' numa_mem as they come on-line. During
5342 * node/memory hotplug, we'll fixup all on-line cpus.
5344 for_each_online_cpu(cpu)
5345 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5352 static noinline void __init
5353 build_all_zonelists_init(void)
5357 __build_all_zonelists(NULL);
5360 * Initialize the boot_pagesets that are going to be used
5361 * for bootstrapping processors. The real pagesets for
5362 * each zone will be allocated later when the per cpu
5363 * allocator is available.
5365 * boot_pagesets are used also for bootstrapping offline
5366 * cpus if the system is already booted because the pagesets
5367 * are needed to initialize allocators on a specific cpu too.
5368 * F.e. the percpu allocator needs the page allocator which
5369 * needs the percpu allocator in order to allocate its pagesets
5370 * (a chicken-egg dilemma).
5372 for_each_possible_cpu(cpu)
5373 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5375 mminit_verify_zonelist();
5376 cpuset_init_current_mems_allowed();
5380 * unless system_state == SYSTEM_BOOTING.
5382 * __ref due to call of __init annotated helper build_all_zonelists_init
5383 * [protected by SYSTEM_BOOTING].
5385 void __ref build_all_zonelists(pg_data_t *pgdat)
5387 if (system_state == SYSTEM_BOOTING) {
5388 build_all_zonelists_init();
5390 __build_all_zonelists(pgdat);
5391 /* cpuset refresh routine should be here */
5393 vm_total_pages = nr_free_pagecache_pages();
5395 * Disable grouping by mobility if the number of pages in the
5396 * system is too low to allow the mechanism to work. It would be
5397 * more accurate, but expensive to check per-zone. This check is
5398 * made on memory-hotadd so a system can start with mobility
5399 * disabled and enable it later
5401 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5402 page_group_by_mobility_disabled = 1;
5404 page_group_by_mobility_disabled = 0;
5406 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5408 page_group_by_mobility_disabled ? "off" : "on",
5411 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5416 * Initially all pages are reserved - free ones are freed
5417 * up by free_all_bootmem() once the early boot process is
5418 * done. Non-atomic initialization, single-pass.
5420 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5421 unsigned long start_pfn, enum memmap_context context,
5422 struct vmem_altmap *altmap)
5424 unsigned long end_pfn = start_pfn + size;
5425 pg_data_t *pgdat = NODE_DATA(nid);
5427 unsigned long nr_initialised = 0;
5429 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5430 struct memblock_region *r = NULL, *tmp;
5433 if (highest_memmap_pfn < end_pfn - 1)
5434 highest_memmap_pfn = end_pfn - 1;
5437 * Honor reservation requested by the driver for this ZONE_DEVICE
5440 if (altmap && start_pfn == altmap->base_pfn)
5441 start_pfn += altmap->reserve;
5443 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5445 * There can be holes in boot-time mem_map[]s handed to this
5446 * function. They do not exist on hotplugged memory.
5448 if (context != MEMMAP_EARLY)
5451 if (!early_pfn_valid(pfn))
5453 if (!early_pfn_in_nid(pfn, nid))
5455 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5458 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5460 * Check given memblock attribute by firmware which can affect
5461 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5462 * mirrored, it's an overlapped memmap init. skip it.
5464 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5465 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5466 for_each_memblock(memory, tmp)
5467 if (pfn < memblock_region_memory_end_pfn(tmp))
5471 if (pfn >= memblock_region_memory_base_pfn(r) &&
5472 memblock_is_mirror(r)) {
5473 /* already initialized as NORMAL */
5474 pfn = memblock_region_memory_end_pfn(r);
5481 page = pfn_to_page(pfn);
5482 __init_single_page(page, pfn, zone, nid);
5483 if (context == MEMMAP_HOTPLUG)
5484 SetPageReserved(page);
5487 * Mark the block movable so that blocks are reserved for
5488 * movable at startup. This will force kernel allocations
5489 * to reserve their blocks rather than leaking throughout
5490 * the address space during boot when many long-lived
5491 * kernel allocations are made.
5493 * bitmap is created for zone's valid pfn range. but memmap
5494 * can be created for invalid pages (for alignment)
5495 * check here not to call set_pageblock_migratetype() against
5498 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5499 * because this is done early in sparse_add_one_section
5501 if (!(pfn & (pageblock_nr_pages - 1))) {
5502 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5508 static void __meminit zone_init_free_lists(struct zone *zone)
5510 unsigned int order, t;
5511 for_each_migratetype_order(order, t) {
5512 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5513 zone->free_area[order].nr_free = 0;
5517 #ifndef __HAVE_ARCH_MEMMAP_INIT
5518 #define memmap_init(size, nid, zone, start_pfn) \
5519 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5522 static int zone_batchsize(struct zone *zone)
5528 * The per-cpu-pages pools are set to around 1000th of the
5529 * size of the zone. But no more than 1/2 of a meg.
5531 * OK, so we don't know how big the cache is. So guess.
5533 batch = zone->managed_pages / 1024;
5534 if (batch * PAGE_SIZE > 512 * 1024)
5535 batch = (512 * 1024) / PAGE_SIZE;
5536 batch /= 4; /* We effectively *= 4 below */
5541 * Clamp the batch to a 2^n - 1 value. Having a power
5542 * of 2 value was found to be more likely to have
5543 * suboptimal cache aliasing properties in some cases.
5545 * For example if 2 tasks are alternately allocating
5546 * batches of pages, one task can end up with a lot
5547 * of pages of one half of the possible page colors
5548 * and the other with pages of the other colors.
5550 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5555 /* The deferral and batching of frees should be suppressed under NOMMU
5558 * The problem is that NOMMU needs to be able to allocate large chunks
5559 * of contiguous memory as there's no hardware page translation to
5560 * assemble apparent contiguous memory from discontiguous pages.
5562 * Queueing large contiguous runs of pages for batching, however,
5563 * causes the pages to actually be freed in smaller chunks. As there
5564 * can be a significant delay between the individual batches being
5565 * recycled, this leads to the once large chunks of space being
5566 * fragmented and becoming unavailable for high-order allocations.
5573 * pcp->high and pcp->batch values are related and dependent on one another:
5574 * ->batch must never be higher then ->high.
5575 * The following function updates them in a safe manner without read side
5578 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5579 * those fields changing asynchronously (acording the the above rule).
5581 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5582 * outside of boot time (or some other assurance that no concurrent updaters
5585 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5586 unsigned long batch)
5588 /* start with a fail safe value for batch */
5592 /* Update high, then batch, in order */
5599 /* a companion to pageset_set_high() */
5600 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5602 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5605 static void pageset_init(struct per_cpu_pageset *p)
5607 struct per_cpu_pages *pcp;
5610 memset(p, 0, sizeof(*p));
5614 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5615 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5618 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5621 pageset_set_batch(p, batch);
5625 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5626 * to the value high for the pageset p.
5628 static void pageset_set_high(struct per_cpu_pageset *p,
5631 unsigned long batch = max(1UL, high / 4);
5632 if ((high / 4) > (PAGE_SHIFT * 8))
5633 batch = PAGE_SHIFT * 8;
5635 pageset_update(&p->pcp, high, batch);
5638 static void pageset_set_high_and_batch(struct zone *zone,
5639 struct per_cpu_pageset *pcp)
5641 if (percpu_pagelist_fraction)
5642 pageset_set_high(pcp,
5643 (zone->managed_pages /
5644 percpu_pagelist_fraction));
5646 pageset_set_batch(pcp, zone_batchsize(zone));
5649 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5651 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5654 pageset_set_high_and_batch(zone, pcp);
5657 void __meminit setup_zone_pageset(struct zone *zone)
5660 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5661 for_each_possible_cpu(cpu)
5662 zone_pageset_init(zone, cpu);
5666 * Allocate per cpu pagesets and initialize them.
5667 * Before this call only boot pagesets were available.
5669 void __init setup_per_cpu_pageset(void)
5671 struct pglist_data *pgdat;
5674 for_each_populated_zone(zone)
5675 setup_zone_pageset(zone);
5677 for_each_online_pgdat(pgdat)
5678 pgdat->per_cpu_nodestats =
5679 alloc_percpu(struct per_cpu_nodestat);
5682 static __meminit void zone_pcp_init(struct zone *zone)
5685 * per cpu subsystem is not up at this point. The following code
5686 * relies on the ability of the linker to provide the
5687 * offset of a (static) per cpu variable into the per cpu area.
5689 zone->pageset = &boot_pageset;
5691 if (populated_zone(zone))
5692 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5693 zone->name, zone->present_pages,
5694 zone_batchsize(zone));
5697 void __meminit init_currently_empty_zone(struct zone *zone,
5698 unsigned long zone_start_pfn,
5701 struct pglist_data *pgdat = zone->zone_pgdat;
5703 pgdat->nr_zones = zone_idx(zone) + 1;
5705 zone->zone_start_pfn = zone_start_pfn;
5707 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5708 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5710 (unsigned long)zone_idx(zone),
5711 zone_start_pfn, (zone_start_pfn + size));
5713 zone_init_free_lists(zone);
5714 zone->initialized = 1;
5717 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5718 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5721 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5723 int __meminit __early_pfn_to_nid(unsigned long pfn,
5724 struct mminit_pfnnid_cache *state)
5726 unsigned long start_pfn, end_pfn;
5729 if (state->last_start <= pfn && pfn < state->last_end)
5730 return state->last_nid;
5732 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5734 state->last_start = start_pfn;
5735 state->last_end = end_pfn;
5736 state->last_nid = nid;
5741 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5744 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5745 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5746 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5748 * If an architecture guarantees that all ranges registered contain no holes
5749 * and may be freed, this this function may be used instead of calling
5750 * memblock_free_early_nid() manually.
5752 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5754 unsigned long start_pfn, end_pfn;
5757 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5758 start_pfn = min(start_pfn, max_low_pfn);
5759 end_pfn = min(end_pfn, max_low_pfn);
5761 if (start_pfn < end_pfn)
5762 memblock_free_early_nid(PFN_PHYS(start_pfn),
5763 (end_pfn - start_pfn) << PAGE_SHIFT,
5769 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5770 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5772 * If an architecture guarantees that all ranges registered contain no holes and may
5773 * be freed, this function may be used instead of calling memory_present() manually.
5775 void __init sparse_memory_present_with_active_regions(int nid)
5777 unsigned long start_pfn, end_pfn;
5780 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5781 memory_present(this_nid, start_pfn, end_pfn);
5785 * get_pfn_range_for_nid - Return the start and end page frames for a node
5786 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5787 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5788 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5790 * It returns the start and end page frame of a node based on information
5791 * provided by memblock_set_node(). If called for a node
5792 * with no available memory, a warning is printed and the start and end
5795 void __meminit get_pfn_range_for_nid(unsigned int nid,
5796 unsigned long *start_pfn, unsigned long *end_pfn)
5798 unsigned long this_start_pfn, this_end_pfn;
5804 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5805 *start_pfn = min(*start_pfn, this_start_pfn);
5806 *end_pfn = max(*end_pfn, this_end_pfn);
5809 if (*start_pfn == -1UL)
5814 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5815 * assumption is made that zones within a node are ordered in monotonic
5816 * increasing memory addresses so that the "highest" populated zone is used
5818 static void __init find_usable_zone_for_movable(void)
5821 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5822 if (zone_index == ZONE_MOVABLE)
5825 if (arch_zone_highest_possible_pfn[zone_index] >
5826 arch_zone_lowest_possible_pfn[zone_index])
5830 VM_BUG_ON(zone_index == -1);
5831 movable_zone = zone_index;
5835 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5836 * because it is sized independent of architecture. Unlike the other zones,
5837 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5838 * in each node depending on the size of each node and how evenly kernelcore
5839 * is distributed. This helper function adjusts the zone ranges
5840 * provided by the architecture for a given node by using the end of the
5841 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5842 * zones within a node are in order of monotonic increases memory addresses
5844 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5845 unsigned long zone_type,
5846 unsigned long node_start_pfn,
5847 unsigned long node_end_pfn,
5848 unsigned long *zone_start_pfn,
5849 unsigned long *zone_end_pfn)
5851 /* Only adjust if ZONE_MOVABLE is on this node */
5852 if (zone_movable_pfn[nid]) {
5853 /* Size ZONE_MOVABLE */
5854 if (zone_type == ZONE_MOVABLE) {
5855 *zone_start_pfn = zone_movable_pfn[nid];
5856 *zone_end_pfn = min(node_end_pfn,
5857 arch_zone_highest_possible_pfn[movable_zone]);
5859 /* Adjust for ZONE_MOVABLE starting within this range */
5860 } else if (!mirrored_kernelcore &&
5861 *zone_start_pfn < zone_movable_pfn[nid] &&
5862 *zone_end_pfn > zone_movable_pfn[nid]) {
5863 *zone_end_pfn = zone_movable_pfn[nid];
5865 /* Check if this whole range is within ZONE_MOVABLE */
5866 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5867 *zone_start_pfn = *zone_end_pfn;
5872 * Return the number of pages a zone spans in a node, including holes
5873 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5875 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5876 unsigned long zone_type,
5877 unsigned long node_start_pfn,
5878 unsigned long node_end_pfn,
5879 unsigned long *zone_start_pfn,
5880 unsigned long *zone_end_pfn,
5881 unsigned long *ignored)
5883 /* When hotadd a new node from cpu_up(), the node should be empty */
5884 if (!node_start_pfn && !node_end_pfn)
5887 /* Get the start and end of the zone */
5888 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5889 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5890 adjust_zone_range_for_zone_movable(nid, zone_type,
5891 node_start_pfn, node_end_pfn,
5892 zone_start_pfn, zone_end_pfn);
5894 /* Check that this node has pages within the zone's required range */
5895 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5898 /* Move the zone boundaries inside the node if necessary */
5899 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5900 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5902 /* Return the spanned pages */
5903 return *zone_end_pfn - *zone_start_pfn;
5907 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5908 * then all holes in the requested range will be accounted for.
5910 unsigned long __meminit __absent_pages_in_range(int nid,
5911 unsigned long range_start_pfn,
5912 unsigned long range_end_pfn)
5914 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5915 unsigned long start_pfn, end_pfn;
5918 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5919 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5920 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5921 nr_absent -= end_pfn - start_pfn;
5927 * absent_pages_in_range - Return number of page frames in holes within a range
5928 * @start_pfn: The start PFN to start searching for holes
5929 * @end_pfn: The end PFN to stop searching for holes
5931 * It returns the number of pages frames in memory holes within a range.
5933 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5934 unsigned long end_pfn)
5936 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5939 /* Return the number of page frames in holes in a zone on a node */
5940 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5941 unsigned long zone_type,
5942 unsigned long node_start_pfn,
5943 unsigned long node_end_pfn,
5944 unsigned long *ignored)
5946 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5947 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5948 unsigned long zone_start_pfn, zone_end_pfn;
5949 unsigned long nr_absent;
5951 /* When hotadd a new node from cpu_up(), the node should be empty */
5952 if (!node_start_pfn && !node_end_pfn)
5955 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5956 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5958 adjust_zone_range_for_zone_movable(nid, zone_type,
5959 node_start_pfn, node_end_pfn,
5960 &zone_start_pfn, &zone_end_pfn);
5961 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5964 * ZONE_MOVABLE handling.
5965 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5968 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5969 unsigned long start_pfn, end_pfn;
5970 struct memblock_region *r;
5972 for_each_memblock(memory, r) {
5973 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5974 zone_start_pfn, zone_end_pfn);
5975 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5976 zone_start_pfn, zone_end_pfn);
5978 if (zone_type == ZONE_MOVABLE &&
5979 memblock_is_mirror(r))
5980 nr_absent += end_pfn - start_pfn;
5982 if (zone_type == ZONE_NORMAL &&
5983 !memblock_is_mirror(r))
5984 nr_absent += end_pfn - start_pfn;
5991 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5992 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5993 unsigned long zone_type,
5994 unsigned long node_start_pfn,
5995 unsigned long node_end_pfn,
5996 unsigned long *zone_start_pfn,
5997 unsigned long *zone_end_pfn,
5998 unsigned long *zones_size)
6002 *zone_start_pfn = node_start_pfn;
6003 for (zone = 0; zone < zone_type; zone++)
6004 *zone_start_pfn += zones_size[zone];
6006 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6008 return zones_size[zone_type];
6011 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6012 unsigned long zone_type,
6013 unsigned long node_start_pfn,
6014 unsigned long node_end_pfn,
6015 unsigned long *zholes_size)
6020 return zholes_size[zone_type];
6023 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6025 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6026 unsigned long node_start_pfn,
6027 unsigned long node_end_pfn,
6028 unsigned long *zones_size,
6029 unsigned long *zholes_size)
6031 unsigned long realtotalpages = 0, totalpages = 0;
6034 for (i = 0; i < MAX_NR_ZONES; i++) {
6035 struct zone *zone = pgdat->node_zones + i;
6036 unsigned long zone_start_pfn, zone_end_pfn;
6037 unsigned long size, real_size;
6039 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6045 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6046 node_start_pfn, node_end_pfn,
6049 zone->zone_start_pfn = zone_start_pfn;
6051 zone->zone_start_pfn = 0;
6052 zone->spanned_pages = size;
6053 zone->present_pages = real_size;
6056 realtotalpages += real_size;
6059 pgdat->node_spanned_pages = totalpages;
6060 pgdat->node_present_pages = realtotalpages;
6061 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6065 #ifndef CONFIG_SPARSEMEM
6067 * Calculate the size of the zone->blockflags rounded to an unsigned long
6068 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6069 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6070 * round what is now in bits to nearest long in bits, then return it in
6073 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6075 unsigned long usemapsize;
6077 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6078 usemapsize = roundup(zonesize, pageblock_nr_pages);
6079 usemapsize = usemapsize >> pageblock_order;
6080 usemapsize *= NR_PAGEBLOCK_BITS;
6081 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6083 return usemapsize / 8;
6086 static void __init setup_usemap(struct pglist_data *pgdat,
6088 unsigned long zone_start_pfn,
6089 unsigned long zonesize)
6091 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6092 zone->pageblock_flags = NULL;
6094 zone->pageblock_flags =
6095 memblock_virt_alloc_node_nopanic(usemapsize,
6099 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6100 unsigned long zone_start_pfn, unsigned long zonesize) {}
6101 #endif /* CONFIG_SPARSEMEM */
6103 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6105 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6106 void __paginginit set_pageblock_order(void)
6110 /* Check that pageblock_nr_pages has not already been setup */
6111 if (pageblock_order)
6114 if (HPAGE_SHIFT > PAGE_SHIFT)
6115 order = HUGETLB_PAGE_ORDER;
6117 order = MAX_ORDER - 1;
6120 * Assume the largest contiguous order of interest is a huge page.
6121 * This value may be variable depending on boot parameters on IA64 and
6124 pageblock_order = order;
6126 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6129 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6130 * is unused as pageblock_order is set at compile-time. See
6131 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6134 void __paginginit set_pageblock_order(void)
6138 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6140 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6141 unsigned long present_pages)
6143 unsigned long pages = spanned_pages;
6146 * Provide a more accurate estimation if there are holes within
6147 * the zone and SPARSEMEM is in use. If there are holes within the
6148 * zone, each populated memory region may cost us one or two extra
6149 * memmap pages due to alignment because memmap pages for each
6150 * populated regions may not be naturally aligned on page boundary.
6151 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6153 if (spanned_pages > present_pages + (present_pages >> 4) &&
6154 IS_ENABLED(CONFIG_SPARSEMEM))
6155 pages = present_pages;
6157 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6161 * Set up the zone data structures:
6162 * - mark all pages reserved
6163 * - mark all memory queues empty
6164 * - clear the memory bitmaps
6166 * NOTE: pgdat should get zeroed by caller.
6168 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6171 int nid = pgdat->node_id;
6173 pgdat_resize_init(pgdat);
6174 #ifdef CONFIG_NUMA_BALANCING
6175 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6176 pgdat->numabalancing_migrate_nr_pages = 0;
6177 pgdat->numabalancing_migrate_next_window = jiffies;
6179 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6180 spin_lock_init(&pgdat->split_queue_lock);
6181 INIT_LIST_HEAD(&pgdat->split_queue);
6182 pgdat->split_queue_len = 0;
6184 init_waitqueue_head(&pgdat->kswapd_wait);
6185 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6186 #ifdef CONFIG_COMPACTION
6187 init_waitqueue_head(&pgdat->kcompactd_wait);
6189 pgdat_page_ext_init(pgdat);
6190 spin_lock_init(&pgdat->lru_lock);
6191 lruvec_init(node_lruvec(pgdat));
6193 pgdat->per_cpu_nodestats = &boot_nodestats;
6195 for (j = 0; j < MAX_NR_ZONES; j++) {
6196 struct zone *zone = pgdat->node_zones + j;
6197 unsigned long size, realsize, freesize, memmap_pages;
6198 unsigned long zone_start_pfn = zone->zone_start_pfn;
6200 size = zone->spanned_pages;
6201 realsize = freesize = zone->present_pages;
6204 * Adjust freesize so that it accounts for how much memory
6205 * is used by this zone for memmap. This affects the watermark
6206 * and per-cpu initialisations
6208 memmap_pages = calc_memmap_size(size, realsize);
6209 if (!is_highmem_idx(j)) {
6210 if (freesize >= memmap_pages) {
6211 freesize -= memmap_pages;
6214 " %s zone: %lu pages used for memmap\n",
6215 zone_names[j], memmap_pages);
6217 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6218 zone_names[j], memmap_pages, freesize);
6221 /* Account for reserved pages */
6222 if (j == 0 && freesize > dma_reserve) {
6223 freesize -= dma_reserve;
6224 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6225 zone_names[0], dma_reserve);
6228 if (!is_highmem_idx(j))
6229 nr_kernel_pages += freesize;
6230 /* Charge for highmem memmap if there are enough kernel pages */
6231 else if (nr_kernel_pages > memmap_pages * 2)
6232 nr_kernel_pages -= memmap_pages;
6233 nr_all_pages += freesize;
6236 * Set an approximate value for lowmem here, it will be adjusted
6237 * when the bootmem allocator frees pages into the buddy system.
6238 * And all highmem pages will be managed by the buddy system.
6240 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6244 zone->name = zone_names[j];
6245 zone->zone_pgdat = pgdat;
6246 spin_lock_init(&zone->lock);
6247 zone_seqlock_init(zone);
6248 zone_pcp_init(zone);
6253 set_pageblock_order();
6254 setup_usemap(pgdat, zone, zone_start_pfn, size);
6255 init_currently_empty_zone(zone, zone_start_pfn, size);
6256 memmap_init(size, nid, j, zone_start_pfn);
6260 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6261 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6263 unsigned long __maybe_unused start = 0;
6264 unsigned long __maybe_unused offset = 0;
6266 /* Skip empty nodes */
6267 if (!pgdat->node_spanned_pages)
6270 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6271 offset = pgdat->node_start_pfn - start;
6272 /* ia64 gets its own node_mem_map, before this, without bootmem */
6273 if (!pgdat->node_mem_map) {
6274 unsigned long size, end;
6278 * The zone's endpoints aren't required to be MAX_ORDER
6279 * aligned but the node_mem_map endpoints must be in order
6280 * for the buddy allocator to function correctly.
6282 end = pgdat_end_pfn(pgdat);
6283 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6284 size = (end - start) * sizeof(struct page);
6285 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6286 pgdat->node_mem_map = map + offset;
6288 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6289 __func__, pgdat->node_id, (unsigned long)pgdat,
6290 (unsigned long)pgdat->node_mem_map);
6291 #ifndef CONFIG_NEED_MULTIPLE_NODES
6293 * With no DISCONTIG, the global mem_map is just set as node 0's
6295 if (pgdat == NODE_DATA(0)) {
6296 mem_map = NODE_DATA(0)->node_mem_map;
6297 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6298 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6300 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6305 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6306 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6308 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6309 unsigned long node_start_pfn, unsigned long *zholes_size)
6311 pg_data_t *pgdat = NODE_DATA(nid);
6312 unsigned long start_pfn = 0;
6313 unsigned long end_pfn = 0;
6315 /* pg_data_t should be reset to zero when it's allocated */
6316 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6318 pgdat->node_id = nid;
6319 pgdat->node_start_pfn = node_start_pfn;
6320 pgdat->per_cpu_nodestats = NULL;
6321 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6322 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6323 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6324 (u64)start_pfn << PAGE_SHIFT,
6325 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6327 start_pfn = node_start_pfn;
6329 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6330 zones_size, zholes_size);
6332 alloc_node_mem_map(pgdat);
6334 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6336 * We start only with one section of pages, more pages are added as
6337 * needed until the rest of deferred pages are initialized.
6339 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6340 pgdat->node_spanned_pages);
6341 pgdat->first_deferred_pfn = ULONG_MAX;
6343 free_area_init_core(pgdat);
6346 #ifdef CONFIG_HAVE_MEMBLOCK
6348 * Only struct pages that are backed by physical memory are zeroed and
6349 * initialized by going through __init_single_page(). But, there are some
6350 * struct pages which are reserved in memblock allocator and their fields
6351 * may be accessed (for example page_to_pfn() on some configuration accesses
6352 * flags). We must explicitly zero those struct pages.
6354 void __paginginit zero_resv_unavail(void)
6356 phys_addr_t start, end;
6361 * Loop through ranges that are reserved, but do not have reported
6362 * physical memory backing.
6365 for_each_resv_unavail_range(i, &start, &end) {
6366 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6367 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages)))
6369 mm_zero_struct_page(pfn_to_page(pfn));
6375 * Struct pages that do not have backing memory. This could be because
6376 * firmware is using some of this memory, or for some other reasons.
6377 * Once memblock is changed so such behaviour is not allowed: i.e.
6378 * list of "reserved" memory must be a subset of list of "memory", then
6379 * this code can be removed.
6382 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6384 #endif /* CONFIG_HAVE_MEMBLOCK */
6386 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6388 #if MAX_NUMNODES > 1
6390 * Figure out the number of possible node ids.
6392 void __init setup_nr_node_ids(void)
6394 unsigned int highest;
6396 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6397 nr_node_ids = highest + 1;
6402 * node_map_pfn_alignment - determine the maximum internode alignment
6404 * This function should be called after node map is populated and sorted.
6405 * It calculates the maximum power of two alignment which can distinguish
6408 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6409 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6410 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6411 * shifted, 1GiB is enough and this function will indicate so.
6413 * This is used to test whether pfn -> nid mapping of the chosen memory
6414 * model has fine enough granularity to avoid incorrect mapping for the
6415 * populated node map.
6417 * Returns the determined alignment in pfn's. 0 if there is no alignment
6418 * requirement (single node).
6420 unsigned long __init node_map_pfn_alignment(void)
6422 unsigned long accl_mask = 0, last_end = 0;
6423 unsigned long start, end, mask;
6427 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6428 if (!start || last_nid < 0 || last_nid == nid) {
6435 * Start with a mask granular enough to pin-point to the
6436 * start pfn and tick off bits one-by-one until it becomes
6437 * too coarse to separate the current node from the last.
6439 mask = ~((1 << __ffs(start)) - 1);
6440 while (mask && last_end <= (start & (mask << 1)))
6443 /* accumulate all internode masks */
6447 /* convert mask to number of pages */
6448 return ~accl_mask + 1;
6451 /* Find the lowest pfn for a node */
6452 static unsigned long __init find_min_pfn_for_node(int nid)
6454 unsigned long min_pfn = ULONG_MAX;
6455 unsigned long start_pfn;
6458 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6459 min_pfn = min(min_pfn, start_pfn);
6461 if (min_pfn == ULONG_MAX) {
6462 pr_warn("Could not find start_pfn for node %d\n", nid);
6470 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6472 * It returns the minimum PFN based on information provided via
6473 * memblock_set_node().
6475 unsigned long __init find_min_pfn_with_active_regions(void)
6477 return find_min_pfn_for_node(MAX_NUMNODES);
6481 * early_calculate_totalpages()
6482 * Sum pages in active regions for movable zone.
6483 * Populate N_MEMORY for calculating usable_nodes.
6485 static unsigned long __init early_calculate_totalpages(void)
6487 unsigned long totalpages = 0;
6488 unsigned long start_pfn, end_pfn;
6491 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6492 unsigned long pages = end_pfn - start_pfn;
6494 totalpages += pages;
6496 node_set_state(nid, N_MEMORY);
6502 * Find the PFN the Movable zone begins in each node. Kernel memory
6503 * is spread evenly between nodes as long as the nodes have enough
6504 * memory. When they don't, some nodes will have more kernelcore than
6507 static void __init find_zone_movable_pfns_for_nodes(void)
6510 unsigned long usable_startpfn;
6511 unsigned long kernelcore_node, kernelcore_remaining;
6512 /* save the state before borrow the nodemask */
6513 nodemask_t saved_node_state = node_states[N_MEMORY];
6514 unsigned long totalpages = early_calculate_totalpages();
6515 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6516 struct memblock_region *r;
6518 /* Need to find movable_zone earlier when movable_node is specified. */
6519 find_usable_zone_for_movable();
6522 * If movable_node is specified, ignore kernelcore and movablecore
6525 if (movable_node_is_enabled()) {
6526 for_each_memblock(memory, r) {
6527 if (!memblock_is_hotpluggable(r))
6532 usable_startpfn = PFN_DOWN(r->base);
6533 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6534 min(usable_startpfn, zone_movable_pfn[nid]) :
6542 * If kernelcore=mirror is specified, ignore movablecore option
6544 if (mirrored_kernelcore) {
6545 bool mem_below_4gb_not_mirrored = false;
6547 for_each_memblock(memory, r) {
6548 if (memblock_is_mirror(r))
6553 usable_startpfn = memblock_region_memory_base_pfn(r);
6555 if (usable_startpfn < 0x100000) {
6556 mem_below_4gb_not_mirrored = true;
6560 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6561 min(usable_startpfn, zone_movable_pfn[nid]) :
6565 if (mem_below_4gb_not_mirrored)
6566 pr_warn("This configuration results in unmirrored kernel memory.");
6572 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6573 * amount of necessary memory.
6575 if (required_kernelcore_percent)
6576 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6578 if (required_movablecore_percent)
6579 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6583 * If movablecore= was specified, calculate what size of
6584 * kernelcore that corresponds so that memory usable for
6585 * any allocation type is evenly spread. If both kernelcore
6586 * and movablecore are specified, then the value of kernelcore
6587 * will be used for required_kernelcore if it's greater than
6588 * what movablecore would have allowed.
6590 if (required_movablecore) {
6591 unsigned long corepages;
6594 * Round-up so that ZONE_MOVABLE is at least as large as what
6595 * was requested by the user
6597 required_movablecore =
6598 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6599 required_movablecore = min(totalpages, required_movablecore);
6600 corepages = totalpages - required_movablecore;
6602 required_kernelcore = max(required_kernelcore, corepages);
6606 * If kernelcore was not specified or kernelcore size is larger
6607 * than totalpages, there is no ZONE_MOVABLE.
6609 if (!required_kernelcore || required_kernelcore >= totalpages)
6612 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6613 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6616 /* Spread kernelcore memory as evenly as possible throughout nodes */
6617 kernelcore_node = required_kernelcore / usable_nodes;
6618 for_each_node_state(nid, N_MEMORY) {
6619 unsigned long start_pfn, end_pfn;
6622 * Recalculate kernelcore_node if the division per node
6623 * now exceeds what is necessary to satisfy the requested
6624 * amount of memory for the kernel
6626 if (required_kernelcore < kernelcore_node)
6627 kernelcore_node = required_kernelcore / usable_nodes;
6630 * As the map is walked, we track how much memory is usable
6631 * by the kernel using kernelcore_remaining. When it is
6632 * 0, the rest of the node is usable by ZONE_MOVABLE
6634 kernelcore_remaining = kernelcore_node;
6636 /* Go through each range of PFNs within this node */
6637 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6638 unsigned long size_pages;
6640 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6641 if (start_pfn >= end_pfn)
6644 /* Account for what is only usable for kernelcore */
6645 if (start_pfn < usable_startpfn) {
6646 unsigned long kernel_pages;
6647 kernel_pages = min(end_pfn, usable_startpfn)
6650 kernelcore_remaining -= min(kernel_pages,
6651 kernelcore_remaining);
6652 required_kernelcore -= min(kernel_pages,
6653 required_kernelcore);
6655 /* Continue if range is now fully accounted */
6656 if (end_pfn <= usable_startpfn) {
6659 * Push zone_movable_pfn to the end so
6660 * that if we have to rebalance
6661 * kernelcore across nodes, we will
6662 * not double account here
6664 zone_movable_pfn[nid] = end_pfn;
6667 start_pfn = usable_startpfn;
6671 * The usable PFN range for ZONE_MOVABLE is from
6672 * start_pfn->end_pfn. Calculate size_pages as the
6673 * number of pages used as kernelcore
6675 size_pages = end_pfn - start_pfn;
6676 if (size_pages > kernelcore_remaining)
6677 size_pages = kernelcore_remaining;
6678 zone_movable_pfn[nid] = start_pfn + size_pages;
6681 * Some kernelcore has been met, update counts and
6682 * break if the kernelcore for this node has been
6685 required_kernelcore -= min(required_kernelcore,
6687 kernelcore_remaining -= size_pages;
6688 if (!kernelcore_remaining)
6694 * If there is still required_kernelcore, we do another pass with one
6695 * less node in the count. This will push zone_movable_pfn[nid] further
6696 * along on the nodes that still have memory until kernelcore is
6700 if (usable_nodes && required_kernelcore > usable_nodes)
6704 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6705 for (nid = 0; nid < MAX_NUMNODES; nid++)
6706 zone_movable_pfn[nid] =
6707 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6710 /* restore the node_state */
6711 node_states[N_MEMORY] = saved_node_state;
6714 /* Any regular or high memory on that node ? */
6715 static void check_for_memory(pg_data_t *pgdat, int nid)
6717 enum zone_type zone_type;
6719 if (N_MEMORY == N_NORMAL_MEMORY)
6722 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6723 struct zone *zone = &pgdat->node_zones[zone_type];
6724 if (populated_zone(zone)) {
6725 node_set_state(nid, N_HIGH_MEMORY);
6726 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6727 zone_type <= ZONE_NORMAL)
6728 node_set_state(nid, N_NORMAL_MEMORY);
6735 * free_area_init_nodes - Initialise all pg_data_t and zone data
6736 * @max_zone_pfn: an array of max PFNs for each zone
6738 * This will call free_area_init_node() for each active node in the system.
6739 * Using the page ranges provided by memblock_set_node(), the size of each
6740 * zone in each node and their holes is calculated. If the maximum PFN
6741 * between two adjacent zones match, it is assumed that the zone is empty.
6742 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6743 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6744 * starts where the previous one ended. For example, ZONE_DMA32 starts
6745 * at arch_max_dma_pfn.
6747 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6749 unsigned long start_pfn, end_pfn;
6752 /* Record where the zone boundaries are */
6753 memset(arch_zone_lowest_possible_pfn, 0,
6754 sizeof(arch_zone_lowest_possible_pfn));
6755 memset(arch_zone_highest_possible_pfn, 0,
6756 sizeof(arch_zone_highest_possible_pfn));
6758 start_pfn = find_min_pfn_with_active_regions();
6760 for (i = 0; i < MAX_NR_ZONES; i++) {
6761 if (i == ZONE_MOVABLE)
6764 end_pfn = max(max_zone_pfn[i], start_pfn);
6765 arch_zone_lowest_possible_pfn[i] = start_pfn;
6766 arch_zone_highest_possible_pfn[i] = end_pfn;
6768 start_pfn = end_pfn;
6771 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6772 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6773 find_zone_movable_pfns_for_nodes();
6775 /* Print out the zone ranges */
6776 pr_info("Zone ranges:\n");
6777 for (i = 0; i < MAX_NR_ZONES; i++) {
6778 if (i == ZONE_MOVABLE)
6780 pr_info(" %-8s ", zone_names[i]);
6781 if (arch_zone_lowest_possible_pfn[i] ==
6782 arch_zone_highest_possible_pfn[i])
6785 pr_cont("[mem %#018Lx-%#018Lx]\n",
6786 (u64)arch_zone_lowest_possible_pfn[i]
6788 ((u64)arch_zone_highest_possible_pfn[i]
6789 << PAGE_SHIFT) - 1);
6792 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6793 pr_info("Movable zone start for each node\n");
6794 for (i = 0; i < MAX_NUMNODES; i++) {
6795 if (zone_movable_pfn[i])
6796 pr_info(" Node %d: %#018Lx\n", i,
6797 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6800 /* Print out the early node map */
6801 pr_info("Early memory node ranges\n");
6802 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6803 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6804 (u64)start_pfn << PAGE_SHIFT,
6805 ((u64)end_pfn << PAGE_SHIFT) - 1);
6807 /* Initialise every node */
6808 mminit_verify_pageflags_layout();
6809 setup_nr_node_ids();
6810 for_each_online_node(nid) {
6811 pg_data_t *pgdat = NODE_DATA(nid);
6812 free_area_init_node(nid, NULL,
6813 find_min_pfn_for_node(nid), NULL);
6815 /* Any memory on that node */
6816 if (pgdat->node_present_pages)
6817 node_set_state(nid, N_MEMORY);
6818 check_for_memory(pgdat, nid);
6820 zero_resv_unavail();
6823 static int __init cmdline_parse_core(char *p, unsigned long *core,
6824 unsigned long *percent)
6826 unsigned long long coremem;
6832 /* Value may be a percentage of total memory, otherwise bytes */
6833 coremem = simple_strtoull(p, &endptr, 0);
6834 if (*endptr == '%') {
6835 /* Paranoid check for percent values greater than 100 */
6836 WARN_ON(coremem > 100);
6840 coremem = memparse(p, &p);
6841 /* Paranoid check that UL is enough for the coremem value */
6842 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6844 *core = coremem >> PAGE_SHIFT;
6851 * kernelcore=size sets the amount of memory for use for allocations that
6852 * cannot be reclaimed or migrated.
6854 static int __init cmdline_parse_kernelcore(char *p)
6856 /* parse kernelcore=mirror */
6857 if (parse_option_str(p, "mirror")) {
6858 mirrored_kernelcore = true;
6862 return cmdline_parse_core(p, &required_kernelcore,
6863 &required_kernelcore_percent);
6867 * movablecore=size sets the amount of memory for use for allocations that
6868 * can be reclaimed or migrated.
6870 static int __init cmdline_parse_movablecore(char *p)
6872 return cmdline_parse_core(p, &required_movablecore,
6873 &required_movablecore_percent);
6876 early_param("kernelcore", cmdline_parse_kernelcore);
6877 early_param("movablecore", cmdline_parse_movablecore);
6879 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6881 void adjust_managed_page_count(struct page *page, long count)
6883 spin_lock(&managed_page_count_lock);
6884 page_zone(page)->managed_pages += count;
6885 totalram_pages += count;
6886 #ifdef CONFIG_HIGHMEM
6887 if (PageHighMem(page))
6888 totalhigh_pages += count;
6890 spin_unlock(&managed_page_count_lock);
6892 EXPORT_SYMBOL(adjust_managed_page_count);
6894 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6897 unsigned long pages = 0;
6899 start = (void *)PAGE_ALIGN((unsigned long)start);
6900 end = (void *)((unsigned long)end & PAGE_MASK);
6901 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6902 if ((unsigned int)poison <= 0xFF)
6903 memset(pos, poison, PAGE_SIZE);
6904 free_reserved_page(virt_to_page(pos));
6908 pr_info("Freeing %s memory: %ldK\n",
6909 s, pages << (PAGE_SHIFT - 10));
6913 EXPORT_SYMBOL(free_reserved_area);
6915 #ifdef CONFIG_HIGHMEM
6916 void free_highmem_page(struct page *page)
6918 __free_reserved_page(page);
6920 page_zone(page)->managed_pages++;
6926 void __init mem_init_print_info(const char *str)
6928 unsigned long physpages, codesize, datasize, rosize, bss_size;
6929 unsigned long init_code_size, init_data_size;
6931 physpages = get_num_physpages();
6932 codesize = _etext - _stext;
6933 datasize = _edata - _sdata;
6934 rosize = __end_rodata - __start_rodata;
6935 bss_size = __bss_stop - __bss_start;
6936 init_data_size = __init_end - __init_begin;
6937 init_code_size = _einittext - _sinittext;
6940 * Detect special cases and adjust section sizes accordingly:
6941 * 1) .init.* may be embedded into .data sections
6942 * 2) .init.text.* may be out of [__init_begin, __init_end],
6943 * please refer to arch/tile/kernel/vmlinux.lds.S.
6944 * 3) .rodata.* may be embedded into .text or .data sections.
6946 #define adj_init_size(start, end, size, pos, adj) \
6948 if (start <= pos && pos < end && size > adj) \
6952 adj_init_size(__init_begin, __init_end, init_data_size,
6953 _sinittext, init_code_size);
6954 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6955 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6956 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6957 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6959 #undef adj_init_size
6961 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6962 #ifdef CONFIG_HIGHMEM
6966 nr_free_pages() << (PAGE_SHIFT - 10),
6967 physpages << (PAGE_SHIFT - 10),
6968 codesize >> 10, datasize >> 10, rosize >> 10,
6969 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6970 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6971 totalcma_pages << (PAGE_SHIFT - 10),
6972 #ifdef CONFIG_HIGHMEM
6973 totalhigh_pages << (PAGE_SHIFT - 10),
6975 str ? ", " : "", str ? str : "");
6979 * set_dma_reserve - set the specified number of pages reserved in the first zone
6980 * @new_dma_reserve: The number of pages to mark reserved
6982 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6983 * In the DMA zone, a significant percentage may be consumed by kernel image
6984 * and other unfreeable allocations which can skew the watermarks badly. This
6985 * function may optionally be used to account for unfreeable pages in the
6986 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6987 * smaller per-cpu batchsize.
6989 void __init set_dma_reserve(unsigned long new_dma_reserve)
6991 dma_reserve = new_dma_reserve;
6994 void __init free_area_init(unsigned long *zones_size)
6996 free_area_init_node(0, zones_size,
6997 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6998 zero_resv_unavail();
7001 static int page_alloc_cpu_dead(unsigned int cpu)
7004 lru_add_drain_cpu(cpu);
7008 * Spill the event counters of the dead processor
7009 * into the current processors event counters.
7010 * This artificially elevates the count of the current
7013 vm_events_fold_cpu(cpu);
7016 * Zero the differential counters of the dead processor
7017 * so that the vm statistics are consistent.
7019 * This is only okay since the processor is dead and cannot
7020 * race with what we are doing.
7022 cpu_vm_stats_fold(cpu);
7026 void __init page_alloc_init(void)
7030 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7031 "mm/page_alloc:dead", NULL,
7032 page_alloc_cpu_dead);
7037 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7038 * or min_free_kbytes changes.
7040 static void calculate_totalreserve_pages(void)
7042 struct pglist_data *pgdat;
7043 unsigned long reserve_pages = 0;
7044 enum zone_type i, j;
7046 for_each_online_pgdat(pgdat) {
7048 pgdat->totalreserve_pages = 0;
7050 for (i = 0; i < MAX_NR_ZONES; i++) {
7051 struct zone *zone = pgdat->node_zones + i;
7054 /* Find valid and maximum lowmem_reserve in the zone */
7055 for (j = i; j < MAX_NR_ZONES; j++) {
7056 if (zone->lowmem_reserve[j] > max)
7057 max = zone->lowmem_reserve[j];
7060 /* we treat the high watermark as reserved pages. */
7061 max += high_wmark_pages(zone);
7063 if (max > zone->managed_pages)
7064 max = zone->managed_pages;
7066 pgdat->totalreserve_pages += max;
7068 reserve_pages += max;
7071 totalreserve_pages = reserve_pages;
7075 * setup_per_zone_lowmem_reserve - called whenever
7076 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7077 * has a correct pages reserved value, so an adequate number of
7078 * pages are left in the zone after a successful __alloc_pages().
7080 static void setup_per_zone_lowmem_reserve(void)
7082 struct pglist_data *pgdat;
7083 enum zone_type j, idx;
7085 for_each_online_pgdat(pgdat) {
7086 for (j = 0; j < MAX_NR_ZONES; j++) {
7087 struct zone *zone = pgdat->node_zones + j;
7088 unsigned long managed_pages = zone->managed_pages;
7090 zone->lowmem_reserve[j] = 0;
7094 struct zone *lower_zone;
7098 if (sysctl_lowmem_reserve_ratio[idx] < 1)
7099 sysctl_lowmem_reserve_ratio[idx] = 1;
7101 lower_zone = pgdat->node_zones + idx;
7102 lower_zone->lowmem_reserve[j] = managed_pages /
7103 sysctl_lowmem_reserve_ratio[idx];
7104 managed_pages += lower_zone->managed_pages;
7109 /* update totalreserve_pages */
7110 calculate_totalreserve_pages();
7113 static void __setup_per_zone_wmarks(void)
7115 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7116 unsigned long lowmem_pages = 0;
7118 unsigned long flags;
7120 /* Calculate total number of !ZONE_HIGHMEM pages */
7121 for_each_zone(zone) {
7122 if (!is_highmem(zone))
7123 lowmem_pages += zone->managed_pages;
7126 for_each_zone(zone) {
7129 spin_lock_irqsave(&zone->lock, flags);
7130 tmp = (u64)pages_min * zone->managed_pages;
7131 do_div(tmp, lowmem_pages);
7132 if (is_highmem(zone)) {
7134 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7135 * need highmem pages, so cap pages_min to a small
7138 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7139 * deltas control asynch page reclaim, and so should
7140 * not be capped for highmem.
7142 unsigned long min_pages;
7144 min_pages = zone->managed_pages / 1024;
7145 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7146 zone->watermark[WMARK_MIN] = min_pages;
7149 * If it's a lowmem zone, reserve a number of pages
7150 * proportionate to the zone's size.
7152 zone->watermark[WMARK_MIN] = tmp;
7156 * Set the kswapd watermarks distance according to the
7157 * scale factor in proportion to available memory, but
7158 * ensure a minimum size on small systems.
7160 tmp = max_t(u64, tmp >> 2,
7161 mult_frac(zone->managed_pages,
7162 watermark_scale_factor, 10000));
7164 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7165 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7167 spin_unlock_irqrestore(&zone->lock, flags);
7170 /* update totalreserve_pages */
7171 calculate_totalreserve_pages();
7175 * setup_per_zone_wmarks - called when min_free_kbytes changes
7176 * or when memory is hot-{added|removed}
7178 * Ensures that the watermark[min,low,high] values for each zone are set
7179 * correctly with respect to min_free_kbytes.
7181 void setup_per_zone_wmarks(void)
7183 static DEFINE_SPINLOCK(lock);
7186 __setup_per_zone_wmarks();
7191 * Initialise min_free_kbytes.
7193 * For small machines we want it small (128k min). For large machines
7194 * we want it large (64MB max). But it is not linear, because network
7195 * bandwidth does not increase linearly with machine size. We use
7197 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7198 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7214 int __meminit init_per_zone_wmark_min(void)
7216 unsigned long lowmem_kbytes;
7217 int new_min_free_kbytes;
7219 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7220 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7222 if (new_min_free_kbytes > user_min_free_kbytes) {
7223 min_free_kbytes = new_min_free_kbytes;
7224 if (min_free_kbytes < 128)
7225 min_free_kbytes = 128;
7226 if (min_free_kbytes > 65536)
7227 min_free_kbytes = 65536;
7229 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7230 new_min_free_kbytes, user_min_free_kbytes);
7232 setup_per_zone_wmarks();
7233 refresh_zone_stat_thresholds();
7234 setup_per_zone_lowmem_reserve();
7237 setup_min_unmapped_ratio();
7238 setup_min_slab_ratio();
7243 core_initcall(init_per_zone_wmark_min)
7246 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7247 * that we can call two helper functions whenever min_free_kbytes
7250 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7251 void __user *buffer, size_t *length, loff_t *ppos)
7255 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7260 user_min_free_kbytes = min_free_kbytes;
7261 setup_per_zone_wmarks();
7266 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7267 void __user *buffer, size_t *length, loff_t *ppos)
7271 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7276 setup_per_zone_wmarks();
7282 static void setup_min_unmapped_ratio(void)
7287 for_each_online_pgdat(pgdat)
7288 pgdat->min_unmapped_pages = 0;
7291 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7292 sysctl_min_unmapped_ratio) / 100;
7296 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7297 void __user *buffer, size_t *length, loff_t *ppos)
7301 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7305 setup_min_unmapped_ratio();
7310 static void setup_min_slab_ratio(void)
7315 for_each_online_pgdat(pgdat)
7316 pgdat->min_slab_pages = 0;
7319 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7320 sysctl_min_slab_ratio) / 100;
7323 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7324 void __user *buffer, size_t *length, loff_t *ppos)
7328 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7332 setup_min_slab_ratio();
7339 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7340 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7341 * whenever sysctl_lowmem_reserve_ratio changes.
7343 * The reserve ratio obviously has absolutely no relation with the
7344 * minimum watermarks. The lowmem reserve ratio can only make sense
7345 * if in function of the boot time zone sizes.
7347 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7348 void __user *buffer, size_t *length, loff_t *ppos)
7350 proc_dointvec_minmax(table, write, buffer, length, ppos);
7351 setup_per_zone_lowmem_reserve();
7356 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7357 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7358 * pagelist can have before it gets flushed back to buddy allocator.
7360 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7361 void __user *buffer, size_t *length, loff_t *ppos)
7364 int old_percpu_pagelist_fraction;
7367 mutex_lock(&pcp_batch_high_lock);
7368 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7370 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7371 if (!write || ret < 0)
7374 /* Sanity checking to avoid pcp imbalance */
7375 if (percpu_pagelist_fraction &&
7376 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7377 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7383 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7386 for_each_populated_zone(zone) {
7389 for_each_possible_cpu(cpu)
7390 pageset_set_high_and_batch(zone,
7391 per_cpu_ptr(zone->pageset, cpu));
7394 mutex_unlock(&pcp_batch_high_lock);
7399 int hashdist = HASHDIST_DEFAULT;
7401 static int __init set_hashdist(char *str)
7405 hashdist = simple_strtoul(str, &str, 0);
7408 __setup("hashdist=", set_hashdist);
7411 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7413 * Returns the number of pages that arch has reserved but
7414 * is not known to alloc_large_system_hash().
7416 static unsigned long __init arch_reserved_kernel_pages(void)
7423 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7424 * machines. As memory size is increased the scale is also increased but at
7425 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7426 * quadruples the scale is increased by one, which means the size of hash table
7427 * only doubles, instead of quadrupling as well.
7428 * Because 32-bit systems cannot have large physical memory, where this scaling
7429 * makes sense, it is disabled on such platforms.
7431 #if __BITS_PER_LONG > 32
7432 #define ADAPT_SCALE_BASE (64ul << 30)
7433 #define ADAPT_SCALE_SHIFT 2
7434 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7438 * allocate a large system hash table from bootmem
7439 * - it is assumed that the hash table must contain an exact power-of-2
7440 * quantity of entries
7441 * - limit is the number of hash buckets, not the total allocation size
7443 void *__init alloc_large_system_hash(const char *tablename,
7444 unsigned long bucketsize,
7445 unsigned long numentries,
7448 unsigned int *_hash_shift,
7449 unsigned int *_hash_mask,
7450 unsigned long low_limit,
7451 unsigned long high_limit)
7453 unsigned long long max = high_limit;
7454 unsigned long log2qty, size;
7458 /* allow the kernel cmdline to have a say */
7460 /* round applicable memory size up to nearest megabyte */
7461 numentries = nr_kernel_pages;
7462 numentries -= arch_reserved_kernel_pages();
7464 /* It isn't necessary when PAGE_SIZE >= 1MB */
7465 if (PAGE_SHIFT < 20)
7466 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7468 #if __BITS_PER_LONG > 32
7470 unsigned long adapt;
7472 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7473 adapt <<= ADAPT_SCALE_SHIFT)
7478 /* limit to 1 bucket per 2^scale bytes of low memory */
7479 if (scale > PAGE_SHIFT)
7480 numentries >>= (scale - PAGE_SHIFT);
7482 numentries <<= (PAGE_SHIFT - scale);
7484 /* Make sure we've got at least a 0-order allocation.. */
7485 if (unlikely(flags & HASH_SMALL)) {
7486 /* Makes no sense without HASH_EARLY */
7487 WARN_ON(!(flags & HASH_EARLY));
7488 if (!(numentries >> *_hash_shift)) {
7489 numentries = 1UL << *_hash_shift;
7490 BUG_ON(!numentries);
7492 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7493 numentries = PAGE_SIZE / bucketsize;
7495 numentries = roundup_pow_of_two(numentries);
7497 /* limit allocation size to 1/16 total memory by default */
7499 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7500 do_div(max, bucketsize);
7502 max = min(max, 0x80000000ULL);
7504 if (numentries < low_limit)
7505 numentries = low_limit;
7506 if (numentries > max)
7509 log2qty = ilog2(numentries);
7511 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7513 size = bucketsize << log2qty;
7514 if (flags & HASH_EARLY) {
7515 if (flags & HASH_ZERO)
7516 table = memblock_virt_alloc_nopanic(size, 0);
7518 table = memblock_virt_alloc_raw(size, 0);
7519 } else if (hashdist) {
7520 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7523 * If bucketsize is not a power-of-two, we may free
7524 * some pages at the end of hash table which
7525 * alloc_pages_exact() automatically does
7527 if (get_order(size) < MAX_ORDER) {
7528 table = alloc_pages_exact(size, gfp_flags);
7529 kmemleak_alloc(table, size, 1, gfp_flags);
7532 } while (!table && size > PAGE_SIZE && --log2qty);
7535 panic("Failed to allocate %s hash table\n", tablename);
7537 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7538 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7541 *_hash_shift = log2qty;
7543 *_hash_mask = (1 << log2qty) - 1;
7549 * This function checks whether pageblock includes unmovable pages or not.
7550 * If @count is not zero, it is okay to include less @count unmovable pages
7552 * PageLRU check without isolation or lru_lock could race so that
7553 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7554 * check without lock_page also may miss some movable non-lru pages at
7555 * race condition. So you can't expect this function should be exact.
7557 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7559 bool skip_hwpoisoned_pages)
7561 unsigned long pfn, iter, found;
7564 * For avoiding noise data, lru_add_drain_all() should be called
7565 * If ZONE_MOVABLE, the zone never contains unmovable pages
7567 if (zone_idx(zone) == ZONE_MOVABLE)
7571 * CMA allocations (alloc_contig_range) really need to mark isolate
7572 * CMA pageblocks even when they are not movable in fact so consider
7573 * them movable here.
7575 if (is_migrate_cma(migratetype) &&
7576 is_migrate_cma(get_pageblock_migratetype(page)))
7579 pfn = page_to_pfn(page);
7580 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7581 unsigned long check = pfn + iter;
7583 if (!pfn_valid_within(check))
7586 page = pfn_to_page(check);
7588 if (PageReserved(page))
7592 * Hugepages are not in LRU lists, but they're movable.
7593 * We need not scan over tail pages bacause we don't
7594 * handle each tail page individually in migration.
7596 if (PageHuge(page)) {
7597 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7602 * We can't use page_count without pin a page
7603 * because another CPU can free compound page.
7604 * This check already skips compound tails of THP
7605 * because their page->_refcount is zero at all time.
7607 if (!page_ref_count(page)) {
7608 if (PageBuddy(page))
7609 iter += (1 << page_order(page)) - 1;
7614 * The HWPoisoned page may be not in buddy system, and
7615 * page_count() is not 0.
7617 if (skip_hwpoisoned_pages && PageHWPoison(page))
7620 if (__PageMovable(page))
7626 * If there are RECLAIMABLE pages, we need to check
7627 * it. But now, memory offline itself doesn't call
7628 * shrink_node_slabs() and it still to be fixed.
7631 * If the page is not RAM, page_count()should be 0.
7632 * we don't need more check. This is an _used_ not-movable page.
7634 * The problematic thing here is PG_reserved pages. PG_reserved
7635 * is set to both of a memory hole page and a _used_ kernel
7644 bool is_pageblock_removable_nolock(struct page *page)
7650 * We have to be careful here because we are iterating over memory
7651 * sections which are not zone aware so we might end up outside of
7652 * the zone but still within the section.
7653 * We have to take care about the node as well. If the node is offline
7654 * its NODE_DATA will be NULL - see page_zone.
7656 if (!node_online(page_to_nid(page)))
7659 zone = page_zone(page);
7660 pfn = page_to_pfn(page);
7661 if (!zone_spans_pfn(zone, pfn))
7664 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7667 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7669 static unsigned long pfn_max_align_down(unsigned long pfn)
7671 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7672 pageblock_nr_pages) - 1);
7675 static unsigned long pfn_max_align_up(unsigned long pfn)
7677 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7678 pageblock_nr_pages));
7681 /* [start, end) must belong to a single zone. */
7682 static int __alloc_contig_migrate_range(struct compact_control *cc,
7683 unsigned long start, unsigned long end)
7685 /* This function is based on compact_zone() from compaction.c. */
7686 unsigned long nr_reclaimed;
7687 unsigned long pfn = start;
7688 unsigned int tries = 0;
7693 while (pfn < end || !list_empty(&cc->migratepages)) {
7694 if (fatal_signal_pending(current)) {
7699 if (list_empty(&cc->migratepages)) {
7700 cc->nr_migratepages = 0;
7701 pfn = isolate_migratepages_range(cc, pfn, end);
7707 } else if (++tries == 5) {
7708 ret = ret < 0 ? ret : -EBUSY;
7712 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7714 cc->nr_migratepages -= nr_reclaimed;
7716 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7717 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7720 putback_movable_pages(&cc->migratepages);
7727 * alloc_contig_range() -- tries to allocate given range of pages
7728 * @start: start PFN to allocate
7729 * @end: one-past-the-last PFN to allocate
7730 * @migratetype: migratetype of the underlaying pageblocks (either
7731 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7732 * in range must have the same migratetype and it must
7733 * be either of the two.
7734 * @gfp_mask: GFP mask to use during compaction
7736 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7737 * aligned, however it's the caller's responsibility to guarantee that
7738 * we are the only thread that changes migrate type of pageblocks the
7741 * The PFN range must belong to a single zone.
7743 * Returns zero on success or negative error code. On success all
7744 * pages which PFN is in [start, end) are allocated for the caller and
7745 * need to be freed with free_contig_range().
7747 int alloc_contig_range(unsigned long start, unsigned long end,
7748 unsigned migratetype, gfp_t gfp_mask)
7750 unsigned long outer_start, outer_end;
7754 struct compact_control cc = {
7755 .nr_migratepages = 0,
7757 .zone = page_zone(pfn_to_page(start)),
7758 .mode = MIGRATE_SYNC,
7759 .ignore_skip_hint = true,
7760 .no_set_skip_hint = true,
7761 .gfp_mask = current_gfp_context(gfp_mask),
7763 INIT_LIST_HEAD(&cc.migratepages);
7766 * What we do here is we mark all pageblocks in range as
7767 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7768 * have different sizes, and due to the way page allocator
7769 * work, we align the range to biggest of the two pages so
7770 * that page allocator won't try to merge buddies from
7771 * different pageblocks and change MIGRATE_ISOLATE to some
7772 * other migration type.
7774 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7775 * migrate the pages from an unaligned range (ie. pages that
7776 * we are interested in). This will put all the pages in
7777 * range back to page allocator as MIGRATE_ISOLATE.
7779 * When this is done, we take the pages in range from page
7780 * allocator removing them from the buddy system. This way
7781 * page allocator will never consider using them.
7783 * This lets us mark the pageblocks back as
7784 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7785 * aligned range but not in the unaligned, original range are
7786 * put back to page allocator so that buddy can use them.
7789 ret = start_isolate_page_range(pfn_max_align_down(start),
7790 pfn_max_align_up(end), migratetype,
7796 * In case of -EBUSY, we'd like to know which page causes problem.
7797 * So, just fall through. test_pages_isolated() has a tracepoint
7798 * which will report the busy page.
7800 * It is possible that busy pages could become available before
7801 * the call to test_pages_isolated, and the range will actually be
7802 * allocated. So, if we fall through be sure to clear ret so that
7803 * -EBUSY is not accidentally used or returned to caller.
7805 ret = __alloc_contig_migrate_range(&cc, start, end);
7806 if (ret && ret != -EBUSY)
7811 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7812 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7813 * more, all pages in [start, end) are free in page allocator.
7814 * What we are going to do is to allocate all pages from
7815 * [start, end) (that is remove them from page allocator).
7817 * The only problem is that pages at the beginning and at the
7818 * end of interesting range may be not aligned with pages that
7819 * page allocator holds, ie. they can be part of higher order
7820 * pages. Because of this, we reserve the bigger range and
7821 * once this is done free the pages we are not interested in.
7823 * We don't have to hold zone->lock here because the pages are
7824 * isolated thus they won't get removed from buddy.
7827 lru_add_drain_all();
7828 drain_all_pages(cc.zone);
7831 outer_start = start;
7832 while (!PageBuddy(pfn_to_page(outer_start))) {
7833 if (++order >= MAX_ORDER) {
7834 outer_start = start;
7837 outer_start &= ~0UL << order;
7840 if (outer_start != start) {
7841 order = page_order(pfn_to_page(outer_start));
7844 * outer_start page could be small order buddy page and
7845 * it doesn't include start page. Adjust outer_start
7846 * in this case to report failed page properly
7847 * on tracepoint in test_pages_isolated()
7849 if (outer_start + (1UL << order) <= start)
7850 outer_start = start;
7853 /* Make sure the range is really isolated. */
7854 if (test_pages_isolated(outer_start, end, false)) {
7855 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7856 __func__, outer_start, end);
7861 /* Grab isolated pages from freelists. */
7862 outer_end = isolate_freepages_range(&cc, outer_start, end);
7868 /* Free head and tail (if any) */
7869 if (start != outer_start)
7870 free_contig_range(outer_start, start - outer_start);
7871 if (end != outer_end)
7872 free_contig_range(end, outer_end - end);
7875 undo_isolate_page_range(pfn_max_align_down(start),
7876 pfn_max_align_up(end), migratetype);
7880 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7882 unsigned int count = 0;
7884 for (; nr_pages--; pfn++) {
7885 struct page *page = pfn_to_page(pfn);
7887 count += page_count(page) != 1;
7890 WARN(count != 0, "%d pages are still in use!\n", count);
7894 #ifdef CONFIG_MEMORY_HOTPLUG
7896 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7897 * page high values need to be recalulated.
7899 void __meminit zone_pcp_update(struct zone *zone)
7902 mutex_lock(&pcp_batch_high_lock);
7903 for_each_possible_cpu(cpu)
7904 pageset_set_high_and_batch(zone,
7905 per_cpu_ptr(zone->pageset, cpu));
7906 mutex_unlock(&pcp_batch_high_lock);
7910 void zone_pcp_reset(struct zone *zone)
7912 unsigned long flags;
7914 struct per_cpu_pageset *pset;
7916 /* avoid races with drain_pages() */
7917 local_irq_save(flags);
7918 if (zone->pageset != &boot_pageset) {
7919 for_each_online_cpu(cpu) {
7920 pset = per_cpu_ptr(zone->pageset, cpu);
7921 drain_zonestat(zone, pset);
7923 free_percpu(zone->pageset);
7924 zone->pageset = &boot_pageset;
7926 local_irq_restore(flags);
7929 #ifdef CONFIG_MEMORY_HOTREMOVE
7931 * All pages in the range must be in a single zone and isolated
7932 * before calling this.
7935 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7939 unsigned int order, i;
7941 unsigned long flags;
7942 /* find the first valid pfn */
7943 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7948 offline_mem_sections(pfn, end_pfn);
7949 zone = page_zone(pfn_to_page(pfn));
7950 spin_lock_irqsave(&zone->lock, flags);
7952 while (pfn < end_pfn) {
7953 if (!pfn_valid(pfn)) {
7957 page = pfn_to_page(pfn);
7959 * The HWPoisoned page may be not in buddy system, and
7960 * page_count() is not 0.
7962 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7964 SetPageReserved(page);
7968 BUG_ON(page_count(page));
7969 BUG_ON(!PageBuddy(page));
7970 order = page_order(page);
7971 #ifdef CONFIG_DEBUG_VM
7972 pr_info("remove from free list %lx %d %lx\n",
7973 pfn, 1 << order, end_pfn);
7975 list_del(&page->lru);
7976 rmv_page_order(page);
7977 zone->free_area[order].nr_free--;
7978 for (i = 0; i < (1 << order); i++)
7979 SetPageReserved((page+i));
7980 pfn += (1 << order);
7982 spin_unlock_irqrestore(&zone->lock, flags);
7986 bool is_free_buddy_page(struct page *page)
7988 struct zone *zone = page_zone(page);
7989 unsigned long pfn = page_to_pfn(page);
7990 unsigned long flags;
7993 spin_lock_irqsave(&zone->lock, flags);
7994 for (order = 0; order < MAX_ORDER; order++) {
7995 struct page *page_head = page - (pfn & ((1 << order) - 1));
7997 if (PageBuddy(page_head) && page_order(page_head) >= order)
8000 spin_unlock_irqrestore(&zone->lock, flags);
8002 return order < MAX_ORDER;