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/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 /* Protect totalram_pages and zone->managed_pages */
126 static DEFINE_SPINLOCK(managed_page_count_lock);
128 unsigned long totalram_pages __read_mostly;
129 unsigned long totalreserve_pages __read_mostly;
130 unsigned long totalcma_pages __read_mostly;
132 int percpu_pagelist_fraction;
133 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
143 static inline int get_pcppage_migratetype(struct page *page)
148 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
150 page->index = migratetype;
153 #ifdef CONFIG_PM_SLEEP
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with system_transition_mutex held
159 * (gfp_allowed_mask also should only be modified with system_transition_mutex
160 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
161 * with that modification).
164 static gfp_t saved_gfp_mask;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&system_transition_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(&system_transition_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] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
222 EXPORT_SYMBOL(totalram_pages);
224 static char * const zone_names[MAX_NR_ZONES] = {
225 #ifdef CONFIG_ZONE_DMA
228 #ifdef CONFIG_ZONE_DMA32
232 #ifdef CONFIG_HIGHMEM
236 #ifdef CONFIG_ZONE_DEVICE
241 char * const migratetype_names[MIGRATE_TYPES] = {
249 #ifdef CONFIG_MEMORY_ISOLATION
254 compound_page_dtor * const compound_page_dtors[] = {
257 #ifdef CONFIG_HUGETLB_PAGE
260 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
265 int min_free_kbytes = 1024;
266 int user_min_free_kbytes = -1;
267 int watermark_scale_factor = 10;
269 static unsigned long nr_kernel_pages __meminitdata;
270 static unsigned long nr_all_pages __meminitdata;
271 static unsigned long dma_reserve __meminitdata;
273 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
274 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
276 static unsigned long required_kernelcore __initdata;
277 static unsigned long required_kernelcore_percent __initdata;
278 static unsigned long required_movablecore __initdata;
279 static unsigned long required_movablecore_percent __initdata;
280 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
281 static bool mirrored_kernelcore __meminitdata;
283 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
285 EXPORT_SYMBOL(movable_zone);
286 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
289 int nr_node_ids __read_mostly = MAX_NUMNODES;
290 int nr_online_nodes __read_mostly = 1;
291 EXPORT_SYMBOL(nr_node_ids);
292 EXPORT_SYMBOL(nr_online_nodes);
295 int page_group_by_mobility_disabled __read_mostly;
297 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 /* Returns true if the struct page for the pfn is uninitialised */
299 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
301 int nid = early_pfn_to_nid(pfn);
303 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
310 * Returns false when the remaining initialisation should be deferred until
311 * later in the boot cycle when it can be parallelised.
313 static inline bool update_defer_init(pg_data_t *pgdat,
314 unsigned long pfn, unsigned long zone_end,
315 unsigned long *nr_initialised)
317 /* Always populate low zones for address-constrained allocations */
318 if (zone_end < pgdat_end_pfn(pgdat))
321 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
322 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
323 pgdat->first_deferred_pfn = pfn;
330 static inline bool early_page_uninitialised(unsigned long pfn)
335 static inline bool update_defer_init(pg_data_t *pgdat,
336 unsigned long pfn, unsigned long zone_end,
337 unsigned long *nr_initialised)
343 /* Return a pointer to the bitmap storing bits affecting a block of pages */
344 static inline unsigned long *get_pageblock_bitmap(struct page *page,
347 #ifdef CONFIG_SPARSEMEM
348 return __pfn_to_section(pfn)->pageblock_flags;
350 return page_zone(page)->pageblock_flags;
351 #endif /* CONFIG_SPARSEMEM */
354 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
356 #ifdef CONFIG_SPARSEMEM
357 pfn &= (PAGES_PER_SECTION-1);
358 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
360 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
361 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362 #endif /* CONFIG_SPARSEMEM */
366 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
367 * @page: The page within the block of interest
368 * @pfn: The target page frame number
369 * @end_bitidx: The last bit of interest to retrieve
370 * @mask: mask of bits that the caller is interested in
372 * Return: pageblock_bits flags
374 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
376 unsigned long end_bitidx,
379 unsigned long *bitmap;
380 unsigned long bitidx, word_bitidx;
383 bitmap = get_pageblock_bitmap(page, pfn);
384 bitidx = pfn_to_bitidx(page, pfn);
385 word_bitidx = bitidx / BITS_PER_LONG;
386 bitidx &= (BITS_PER_LONG-1);
388 word = bitmap[word_bitidx];
389 bitidx += end_bitidx;
390 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
393 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
394 unsigned long end_bitidx,
397 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
400 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
402 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
406 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
407 * @page: The page within the block of interest
408 * @flags: The flags to set
409 * @pfn: The target page frame number
410 * @end_bitidx: The last bit of interest
411 * @mask: mask of bits that the caller is interested in
413 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
415 unsigned long end_bitidx,
418 unsigned long *bitmap;
419 unsigned long bitidx, word_bitidx;
420 unsigned long old_word, word;
422 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
424 bitmap = get_pageblock_bitmap(page, pfn);
425 bitidx = pfn_to_bitidx(page, pfn);
426 word_bitidx = bitidx / BITS_PER_LONG;
427 bitidx &= (BITS_PER_LONG-1);
429 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
431 bitidx += end_bitidx;
432 mask <<= (BITS_PER_LONG - bitidx - 1);
433 flags <<= (BITS_PER_LONG - bitidx - 1);
435 word = READ_ONCE(bitmap[word_bitidx]);
437 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
438 if (word == old_word)
444 void set_pageblock_migratetype(struct page *page, int migratetype)
446 if (unlikely(page_group_by_mobility_disabled &&
447 migratetype < MIGRATE_PCPTYPES))
448 migratetype = MIGRATE_UNMOVABLE;
450 set_pageblock_flags_group(page, (unsigned long)migratetype,
451 PB_migrate, PB_migrate_end);
454 #ifdef CONFIG_DEBUG_VM
455 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
459 unsigned long pfn = page_to_pfn(page);
460 unsigned long sp, start_pfn;
463 seq = zone_span_seqbegin(zone);
464 start_pfn = zone->zone_start_pfn;
465 sp = zone->spanned_pages;
466 if (!zone_spans_pfn(zone, pfn))
468 } while (zone_span_seqretry(zone, seq));
471 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
472 pfn, zone_to_nid(zone), zone->name,
473 start_pfn, start_pfn + sp);
478 static int page_is_consistent(struct zone *zone, struct page *page)
480 if (!pfn_valid_within(page_to_pfn(page)))
482 if (zone != page_zone(page))
488 * Temporary debugging check for pages not lying within a given zone.
490 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
492 if (page_outside_zone_boundaries(zone, page))
494 if (!page_is_consistent(zone, page))
500 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
506 static void bad_page(struct page *page, const char *reason,
507 unsigned long bad_flags)
509 static unsigned long resume;
510 static unsigned long nr_shown;
511 static unsigned long nr_unshown;
514 * Allow a burst of 60 reports, then keep quiet for that minute;
515 * or allow a steady drip of one report per second.
517 if (nr_shown == 60) {
518 if (time_before(jiffies, resume)) {
524 "BUG: Bad page state: %lu messages suppressed\n",
531 resume = jiffies + 60 * HZ;
533 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
534 current->comm, page_to_pfn(page));
535 __dump_page(page, reason);
536 bad_flags &= page->flags;
538 pr_alert("bad because of flags: %#lx(%pGp)\n",
539 bad_flags, &bad_flags);
540 dump_page_owner(page);
545 /* Leave bad fields for debug, except PageBuddy could make trouble */
546 page_mapcount_reset(page); /* remove PageBuddy */
547 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
551 * Higher-order pages are called "compound pages". They are structured thusly:
553 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
555 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
556 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
558 * The first tail page's ->compound_dtor holds the offset in array of compound
559 * page destructors. See compound_page_dtors.
561 * The first tail page's ->compound_order holds the order of allocation.
562 * This usage means that zero-order pages may not be compound.
565 void free_compound_page(struct page *page)
567 __free_pages_ok(page, compound_order(page));
570 void prep_compound_page(struct page *page, unsigned int order)
573 int nr_pages = 1 << order;
575 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
576 set_compound_order(page, order);
578 for (i = 1; i < nr_pages; i++) {
579 struct page *p = page + i;
580 set_page_count(p, 0);
581 p->mapping = TAIL_MAPPING;
582 set_compound_head(p, page);
584 atomic_set(compound_mapcount_ptr(page), -1);
587 #ifdef CONFIG_DEBUG_PAGEALLOC
588 unsigned int _debug_guardpage_minorder;
589 bool _debug_pagealloc_enabled __read_mostly
590 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
591 EXPORT_SYMBOL(_debug_pagealloc_enabled);
592 bool _debug_guardpage_enabled __read_mostly;
594 static int __init early_debug_pagealloc(char *buf)
598 return kstrtobool(buf, &_debug_pagealloc_enabled);
600 early_param("debug_pagealloc", early_debug_pagealloc);
602 static bool need_debug_guardpage(void)
604 /* If we don't use debug_pagealloc, we don't need guard page */
605 if (!debug_pagealloc_enabled())
608 if (!debug_guardpage_minorder())
614 static void init_debug_guardpage(void)
616 if (!debug_pagealloc_enabled())
619 if (!debug_guardpage_minorder())
622 _debug_guardpage_enabled = true;
625 struct page_ext_operations debug_guardpage_ops = {
626 .need = need_debug_guardpage,
627 .init = init_debug_guardpage,
630 static int __init debug_guardpage_minorder_setup(char *buf)
634 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
635 pr_err("Bad debug_guardpage_minorder value\n");
638 _debug_guardpage_minorder = res;
639 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
642 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
644 static inline bool set_page_guard(struct zone *zone, struct page *page,
645 unsigned int order, int migratetype)
647 struct page_ext *page_ext;
649 if (!debug_guardpage_enabled())
652 if (order >= debug_guardpage_minorder())
655 page_ext = lookup_page_ext(page);
656 if (unlikely(!page_ext))
659 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
661 INIT_LIST_HEAD(&page->lru);
662 set_page_private(page, order);
663 /* Guard pages are not available for any usage */
664 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
669 static inline void clear_page_guard(struct zone *zone, struct page *page,
670 unsigned int order, int migratetype)
672 struct page_ext *page_ext;
674 if (!debug_guardpage_enabled())
677 page_ext = lookup_page_ext(page);
678 if (unlikely(!page_ext))
681 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
683 set_page_private(page, 0);
684 if (!is_migrate_isolate(migratetype))
685 __mod_zone_freepage_state(zone, (1 << order), migratetype);
688 struct page_ext_operations debug_guardpage_ops;
689 static inline bool set_page_guard(struct zone *zone, struct page *page,
690 unsigned int order, int migratetype) { return false; }
691 static inline void clear_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype) {}
695 static inline void set_page_order(struct page *page, unsigned int order)
697 set_page_private(page, order);
698 __SetPageBuddy(page);
701 static inline void rmv_page_order(struct page *page)
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
708 * This function checks whether a page is free && is the buddy
709 * we can coalesce a page and its buddy if
710 * (a) the buddy is not in a hole (check before calling!) &&
711 * (b) the buddy is in the buddy system &&
712 * (c) a page and its buddy have the same order &&
713 * (d) a page and its buddy are in the same zone.
715 * For recording whether a page is in the buddy system, we set PageBuddy.
716 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
718 * For recording page's order, we use page_private(page).
720 static inline int page_is_buddy(struct page *page, struct page *buddy,
723 if (page_is_guard(buddy) && page_order(buddy) == order) {
724 if (page_zone_id(page) != page_zone_id(buddy))
727 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
732 if (PageBuddy(buddy) && page_order(buddy) == order) {
734 * zone check is done late to avoid uselessly
735 * calculating zone/node ids for pages that could
738 if (page_zone_id(page) != page_zone_id(buddy))
741 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
749 * Freeing function for a buddy system allocator.
751 * The concept of a buddy system is to maintain direct-mapped table
752 * (containing bit values) for memory blocks of various "orders".
753 * The bottom level table contains the map for the smallest allocatable
754 * units of memory (here, pages), and each level above it describes
755 * pairs of units from the levels below, hence, "buddies".
756 * At a high level, all that happens here is marking the table entry
757 * at the bottom level available, and propagating the changes upward
758 * as necessary, plus some accounting needed to play nicely with other
759 * parts of the VM system.
760 * At each level, we keep a list of pages, which are heads of continuous
761 * free pages of length of (1 << order) and marked with PageBuddy.
762 * Page's order is recorded in page_private(page) field.
763 * So when we are allocating or freeing one, we can derive the state of the
764 * other. That is, if we allocate a small block, and both were
765 * free, the remainder of the region must be split into blocks.
766 * If a block is freed, and its buddy is also free, then this
767 * triggers coalescing into a block of larger size.
772 static inline void __free_one_page(struct page *page,
774 struct zone *zone, unsigned int order,
777 unsigned long combined_pfn;
778 unsigned long uninitialized_var(buddy_pfn);
780 unsigned int max_order;
782 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
784 VM_BUG_ON(!zone_is_initialized(zone));
785 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
787 VM_BUG_ON(migratetype == -1);
788 if (likely(!is_migrate_isolate(migratetype)))
789 __mod_zone_freepage_state(zone, 1 << order, migratetype);
791 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
792 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795 while (order < max_order - 1) {
796 buddy_pfn = __find_buddy_pfn(pfn, order);
797 buddy = page + (buddy_pfn - pfn);
799 if (!pfn_valid_within(buddy_pfn))
801 if (!page_is_buddy(page, buddy, order))
804 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
805 * merge with it and move up one order.
807 if (page_is_guard(buddy)) {
808 clear_page_guard(zone, buddy, order, migratetype);
810 list_del(&buddy->lru);
811 zone->free_area[order].nr_free--;
812 rmv_page_order(buddy);
814 combined_pfn = buddy_pfn & pfn;
815 page = page + (combined_pfn - pfn);
819 if (max_order < MAX_ORDER) {
820 /* If we are here, it means order is >= pageblock_order.
821 * We want to prevent merge between freepages on isolate
822 * pageblock and normal pageblock. Without this, pageblock
823 * isolation could cause incorrect freepage or CMA accounting.
825 * We don't want to hit this code for the more frequent
828 if (unlikely(has_isolate_pageblock(zone))) {
831 buddy_pfn = __find_buddy_pfn(pfn, order);
832 buddy = page + (buddy_pfn - pfn);
833 buddy_mt = get_pageblock_migratetype(buddy);
835 if (migratetype != buddy_mt
836 && (is_migrate_isolate(migratetype) ||
837 is_migrate_isolate(buddy_mt)))
841 goto continue_merging;
845 set_page_order(page, order);
848 * If this is not the largest possible page, check if the buddy
849 * of the next-highest order is free. If it is, it's possible
850 * that pages are being freed that will coalesce soon. In case,
851 * that is happening, add the free page to the tail of the list
852 * so it's less likely to be used soon and more likely to be merged
853 * as a higher order page
855 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
856 struct page *higher_page, *higher_buddy;
857 combined_pfn = buddy_pfn & pfn;
858 higher_page = page + (combined_pfn - pfn);
859 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
860 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
861 if (pfn_valid_within(buddy_pfn) &&
862 page_is_buddy(higher_page, higher_buddy, order + 1)) {
863 list_add_tail(&page->lru,
864 &zone->free_area[order].free_list[migratetype]);
869 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
871 zone->free_area[order].nr_free++;
875 * A bad page could be due to a number of fields. Instead of multiple branches,
876 * try and check multiple fields with one check. The caller must do a detailed
877 * check if necessary.
879 static inline bool page_expected_state(struct page *page,
880 unsigned long check_flags)
882 if (unlikely(atomic_read(&page->_mapcount) != -1))
885 if (unlikely((unsigned long)page->mapping |
886 page_ref_count(page) |
888 (unsigned long)page->mem_cgroup |
890 (page->flags & check_flags)))
896 static void free_pages_check_bad(struct page *page)
898 const char *bad_reason;
899 unsigned long bad_flags;
904 if (unlikely(atomic_read(&page->_mapcount) != -1))
905 bad_reason = "nonzero mapcount";
906 if (unlikely(page->mapping != NULL))
907 bad_reason = "non-NULL mapping";
908 if (unlikely(page_ref_count(page) != 0))
909 bad_reason = "nonzero _refcount";
910 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
911 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
912 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
915 if (unlikely(page->mem_cgroup))
916 bad_reason = "page still charged to cgroup";
918 bad_page(page, bad_reason, bad_flags);
921 static inline int free_pages_check(struct page *page)
923 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
926 /* Something has gone sideways, find it */
927 free_pages_check_bad(page);
931 static int free_tail_pages_check(struct page *head_page, struct page *page)
936 * We rely page->lru.next never has bit 0 set, unless the page
937 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
939 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
941 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
945 switch (page - head_page) {
947 /* the first tail page: ->mapping may be compound_mapcount() */
948 if (unlikely(compound_mapcount(page))) {
949 bad_page(page, "nonzero compound_mapcount", 0);
955 * the second tail page: ->mapping is
956 * deferred_list.next -- ignore value.
960 if (page->mapping != TAIL_MAPPING) {
961 bad_page(page, "corrupted mapping in tail page", 0);
966 if (unlikely(!PageTail(page))) {
967 bad_page(page, "PageTail not set", 0);
970 if (unlikely(compound_head(page) != head_page)) {
971 bad_page(page, "compound_head not consistent", 0);
976 page->mapping = NULL;
977 clear_compound_head(page);
981 static __always_inline bool free_pages_prepare(struct page *page,
982 unsigned int order, bool check_free)
986 VM_BUG_ON_PAGE(PageTail(page), page);
988 trace_mm_page_free(page, order);
991 * Check tail pages before head page information is cleared to
992 * avoid checking PageCompound for order-0 pages.
994 if (unlikely(order)) {
995 bool compound = PageCompound(page);
998 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1001 ClearPageDoubleMap(page);
1002 for (i = 1; i < (1 << order); i++) {
1004 bad += free_tail_pages_check(page, page + i);
1005 if (unlikely(free_pages_check(page + i))) {
1009 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 if (PageMappingFlags(page))
1013 page->mapping = NULL;
1014 if (memcg_kmem_enabled() && PageKmemcg(page))
1015 memcg_kmem_uncharge(page, order);
1017 bad += free_pages_check(page);
1021 page_cpupid_reset_last(page);
1022 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 reset_page_owner(page, order);
1025 if (!PageHighMem(page)) {
1026 debug_check_no_locks_freed(page_address(page),
1027 PAGE_SIZE << order);
1028 debug_check_no_obj_freed(page_address(page),
1029 PAGE_SIZE << order);
1031 arch_free_page(page, order);
1032 kernel_poison_pages(page, 1 << order, 0);
1033 kernel_map_pages(page, 1 << order, 0);
1034 kasan_free_pages(page, order);
1039 #ifdef CONFIG_DEBUG_VM
1040 static inline bool free_pcp_prepare(struct page *page)
1042 return free_pages_prepare(page, 0, true);
1045 static inline bool bulkfree_pcp_prepare(struct page *page)
1050 static bool free_pcp_prepare(struct page *page)
1052 return free_pages_prepare(page, 0, false);
1055 static bool bulkfree_pcp_prepare(struct page *page)
1057 return free_pages_check(page);
1059 #endif /* CONFIG_DEBUG_VM */
1061 static inline void prefetch_buddy(struct page *page)
1063 unsigned long pfn = page_to_pfn(page);
1064 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1065 struct page *buddy = page + (buddy_pfn - pfn);
1071 * Frees a number of pages from the PCP lists
1072 * Assumes all pages on list are in same zone, and of same order.
1073 * count is the number of pages to free.
1075 * If the zone was previously in an "all pages pinned" state then look to
1076 * see if this freeing clears that state.
1078 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1079 * pinned" detection logic.
1081 static void free_pcppages_bulk(struct zone *zone, int count,
1082 struct per_cpu_pages *pcp)
1084 int migratetype = 0;
1086 int prefetch_nr = 0;
1087 bool isolated_pageblocks;
1088 struct page *page, *tmp;
1092 struct list_head *list;
1095 * Remove pages from lists in a round-robin fashion. A
1096 * batch_free count is maintained that is incremented when an
1097 * empty list is encountered. This is so more pages are freed
1098 * off fuller lists instead of spinning excessively around empty
1103 if (++migratetype == MIGRATE_PCPTYPES)
1105 list = &pcp->lists[migratetype];
1106 } while (list_empty(list));
1108 /* This is the only non-empty list. Free them all. */
1109 if (batch_free == MIGRATE_PCPTYPES)
1113 page = list_last_entry(list, struct page, lru);
1114 /* must delete to avoid corrupting pcp list */
1115 list_del(&page->lru);
1118 if (bulkfree_pcp_prepare(page))
1121 list_add_tail(&page->lru, &head);
1124 * We are going to put the page back to the global
1125 * pool, prefetch its buddy to speed up later access
1126 * under zone->lock. It is believed the overhead of
1127 * an additional test and calculating buddy_pfn here
1128 * can be offset by reduced memory latency later. To
1129 * avoid excessive prefetching due to large count, only
1130 * prefetch buddy for the first pcp->batch nr of pages.
1132 if (prefetch_nr++ < pcp->batch)
1133 prefetch_buddy(page);
1134 } while (--count && --batch_free && !list_empty(list));
1137 spin_lock(&zone->lock);
1138 isolated_pageblocks = has_isolate_pageblock(zone);
1141 * Use safe version since after __free_one_page(),
1142 * page->lru.next will not point to original list.
1144 list_for_each_entry_safe(page, tmp, &head, lru) {
1145 int mt = get_pcppage_migratetype(page);
1146 /* MIGRATE_ISOLATE page should not go to pcplists */
1147 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1148 /* Pageblock could have been isolated meanwhile */
1149 if (unlikely(isolated_pageblocks))
1150 mt = get_pageblock_migratetype(page);
1152 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1153 trace_mm_page_pcpu_drain(page, 0, mt);
1155 spin_unlock(&zone->lock);
1158 static void free_one_page(struct zone *zone,
1159 struct page *page, unsigned long pfn,
1163 spin_lock(&zone->lock);
1164 if (unlikely(has_isolate_pageblock(zone) ||
1165 is_migrate_isolate(migratetype))) {
1166 migratetype = get_pfnblock_migratetype(page, pfn);
1168 __free_one_page(page, pfn, zone, order, migratetype);
1169 spin_unlock(&zone->lock);
1172 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1173 unsigned long zone, int nid)
1175 mm_zero_struct_page(page);
1176 set_page_links(page, zone, nid, pfn);
1177 init_page_count(page);
1178 page_mapcount_reset(page);
1179 page_cpupid_reset_last(page);
1181 INIT_LIST_HEAD(&page->lru);
1182 #ifdef WANT_PAGE_VIRTUAL
1183 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1184 if (!is_highmem_idx(zone))
1185 set_page_address(page, __va(pfn << PAGE_SHIFT));
1189 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1190 static void __meminit init_reserved_page(unsigned long pfn)
1195 if (!early_page_uninitialised(pfn))
1198 nid = early_pfn_to_nid(pfn);
1199 pgdat = NODE_DATA(nid);
1201 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1202 struct zone *zone = &pgdat->node_zones[zid];
1204 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1207 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1210 static inline void init_reserved_page(unsigned long pfn)
1213 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216 * Initialised pages do not have PageReserved set. This function is
1217 * called for each range allocated by the bootmem allocator and
1218 * marks the pages PageReserved. The remaining valid pages are later
1219 * sent to the buddy page allocator.
1221 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1223 unsigned long start_pfn = PFN_DOWN(start);
1224 unsigned long end_pfn = PFN_UP(end);
1226 for (; start_pfn < end_pfn; start_pfn++) {
1227 if (pfn_valid(start_pfn)) {
1228 struct page *page = pfn_to_page(start_pfn);
1230 init_reserved_page(start_pfn);
1232 /* Avoid false-positive PageTail() */
1233 INIT_LIST_HEAD(&page->lru);
1236 * no need for atomic set_bit because the struct
1237 * page is not visible yet so nobody should
1240 __SetPageReserved(page);
1245 static void __free_pages_ok(struct page *page, unsigned int order)
1247 unsigned long flags;
1249 unsigned long pfn = page_to_pfn(page);
1251 if (!free_pages_prepare(page, order, true))
1254 migratetype = get_pfnblock_migratetype(page, pfn);
1255 local_irq_save(flags);
1256 __count_vm_events(PGFREE, 1 << order);
1257 free_one_page(page_zone(page), page, pfn, order, migratetype);
1258 local_irq_restore(flags);
1261 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1263 unsigned int nr_pages = 1 << order;
1264 struct page *p = page;
1268 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1273 __ClearPageReserved(p);
1274 set_page_count(p, 0);
1276 page_zone(page)->managed_pages += nr_pages;
1277 set_page_refcounted(page);
1278 __free_pages(page, order);
1281 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1282 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1284 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1286 int __meminit early_pfn_to_nid(unsigned long pfn)
1288 static DEFINE_SPINLOCK(early_pfn_lock);
1291 spin_lock(&early_pfn_lock);
1292 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1294 nid = first_online_node;
1295 spin_unlock(&early_pfn_lock);
1301 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1302 static inline bool __meminit __maybe_unused
1303 meminit_pfn_in_nid(unsigned long pfn, int node,
1304 struct mminit_pfnnid_cache *state)
1308 nid = __early_pfn_to_nid(pfn, state);
1309 if (nid >= 0 && nid != node)
1314 /* Only safe to use early in boot when initialisation is single-threaded */
1315 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1317 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1322 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1326 static inline bool __meminit __maybe_unused
1327 meminit_pfn_in_nid(unsigned long pfn, int node,
1328 struct mminit_pfnnid_cache *state)
1335 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1338 if (early_page_uninitialised(pfn))
1340 return __free_pages_boot_core(page, order);
1344 * Check that the whole (or subset of) a pageblock given by the interval of
1345 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1346 * with the migration of free compaction scanner. The scanners then need to
1347 * use only pfn_valid_within() check for arches that allow holes within
1350 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1352 * It's possible on some configurations to have a setup like node0 node1 node0
1353 * i.e. it's possible that all pages within a zones range of pages do not
1354 * belong to a single zone. We assume that a border between node0 and node1
1355 * can occur within a single pageblock, but not a node0 node1 node0
1356 * interleaving within a single pageblock. It is therefore sufficient to check
1357 * the first and last page of a pageblock and avoid checking each individual
1358 * page in a pageblock.
1360 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1361 unsigned long end_pfn, struct zone *zone)
1363 struct page *start_page;
1364 struct page *end_page;
1366 /* end_pfn is one past the range we are checking */
1369 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1372 start_page = pfn_to_online_page(start_pfn);
1376 if (page_zone(start_page) != zone)
1379 end_page = pfn_to_page(end_pfn);
1381 /* This gives a shorter code than deriving page_zone(end_page) */
1382 if (page_zone_id(start_page) != page_zone_id(end_page))
1388 void set_zone_contiguous(struct zone *zone)
1390 unsigned long block_start_pfn = zone->zone_start_pfn;
1391 unsigned long block_end_pfn;
1393 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1394 for (; block_start_pfn < zone_end_pfn(zone);
1395 block_start_pfn = block_end_pfn,
1396 block_end_pfn += pageblock_nr_pages) {
1398 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1400 if (!__pageblock_pfn_to_page(block_start_pfn,
1401 block_end_pfn, zone))
1405 /* We confirm that there is no hole */
1406 zone->contiguous = true;
1409 void clear_zone_contiguous(struct zone *zone)
1411 zone->contiguous = false;
1414 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1415 static void __init deferred_free_range(unsigned long pfn,
1416 unsigned long nr_pages)
1424 page = pfn_to_page(pfn);
1426 /* Free a large naturally-aligned chunk if possible */
1427 if (nr_pages == pageblock_nr_pages &&
1428 (pfn & (pageblock_nr_pages - 1)) == 0) {
1429 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1430 __free_pages_boot_core(page, pageblock_order);
1434 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1435 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1436 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1437 __free_pages_boot_core(page, 0);
1441 /* Completion tracking for deferred_init_memmap() threads */
1442 static atomic_t pgdat_init_n_undone __initdata;
1443 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1445 static inline void __init pgdat_init_report_one_done(void)
1447 if (atomic_dec_and_test(&pgdat_init_n_undone))
1448 complete(&pgdat_init_all_done_comp);
1452 * Returns true if page needs to be initialized or freed to buddy allocator.
1454 * First we check if pfn is valid on architectures where it is possible to have
1455 * holes within pageblock_nr_pages. On systems where it is not possible, this
1456 * function is optimized out.
1458 * Then, we check if a current large page is valid by only checking the validity
1461 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1462 * within a node: a pfn is between start and end of a node, but does not belong
1463 * to this memory node.
1465 static inline bool __init
1466 deferred_pfn_valid(int nid, unsigned long pfn,
1467 struct mminit_pfnnid_cache *nid_init_state)
1469 if (!pfn_valid_within(pfn))
1471 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1473 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1479 * Free pages to buddy allocator. Try to free aligned pages in
1480 * pageblock_nr_pages sizes.
1482 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
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_free = 0;
1489 for (; pfn < end_pfn; pfn++) {
1490 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1491 deferred_free_range(pfn - nr_free, nr_free);
1493 } else if (!(pfn & nr_pgmask)) {
1494 deferred_free_range(pfn - nr_free, nr_free);
1496 touch_nmi_watchdog();
1501 /* Free the last block of pages to allocator */
1502 deferred_free_range(pfn - nr_free, nr_free);
1506 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1507 * by performing it only once every pageblock_nr_pages.
1508 * Return number of pages initialized.
1510 static unsigned long __init deferred_init_pages(int nid, int zid,
1512 unsigned long end_pfn)
1514 struct mminit_pfnnid_cache nid_init_state = { };
1515 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1516 unsigned long nr_pages = 0;
1517 struct page *page = NULL;
1519 for (; pfn < end_pfn; pfn++) {
1520 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1523 } else if (!page || !(pfn & nr_pgmask)) {
1524 page = pfn_to_page(pfn);
1525 touch_nmi_watchdog();
1529 __init_single_page(page, pfn, zid, nid);
1535 /* Initialise remaining memory on a node */
1536 static int __init deferred_init_memmap(void *data)
1538 pg_data_t *pgdat = data;
1539 int nid = pgdat->node_id;
1540 unsigned long start = jiffies;
1541 unsigned long nr_pages = 0;
1542 unsigned long spfn, epfn, first_init_pfn, flags;
1543 phys_addr_t spa, epa;
1546 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1549 /* Bind memory initialisation thread to a local node if possible */
1550 if (!cpumask_empty(cpumask))
1551 set_cpus_allowed_ptr(current, cpumask);
1553 pgdat_resize_lock(pgdat, &flags);
1554 first_init_pfn = pgdat->first_deferred_pfn;
1555 if (first_init_pfn == ULONG_MAX) {
1556 pgdat_resize_unlock(pgdat, &flags);
1557 pgdat_init_report_one_done();
1561 /* Sanity check boundaries */
1562 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1563 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1564 pgdat->first_deferred_pfn = ULONG_MAX;
1566 /* Only the highest zone is deferred so find it */
1567 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1568 zone = pgdat->node_zones + zid;
1569 if (first_init_pfn < zone_end_pfn(zone))
1572 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1575 * Initialize and free pages. We do it in two loops: first we initialize
1576 * struct page, than free to buddy allocator, because while we are
1577 * freeing pages we can access pages that are ahead (computing buddy
1578 * page in __free_one_page()).
1580 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1581 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1582 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1583 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1585 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1586 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1587 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1588 deferred_free_pages(nid, zid, spfn, epfn);
1590 pgdat_resize_unlock(pgdat, &flags);
1592 /* Sanity check that the next zone really is unpopulated */
1593 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1595 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1596 jiffies_to_msecs(jiffies - start));
1598 pgdat_init_report_one_done();
1603 * During boot we initialize deferred pages on-demand, as needed, but once
1604 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1605 * and we can permanently disable that path.
1607 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1610 * If this zone has deferred pages, try to grow it by initializing enough
1611 * deferred pages to satisfy the allocation specified by order, rounded up to
1612 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1613 * of SECTION_SIZE bytes by initializing struct pages in increments of
1614 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1616 * Return true when zone was grown, otherwise return false. We return true even
1617 * when we grow less than requested, to let the caller decide if there are
1618 * enough pages to satisfy the allocation.
1620 * Note: We use noinline because this function is needed only during boot, and
1621 * it is called from a __ref function _deferred_grow_zone. This way we are
1622 * making sure that it is not inlined into permanent text section.
1624 static noinline bool __init
1625 deferred_grow_zone(struct zone *zone, unsigned int order)
1627 int zid = zone_idx(zone);
1628 int nid = zone_to_nid(zone);
1629 pg_data_t *pgdat = NODE_DATA(nid);
1630 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1631 unsigned long nr_pages = 0;
1632 unsigned long first_init_pfn, spfn, epfn, t, flags;
1633 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1634 phys_addr_t spa, epa;
1637 /* Only the last zone may have deferred pages */
1638 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1641 pgdat_resize_lock(pgdat, &flags);
1644 * If deferred pages have been initialized while we were waiting for
1645 * the lock, return true, as the zone was grown. The caller will retry
1646 * this zone. We won't return to this function since the caller also
1647 * has this static branch.
1649 if (!static_branch_unlikely(&deferred_pages)) {
1650 pgdat_resize_unlock(pgdat, &flags);
1655 * If someone grew this zone while we were waiting for spinlock, return
1656 * true, as there might be enough pages already.
1658 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1659 pgdat_resize_unlock(pgdat, &flags);
1663 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1665 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1666 pgdat_resize_unlock(pgdat, &flags);
1670 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1671 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1672 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1674 while (spfn < epfn && nr_pages < nr_pages_needed) {
1675 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1676 first_deferred_pfn = min(t, epfn);
1677 nr_pages += deferred_init_pages(nid, zid, spfn,
1678 first_deferred_pfn);
1679 spfn = first_deferred_pfn;
1682 if (nr_pages >= nr_pages_needed)
1686 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1687 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1688 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1689 deferred_free_pages(nid, zid, spfn, epfn);
1691 if (first_deferred_pfn == epfn)
1694 pgdat->first_deferred_pfn = first_deferred_pfn;
1695 pgdat_resize_unlock(pgdat, &flags);
1697 return nr_pages > 0;
1701 * deferred_grow_zone() is __init, but it is called from
1702 * get_page_from_freelist() during early boot until deferred_pages permanently
1703 * disables this call. This is why we have refdata wrapper to avoid warning,
1704 * and to ensure that the function body gets unloaded.
1707 _deferred_grow_zone(struct zone *zone, unsigned int order)
1709 return deferred_grow_zone(zone, order);
1712 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1714 void __init page_alloc_init_late(void)
1718 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1721 /* There will be num_node_state(N_MEMORY) threads */
1722 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1723 for_each_node_state(nid, N_MEMORY) {
1724 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1727 /* Block until all are initialised */
1728 wait_for_completion(&pgdat_init_all_done_comp);
1731 * We initialized the rest of the deferred pages. Permanently disable
1732 * on-demand struct page initialization.
1734 static_branch_disable(&deferred_pages);
1736 /* Reinit limits that are based on free pages after the kernel is up */
1737 files_maxfiles_init();
1739 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1740 /* Discard memblock private memory */
1744 for_each_populated_zone(zone)
1745 set_zone_contiguous(zone);
1749 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1750 void __init init_cma_reserved_pageblock(struct page *page)
1752 unsigned i = pageblock_nr_pages;
1753 struct page *p = page;
1756 __ClearPageReserved(p);
1757 set_page_count(p, 0);
1760 set_pageblock_migratetype(page, MIGRATE_CMA);
1762 if (pageblock_order >= MAX_ORDER) {
1763 i = pageblock_nr_pages;
1766 set_page_refcounted(p);
1767 __free_pages(p, MAX_ORDER - 1);
1768 p += MAX_ORDER_NR_PAGES;
1769 } while (i -= MAX_ORDER_NR_PAGES);
1771 set_page_refcounted(page);
1772 __free_pages(page, pageblock_order);
1775 adjust_managed_page_count(page, pageblock_nr_pages);
1780 * The order of subdivision here is critical for the IO subsystem.
1781 * Please do not alter this order without good reasons and regression
1782 * testing. Specifically, as large blocks of memory are subdivided,
1783 * the order in which smaller blocks are delivered depends on the order
1784 * they're subdivided in this function. This is the primary factor
1785 * influencing the order in which pages are delivered to the IO
1786 * subsystem according to empirical testing, and this is also justified
1787 * by considering the behavior of a buddy system containing a single
1788 * large block of memory acted on by a series of small allocations.
1789 * This behavior is a critical factor in sglist merging's success.
1793 static inline void expand(struct zone *zone, struct page *page,
1794 int low, int high, struct free_area *area,
1797 unsigned long size = 1 << high;
1799 while (high > low) {
1803 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1806 * Mark as guard pages (or page), that will allow to
1807 * merge back to allocator when buddy will be freed.
1808 * Corresponding page table entries will not be touched,
1809 * pages will stay not present in virtual address space
1811 if (set_page_guard(zone, &page[size], high, migratetype))
1814 list_add(&page[size].lru, &area->free_list[migratetype]);
1816 set_page_order(&page[size], high);
1820 static void check_new_page_bad(struct page *page)
1822 const char *bad_reason = NULL;
1823 unsigned long bad_flags = 0;
1825 if (unlikely(atomic_read(&page->_mapcount) != -1))
1826 bad_reason = "nonzero mapcount";
1827 if (unlikely(page->mapping != NULL))
1828 bad_reason = "non-NULL mapping";
1829 if (unlikely(page_ref_count(page) != 0))
1830 bad_reason = "nonzero _count";
1831 if (unlikely(page->flags & __PG_HWPOISON)) {
1832 bad_reason = "HWPoisoned (hardware-corrupted)";
1833 bad_flags = __PG_HWPOISON;
1834 /* Don't complain about hwpoisoned pages */
1835 page_mapcount_reset(page); /* remove PageBuddy */
1838 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1839 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1840 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1843 if (unlikely(page->mem_cgroup))
1844 bad_reason = "page still charged to cgroup";
1846 bad_page(page, bad_reason, bad_flags);
1850 * This page is about to be returned from the page allocator
1852 static inline int check_new_page(struct page *page)
1854 if (likely(page_expected_state(page,
1855 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1858 check_new_page_bad(page);
1862 static inline bool free_pages_prezeroed(void)
1864 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1865 page_poisoning_enabled();
1868 #ifdef CONFIG_DEBUG_VM
1869 static bool check_pcp_refill(struct page *page)
1874 static bool check_new_pcp(struct page *page)
1876 return check_new_page(page);
1879 static bool check_pcp_refill(struct page *page)
1881 return check_new_page(page);
1883 static bool check_new_pcp(struct page *page)
1887 #endif /* CONFIG_DEBUG_VM */
1889 static bool check_new_pages(struct page *page, unsigned int order)
1892 for (i = 0; i < (1 << order); i++) {
1893 struct page *p = page + i;
1895 if (unlikely(check_new_page(p)))
1902 inline void post_alloc_hook(struct page *page, unsigned int order,
1905 set_page_private(page, 0);
1906 set_page_refcounted(page);
1908 arch_alloc_page(page, order);
1909 kernel_map_pages(page, 1 << order, 1);
1910 kernel_poison_pages(page, 1 << order, 1);
1911 kasan_alloc_pages(page, order);
1912 set_page_owner(page, order, gfp_flags);
1915 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1916 unsigned int alloc_flags)
1920 post_alloc_hook(page, order, gfp_flags);
1922 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1923 for (i = 0; i < (1 << order); i++)
1924 clear_highpage(page + i);
1926 if (order && (gfp_flags & __GFP_COMP))
1927 prep_compound_page(page, order);
1930 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1931 * allocate the page. The expectation is that the caller is taking
1932 * steps that will free more memory. The caller should avoid the page
1933 * being used for !PFMEMALLOC purposes.
1935 if (alloc_flags & ALLOC_NO_WATERMARKS)
1936 set_page_pfmemalloc(page);
1938 clear_page_pfmemalloc(page);
1942 * Go through the free lists for the given migratetype and remove
1943 * the smallest available page from the freelists
1945 static __always_inline
1946 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1949 unsigned int current_order;
1950 struct free_area *area;
1953 /* Find a page of the appropriate size in the preferred list */
1954 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1955 area = &(zone->free_area[current_order]);
1956 page = list_first_entry_or_null(&area->free_list[migratetype],
1960 list_del(&page->lru);
1961 rmv_page_order(page);
1963 expand(zone, page, order, current_order, area, migratetype);
1964 set_pcppage_migratetype(page, migratetype);
1973 * This array describes the order lists are fallen back to when
1974 * the free lists for the desirable migrate type are depleted
1976 static int fallbacks[MIGRATE_TYPES][4] = {
1977 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1978 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1979 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1981 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1983 #ifdef CONFIG_MEMORY_ISOLATION
1984 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1989 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1992 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1995 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1996 unsigned int order) { return NULL; }
2000 * Move the free pages in a range to the free lists of the requested type.
2001 * Note that start_page and end_pages are not aligned on a pageblock
2002 * boundary. If alignment is required, use move_freepages_block()
2004 static int move_freepages(struct zone *zone,
2005 struct page *start_page, struct page *end_page,
2006 int migratetype, int *num_movable)
2010 int pages_moved = 0;
2012 #ifndef CONFIG_HOLES_IN_ZONE
2014 * page_zone is not safe to call in this context when
2015 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2016 * anyway as we check zone boundaries in move_freepages_block().
2017 * Remove at a later date when no bug reports exist related to
2018 * grouping pages by mobility
2020 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2021 pfn_valid(page_to_pfn(end_page)) &&
2022 page_zone(start_page) != page_zone(end_page));
2028 for (page = start_page; page <= end_page;) {
2029 if (!pfn_valid_within(page_to_pfn(page))) {
2034 /* Make sure we are not inadvertently changing nodes */
2035 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2037 if (!PageBuddy(page)) {
2039 * We assume that pages that could be isolated for
2040 * migration are movable. But we don't actually try
2041 * isolating, as that would be expensive.
2044 (PageLRU(page) || __PageMovable(page)))
2051 order = page_order(page);
2052 list_move(&page->lru,
2053 &zone->free_area[order].free_list[migratetype]);
2055 pages_moved += 1 << order;
2061 int move_freepages_block(struct zone *zone, struct page *page,
2062 int migratetype, int *num_movable)
2064 unsigned long start_pfn, end_pfn;
2065 struct page *start_page, *end_page;
2067 start_pfn = page_to_pfn(page);
2068 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2069 start_page = pfn_to_page(start_pfn);
2070 end_page = start_page + pageblock_nr_pages - 1;
2071 end_pfn = start_pfn + pageblock_nr_pages - 1;
2073 /* Do not cross zone boundaries */
2074 if (!zone_spans_pfn(zone, start_pfn))
2076 if (!zone_spans_pfn(zone, end_pfn))
2079 return move_freepages(zone, start_page, end_page, migratetype,
2083 static void change_pageblock_range(struct page *pageblock_page,
2084 int start_order, int migratetype)
2086 int nr_pageblocks = 1 << (start_order - pageblock_order);
2088 while (nr_pageblocks--) {
2089 set_pageblock_migratetype(pageblock_page, migratetype);
2090 pageblock_page += pageblock_nr_pages;
2095 * When we are falling back to another migratetype during allocation, try to
2096 * steal extra free pages from the same pageblocks to satisfy further
2097 * allocations, instead of polluting multiple pageblocks.
2099 * If we are stealing a relatively large buddy page, it is likely there will
2100 * be more free pages in the pageblock, so try to steal them all. For
2101 * reclaimable and unmovable allocations, we steal regardless of page size,
2102 * as fragmentation caused by those allocations polluting movable pageblocks
2103 * is worse than movable allocations stealing from unmovable and reclaimable
2106 static bool can_steal_fallback(unsigned int order, int start_mt)
2109 * Leaving this order check is intended, although there is
2110 * relaxed order check in next check. The reason is that
2111 * we can actually steal whole pageblock if this condition met,
2112 * but, below check doesn't guarantee it and that is just heuristic
2113 * so could be changed anytime.
2115 if (order >= pageblock_order)
2118 if (order >= pageblock_order / 2 ||
2119 start_mt == MIGRATE_RECLAIMABLE ||
2120 start_mt == MIGRATE_UNMOVABLE ||
2121 page_group_by_mobility_disabled)
2128 * This function implements actual steal behaviour. If order is large enough,
2129 * we can steal whole pageblock. If not, we first move freepages in this
2130 * pageblock to our migratetype and determine how many already-allocated pages
2131 * are there in the pageblock with a compatible migratetype. If at least half
2132 * of pages are free or compatible, we can change migratetype of the pageblock
2133 * itself, so pages freed in the future will be put on the correct free list.
2135 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2136 int start_type, bool whole_block)
2138 unsigned int current_order = page_order(page);
2139 struct free_area *area;
2140 int free_pages, movable_pages, alike_pages;
2143 old_block_type = get_pageblock_migratetype(page);
2146 * This can happen due to races and we want to prevent broken
2147 * highatomic accounting.
2149 if (is_migrate_highatomic(old_block_type))
2152 /* Take ownership for orders >= pageblock_order */
2153 if (current_order >= pageblock_order) {
2154 change_pageblock_range(page, current_order, start_type);
2158 /* We are not allowed to try stealing from the whole block */
2162 free_pages = move_freepages_block(zone, page, start_type,
2165 * Determine how many pages are compatible with our allocation.
2166 * For movable allocation, it's the number of movable pages which
2167 * we just obtained. For other types it's a bit more tricky.
2169 if (start_type == MIGRATE_MOVABLE) {
2170 alike_pages = movable_pages;
2173 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2174 * to MOVABLE pageblock, consider all non-movable pages as
2175 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2176 * vice versa, be conservative since we can't distinguish the
2177 * exact migratetype of non-movable pages.
2179 if (old_block_type == MIGRATE_MOVABLE)
2180 alike_pages = pageblock_nr_pages
2181 - (free_pages + movable_pages);
2186 /* moving whole block can fail due to zone boundary conditions */
2191 * If a sufficient number of pages in the block are either free or of
2192 * comparable migratability as our allocation, claim the whole block.
2194 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2195 page_group_by_mobility_disabled)
2196 set_pageblock_migratetype(page, start_type);
2201 area = &zone->free_area[current_order];
2202 list_move(&page->lru, &area->free_list[start_type]);
2206 * Check whether there is a suitable fallback freepage with requested order.
2207 * If only_stealable is true, this function returns fallback_mt only if
2208 * we can steal other freepages all together. This would help to reduce
2209 * fragmentation due to mixed migratetype pages in one pageblock.
2211 int find_suitable_fallback(struct free_area *area, unsigned int order,
2212 int migratetype, bool only_stealable, bool *can_steal)
2217 if (area->nr_free == 0)
2222 fallback_mt = fallbacks[migratetype][i];
2223 if (fallback_mt == MIGRATE_TYPES)
2226 if (list_empty(&area->free_list[fallback_mt]))
2229 if (can_steal_fallback(order, migratetype))
2232 if (!only_stealable)
2243 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2244 * there are no empty page blocks that contain a page with a suitable order
2246 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2247 unsigned int alloc_order)
2250 unsigned long max_managed, flags;
2253 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2254 * Check is race-prone but harmless.
2256 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2257 if (zone->nr_reserved_highatomic >= max_managed)
2260 spin_lock_irqsave(&zone->lock, flags);
2262 /* Recheck the nr_reserved_highatomic limit under the lock */
2263 if (zone->nr_reserved_highatomic >= max_managed)
2267 mt = get_pageblock_migratetype(page);
2268 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2269 && !is_migrate_cma(mt)) {
2270 zone->nr_reserved_highatomic += pageblock_nr_pages;
2271 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2272 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2276 spin_unlock_irqrestore(&zone->lock, flags);
2280 * Used when an allocation is about to fail under memory pressure. This
2281 * potentially hurts the reliability of high-order allocations when under
2282 * intense memory pressure but failed atomic allocations should be easier
2283 * to recover from than an OOM.
2285 * If @force is true, try to unreserve a pageblock even though highatomic
2286 * pageblock is exhausted.
2288 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2291 struct zonelist *zonelist = ac->zonelist;
2292 unsigned long flags;
2299 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2302 * Preserve at least one pageblock unless memory pressure
2305 if (!force && zone->nr_reserved_highatomic <=
2309 spin_lock_irqsave(&zone->lock, flags);
2310 for (order = 0; order < MAX_ORDER; order++) {
2311 struct free_area *area = &(zone->free_area[order]);
2313 page = list_first_entry_or_null(
2314 &area->free_list[MIGRATE_HIGHATOMIC],
2320 * In page freeing path, migratetype change is racy so
2321 * we can counter several free pages in a pageblock
2322 * in this loop althoug we changed the pageblock type
2323 * from highatomic to ac->migratetype. So we should
2324 * adjust the count once.
2326 if (is_migrate_highatomic_page(page)) {
2328 * It should never happen but changes to
2329 * locking could inadvertently allow a per-cpu
2330 * drain to add pages to MIGRATE_HIGHATOMIC
2331 * while unreserving so be safe and watch for
2334 zone->nr_reserved_highatomic -= min(
2336 zone->nr_reserved_highatomic);
2340 * Convert to ac->migratetype and avoid the normal
2341 * pageblock stealing heuristics. Minimally, the caller
2342 * is doing the work and needs the pages. More
2343 * importantly, if the block was always converted to
2344 * MIGRATE_UNMOVABLE or another type then the number
2345 * of pageblocks that cannot be completely freed
2348 set_pageblock_migratetype(page, ac->migratetype);
2349 ret = move_freepages_block(zone, page, ac->migratetype,
2352 spin_unlock_irqrestore(&zone->lock, flags);
2356 spin_unlock_irqrestore(&zone->lock, flags);
2363 * Try finding a free buddy page on the fallback list and put it on the free
2364 * list of requested migratetype, possibly along with other pages from the same
2365 * block, depending on fragmentation avoidance heuristics. Returns true if
2366 * fallback was found so that __rmqueue_smallest() can grab it.
2368 * The use of signed ints for order and current_order is a deliberate
2369 * deviation from the rest of this file, to make the for loop
2370 * condition simpler.
2372 static __always_inline bool
2373 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2375 struct free_area *area;
2382 * Find the largest available free page in the other list. This roughly
2383 * approximates finding the pageblock with the most free pages, which
2384 * would be too costly to do exactly.
2386 for (current_order = MAX_ORDER - 1; current_order >= order;
2388 area = &(zone->free_area[current_order]);
2389 fallback_mt = find_suitable_fallback(area, current_order,
2390 start_migratetype, false, &can_steal);
2391 if (fallback_mt == -1)
2395 * We cannot steal all free pages from the pageblock and the
2396 * requested migratetype is movable. In that case it's better to
2397 * steal and split the smallest available page instead of the
2398 * largest available page, because even if the next movable
2399 * allocation falls back into a different pageblock than this
2400 * one, it won't cause permanent fragmentation.
2402 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2403 && current_order > order)
2412 for (current_order = order; current_order < MAX_ORDER;
2414 area = &(zone->free_area[current_order]);
2415 fallback_mt = find_suitable_fallback(area, current_order,
2416 start_migratetype, false, &can_steal);
2417 if (fallback_mt != -1)
2422 * This should not happen - we already found a suitable fallback
2423 * when looking for the largest page.
2425 VM_BUG_ON(current_order == MAX_ORDER);
2428 page = list_first_entry(&area->free_list[fallback_mt],
2431 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2433 trace_mm_page_alloc_extfrag(page, order, current_order,
2434 start_migratetype, fallback_mt);
2441 * Do the hard work of removing an element from the buddy allocator.
2442 * Call me with the zone->lock already held.
2444 static __always_inline struct page *
2445 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2450 page = __rmqueue_smallest(zone, order, migratetype);
2451 if (unlikely(!page)) {
2452 if (migratetype == MIGRATE_MOVABLE)
2453 page = __rmqueue_cma_fallback(zone, order);
2455 if (!page && __rmqueue_fallback(zone, order, migratetype))
2459 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2464 * Obtain a specified number of elements from the buddy allocator, all under
2465 * a single hold of the lock, for efficiency. Add them to the supplied list.
2466 * Returns the number of new pages which were placed at *list.
2468 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2469 unsigned long count, struct list_head *list,
2474 spin_lock(&zone->lock);
2475 for (i = 0; i < count; ++i) {
2476 struct page *page = __rmqueue(zone, order, migratetype);
2477 if (unlikely(page == NULL))
2480 if (unlikely(check_pcp_refill(page)))
2484 * Split buddy pages returned by expand() are received here in
2485 * physical page order. The page is added to the tail of
2486 * caller's list. From the callers perspective, the linked list
2487 * is ordered by page number under some conditions. This is
2488 * useful for IO devices that can forward direction from the
2489 * head, thus also in the physical page order. This is useful
2490 * for IO devices that can merge IO requests if the physical
2491 * pages are ordered properly.
2493 list_add_tail(&page->lru, list);
2495 if (is_migrate_cma(get_pcppage_migratetype(page)))
2496 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2501 * i pages were removed from the buddy list even if some leak due
2502 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2503 * on i. Do not confuse with 'alloced' which is the number of
2504 * pages added to the pcp list.
2506 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2507 spin_unlock(&zone->lock);
2513 * Called from the vmstat counter updater to drain pagesets of this
2514 * currently executing processor on remote nodes after they have
2517 * Note that this function must be called with the thread pinned to
2518 * a single processor.
2520 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2522 unsigned long flags;
2523 int to_drain, batch;
2525 local_irq_save(flags);
2526 batch = READ_ONCE(pcp->batch);
2527 to_drain = min(pcp->count, batch);
2529 free_pcppages_bulk(zone, to_drain, pcp);
2530 local_irq_restore(flags);
2535 * Drain pcplists of the indicated processor and zone.
2537 * The processor must either be the current processor and the
2538 * thread pinned to the current processor or a processor that
2541 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2543 unsigned long flags;
2544 struct per_cpu_pageset *pset;
2545 struct per_cpu_pages *pcp;
2547 local_irq_save(flags);
2548 pset = per_cpu_ptr(zone->pageset, cpu);
2552 free_pcppages_bulk(zone, pcp->count, pcp);
2553 local_irq_restore(flags);
2557 * Drain pcplists of all zones on the indicated processor.
2559 * The processor must either be the current processor and the
2560 * thread pinned to the current processor or a processor that
2563 static void drain_pages(unsigned int cpu)
2567 for_each_populated_zone(zone) {
2568 drain_pages_zone(cpu, zone);
2573 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2575 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2576 * the single zone's pages.
2578 void drain_local_pages(struct zone *zone)
2580 int cpu = smp_processor_id();
2583 drain_pages_zone(cpu, zone);
2588 static void drain_local_pages_wq(struct work_struct *work)
2591 * drain_all_pages doesn't use proper cpu hotplug protection so
2592 * we can race with cpu offline when the WQ can move this from
2593 * a cpu pinned worker to an unbound one. We can operate on a different
2594 * cpu which is allright but we also have to make sure to not move to
2598 drain_local_pages(NULL);
2603 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2605 * When zone parameter is non-NULL, spill just the single zone's pages.
2607 * Note that this can be extremely slow as the draining happens in a workqueue.
2609 void drain_all_pages(struct zone *zone)
2614 * Allocate in the BSS so we wont require allocation in
2615 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2617 static cpumask_t cpus_with_pcps;
2620 * Make sure nobody triggers this path before mm_percpu_wq is fully
2623 if (WARN_ON_ONCE(!mm_percpu_wq))
2627 * Do not drain if one is already in progress unless it's specific to
2628 * a zone. Such callers are primarily CMA and memory hotplug and need
2629 * the drain to be complete when the call returns.
2631 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2634 mutex_lock(&pcpu_drain_mutex);
2638 * We don't care about racing with CPU hotplug event
2639 * as offline notification will cause the notified
2640 * cpu to drain that CPU pcps and on_each_cpu_mask
2641 * disables preemption as part of its processing
2643 for_each_online_cpu(cpu) {
2644 struct per_cpu_pageset *pcp;
2646 bool has_pcps = false;
2649 pcp = per_cpu_ptr(zone->pageset, cpu);
2653 for_each_populated_zone(z) {
2654 pcp = per_cpu_ptr(z->pageset, cpu);
2655 if (pcp->pcp.count) {
2663 cpumask_set_cpu(cpu, &cpus_with_pcps);
2665 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2668 for_each_cpu(cpu, &cpus_with_pcps) {
2669 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2670 INIT_WORK(work, drain_local_pages_wq);
2671 queue_work_on(cpu, mm_percpu_wq, work);
2673 for_each_cpu(cpu, &cpus_with_pcps)
2674 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2676 mutex_unlock(&pcpu_drain_mutex);
2679 #ifdef CONFIG_HIBERNATION
2682 * Touch the watchdog for every WD_PAGE_COUNT pages.
2684 #define WD_PAGE_COUNT (128*1024)
2686 void mark_free_pages(struct zone *zone)
2688 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2689 unsigned long flags;
2690 unsigned int order, t;
2693 if (zone_is_empty(zone))
2696 spin_lock_irqsave(&zone->lock, flags);
2698 max_zone_pfn = zone_end_pfn(zone);
2699 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2700 if (pfn_valid(pfn)) {
2701 page = pfn_to_page(pfn);
2703 if (!--page_count) {
2704 touch_nmi_watchdog();
2705 page_count = WD_PAGE_COUNT;
2708 if (page_zone(page) != zone)
2711 if (!swsusp_page_is_forbidden(page))
2712 swsusp_unset_page_free(page);
2715 for_each_migratetype_order(order, t) {
2716 list_for_each_entry(page,
2717 &zone->free_area[order].free_list[t], lru) {
2720 pfn = page_to_pfn(page);
2721 for (i = 0; i < (1UL << order); i++) {
2722 if (!--page_count) {
2723 touch_nmi_watchdog();
2724 page_count = WD_PAGE_COUNT;
2726 swsusp_set_page_free(pfn_to_page(pfn + i));
2730 spin_unlock_irqrestore(&zone->lock, flags);
2732 #endif /* CONFIG_PM */
2734 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2738 if (!free_pcp_prepare(page))
2741 migratetype = get_pfnblock_migratetype(page, pfn);
2742 set_pcppage_migratetype(page, migratetype);
2746 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2748 struct zone *zone = page_zone(page);
2749 struct per_cpu_pages *pcp;
2752 migratetype = get_pcppage_migratetype(page);
2753 __count_vm_event(PGFREE);
2756 * We only track unmovable, reclaimable and movable on pcp lists.
2757 * Free ISOLATE pages back to the allocator because they are being
2758 * offlined but treat HIGHATOMIC as movable pages so we can get those
2759 * areas back if necessary. Otherwise, we may have to free
2760 * excessively into the page allocator
2762 if (migratetype >= MIGRATE_PCPTYPES) {
2763 if (unlikely(is_migrate_isolate(migratetype))) {
2764 free_one_page(zone, page, pfn, 0, migratetype);
2767 migratetype = MIGRATE_MOVABLE;
2770 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2771 list_add(&page->lru, &pcp->lists[migratetype]);
2773 if (pcp->count >= pcp->high) {
2774 unsigned long batch = READ_ONCE(pcp->batch);
2775 free_pcppages_bulk(zone, batch, pcp);
2780 * Free a 0-order page
2782 void free_unref_page(struct page *page)
2784 unsigned long flags;
2785 unsigned long pfn = page_to_pfn(page);
2787 if (!free_unref_page_prepare(page, pfn))
2790 local_irq_save(flags);
2791 free_unref_page_commit(page, pfn);
2792 local_irq_restore(flags);
2796 * Free a list of 0-order pages
2798 void free_unref_page_list(struct list_head *list)
2800 struct page *page, *next;
2801 unsigned long flags, pfn;
2802 int batch_count = 0;
2804 /* Prepare pages for freeing */
2805 list_for_each_entry_safe(page, next, list, lru) {
2806 pfn = page_to_pfn(page);
2807 if (!free_unref_page_prepare(page, pfn))
2808 list_del(&page->lru);
2809 set_page_private(page, pfn);
2812 local_irq_save(flags);
2813 list_for_each_entry_safe(page, next, list, lru) {
2814 unsigned long pfn = page_private(page);
2816 set_page_private(page, 0);
2817 trace_mm_page_free_batched(page);
2818 free_unref_page_commit(page, pfn);
2821 * Guard against excessive IRQ disabled times when we get
2822 * a large list of pages to free.
2824 if (++batch_count == SWAP_CLUSTER_MAX) {
2825 local_irq_restore(flags);
2827 local_irq_save(flags);
2830 local_irq_restore(flags);
2834 * split_page takes a non-compound higher-order page, and splits it into
2835 * n (1<<order) sub-pages: page[0..n]
2836 * Each sub-page must be freed individually.
2838 * Note: this is probably too low level an operation for use in drivers.
2839 * Please consult with lkml before using this in your driver.
2841 void split_page(struct page *page, unsigned int order)
2845 VM_BUG_ON_PAGE(PageCompound(page), page);
2846 VM_BUG_ON_PAGE(!page_count(page), page);
2848 for (i = 1; i < (1 << order); i++)
2849 set_page_refcounted(page + i);
2850 split_page_owner(page, order);
2852 EXPORT_SYMBOL_GPL(split_page);
2854 int __isolate_free_page(struct page *page, unsigned int order)
2856 unsigned long watermark;
2860 BUG_ON(!PageBuddy(page));
2862 zone = page_zone(page);
2863 mt = get_pageblock_migratetype(page);
2865 if (!is_migrate_isolate(mt)) {
2867 * Obey watermarks as if the page was being allocated. We can
2868 * emulate a high-order watermark check with a raised order-0
2869 * watermark, because we already know our high-order page
2872 watermark = min_wmark_pages(zone) + (1UL << order);
2873 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2876 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2879 /* Remove page from free list */
2880 list_del(&page->lru);
2881 zone->free_area[order].nr_free--;
2882 rmv_page_order(page);
2885 * Set the pageblock if the isolated page is at least half of a
2888 if (order >= pageblock_order - 1) {
2889 struct page *endpage = page + (1 << order) - 1;
2890 for (; page < endpage; page += pageblock_nr_pages) {
2891 int mt = get_pageblock_migratetype(page);
2892 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2893 && !is_migrate_highatomic(mt))
2894 set_pageblock_migratetype(page,
2900 return 1UL << order;
2904 * Update NUMA hit/miss statistics
2906 * Must be called with interrupts disabled.
2908 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2911 enum numa_stat_item local_stat = NUMA_LOCAL;
2913 /* skip numa counters update if numa stats is disabled */
2914 if (!static_branch_likely(&vm_numa_stat_key))
2917 if (zone_to_nid(z) != numa_node_id())
2918 local_stat = NUMA_OTHER;
2920 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2921 __inc_numa_state(z, NUMA_HIT);
2923 __inc_numa_state(z, NUMA_MISS);
2924 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2926 __inc_numa_state(z, local_stat);
2930 /* Remove page from the per-cpu list, caller must protect the list */
2931 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2932 struct per_cpu_pages *pcp,
2933 struct list_head *list)
2938 if (list_empty(list)) {
2939 pcp->count += rmqueue_bulk(zone, 0,
2942 if (unlikely(list_empty(list)))
2946 page = list_first_entry(list, struct page, lru);
2947 list_del(&page->lru);
2949 } while (check_new_pcp(page));
2954 /* Lock and remove page from the per-cpu list */
2955 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2956 struct zone *zone, unsigned int order,
2957 gfp_t gfp_flags, int migratetype)
2959 struct per_cpu_pages *pcp;
2960 struct list_head *list;
2962 unsigned long flags;
2964 local_irq_save(flags);
2965 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2966 list = &pcp->lists[migratetype];
2967 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2969 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2970 zone_statistics(preferred_zone, zone);
2972 local_irq_restore(flags);
2977 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2980 struct page *rmqueue(struct zone *preferred_zone,
2981 struct zone *zone, unsigned int order,
2982 gfp_t gfp_flags, unsigned int alloc_flags,
2985 unsigned long flags;
2988 if (likely(order == 0)) {
2989 page = rmqueue_pcplist(preferred_zone, zone, order,
2990 gfp_flags, migratetype);
2995 * We most definitely don't want callers attempting to
2996 * allocate greater than order-1 page units with __GFP_NOFAIL.
2998 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2999 spin_lock_irqsave(&zone->lock, flags);
3003 if (alloc_flags & ALLOC_HARDER) {
3004 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3006 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3009 page = __rmqueue(zone, order, migratetype);
3010 } while (page && check_new_pages(page, order));
3011 spin_unlock(&zone->lock);
3014 __mod_zone_freepage_state(zone, -(1 << order),
3015 get_pcppage_migratetype(page));
3017 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3018 zone_statistics(preferred_zone, zone);
3019 local_irq_restore(flags);
3022 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3026 local_irq_restore(flags);
3030 #ifdef CONFIG_FAIL_PAGE_ALLOC
3033 struct fault_attr attr;
3035 bool ignore_gfp_highmem;
3036 bool ignore_gfp_reclaim;
3038 } fail_page_alloc = {
3039 .attr = FAULT_ATTR_INITIALIZER,
3040 .ignore_gfp_reclaim = true,
3041 .ignore_gfp_highmem = true,
3045 static int __init setup_fail_page_alloc(char *str)
3047 return setup_fault_attr(&fail_page_alloc.attr, str);
3049 __setup("fail_page_alloc=", setup_fail_page_alloc);
3051 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3053 if (order < fail_page_alloc.min_order)
3055 if (gfp_mask & __GFP_NOFAIL)
3057 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3059 if (fail_page_alloc.ignore_gfp_reclaim &&
3060 (gfp_mask & __GFP_DIRECT_RECLAIM))
3063 return should_fail(&fail_page_alloc.attr, 1 << order);
3066 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3068 static int __init fail_page_alloc_debugfs(void)
3070 umode_t mode = S_IFREG | 0600;
3073 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3074 &fail_page_alloc.attr);
3076 return PTR_ERR(dir);
3078 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3079 &fail_page_alloc.ignore_gfp_reclaim))
3081 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3082 &fail_page_alloc.ignore_gfp_highmem))
3084 if (!debugfs_create_u32("min-order", mode, dir,
3085 &fail_page_alloc.min_order))
3090 debugfs_remove_recursive(dir);
3095 late_initcall(fail_page_alloc_debugfs);
3097 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3099 #else /* CONFIG_FAIL_PAGE_ALLOC */
3101 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3106 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3109 * Return true if free base pages are above 'mark'. For high-order checks it
3110 * will return true of the order-0 watermark is reached and there is at least
3111 * one free page of a suitable size. Checking now avoids taking the zone lock
3112 * to check in the allocation paths if no pages are free.
3114 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3115 int classzone_idx, unsigned int alloc_flags,
3120 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3122 /* free_pages may go negative - that's OK */
3123 free_pages -= (1 << order) - 1;
3125 if (alloc_flags & ALLOC_HIGH)
3129 * If the caller does not have rights to ALLOC_HARDER then subtract
3130 * the high-atomic reserves. This will over-estimate the size of the
3131 * atomic reserve but it avoids a search.
3133 if (likely(!alloc_harder)) {
3134 free_pages -= z->nr_reserved_highatomic;
3137 * OOM victims can try even harder than normal ALLOC_HARDER
3138 * users on the grounds that it's definitely going to be in
3139 * the exit path shortly and free memory. Any allocation it
3140 * makes during the free path will be small and short-lived.
3142 if (alloc_flags & ALLOC_OOM)
3150 /* If allocation can't use CMA areas don't use free CMA pages */
3151 if (!(alloc_flags & ALLOC_CMA))
3152 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3156 * Check watermarks for an order-0 allocation request. If these
3157 * are not met, then a high-order request also cannot go ahead
3158 * even if a suitable page happened to be free.
3160 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3163 /* If this is an order-0 request then the watermark is fine */
3167 /* For a high-order request, check at least one suitable page is free */
3168 for (o = order; o < MAX_ORDER; o++) {
3169 struct free_area *area = &z->free_area[o];
3175 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3176 if (!list_empty(&area->free_list[mt]))
3181 if ((alloc_flags & ALLOC_CMA) &&
3182 !list_empty(&area->free_list[MIGRATE_CMA])) {
3187 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3193 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3194 int classzone_idx, unsigned int alloc_flags)
3196 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3197 zone_page_state(z, NR_FREE_PAGES));
3200 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3201 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3203 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3207 /* If allocation can't use CMA areas don't use free CMA pages */
3208 if (!(alloc_flags & ALLOC_CMA))
3209 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3213 * Fast check for order-0 only. If this fails then the reserves
3214 * need to be calculated. There is a corner case where the check
3215 * passes but only the high-order atomic reserve are free. If
3216 * the caller is !atomic then it'll uselessly search the free
3217 * list. That corner case is then slower but it is harmless.
3219 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3222 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3226 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3227 unsigned long mark, int classzone_idx)
3229 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3231 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3232 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3234 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3239 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3241 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3244 #else /* CONFIG_NUMA */
3245 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3249 #endif /* CONFIG_NUMA */
3252 * get_page_from_freelist goes through the zonelist trying to allocate
3255 static struct page *
3256 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3257 const struct alloc_context *ac)
3259 struct zoneref *z = ac->preferred_zoneref;
3261 struct pglist_data *last_pgdat_dirty_limit = NULL;
3264 * Scan zonelist, looking for a zone with enough free.
3265 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3267 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3272 if (cpusets_enabled() &&
3273 (alloc_flags & ALLOC_CPUSET) &&
3274 !__cpuset_zone_allowed(zone, gfp_mask))
3277 * When allocating a page cache page for writing, we
3278 * want to get it from a node that is within its dirty
3279 * limit, such that no single node holds more than its
3280 * proportional share of globally allowed dirty pages.
3281 * The dirty limits take into account the node's
3282 * lowmem reserves and high watermark so that kswapd
3283 * should be able to balance it without having to
3284 * write pages from its LRU list.
3286 * XXX: For now, allow allocations to potentially
3287 * exceed the per-node dirty limit in the slowpath
3288 * (spread_dirty_pages unset) before going into reclaim,
3289 * which is important when on a NUMA setup the allowed
3290 * nodes are together not big enough to reach the
3291 * global limit. The proper fix for these situations
3292 * will require awareness of nodes in the
3293 * dirty-throttling and the flusher threads.
3295 if (ac->spread_dirty_pages) {
3296 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3299 if (!node_dirty_ok(zone->zone_pgdat)) {
3300 last_pgdat_dirty_limit = zone->zone_pgdat;
3305 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3306 if (!zone_watermark_fast(zone, order, mark,
3307 ac_classzone_idx(ac), alloc_flags)) {
3310 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3312 * Watermark failed for this zone, but see if we can
3313 * grow this zone if it contains deferred pages.
3315 if (static_branch_unlikely(&deferred_pages)) {
3316 if (_deferred_grow_zone(zone, order))
3320 /* Checked here to keep the fast path fast */
3321 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3322 if (alloc_flags & ALLOC_NO_WATERMARKS)
3325 if (node_reclaim_mode == 0 ||
3326 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3329 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3331 case NODE_RECLAIM_NOSCAN:
3334 case NODE_RECLAIM_FULL:
3335 /* scanned but unreclaimable */
3338 /* did we reclaim enough */
3339 if (zone_watermark_ok(zone, order, mark,
3340 ac_classzone_idx(ac), alloc_flags))
3348 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3349 gfp_mask, alloc_flags, ac->migratetype);
3351 prep_new_page(page, order, gfp_mask, alloc_flags);
3354 * If this is a high-order atomic allocation then check
3355 * if the pageblock should be reserved for the future
3357 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3358 reserve_highatomic_pageblock(page, zone, order);
3362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3363 /* Try again if zone has deferred pages */
3364 if (static_branch_unlikely(&deferred_pages)) {
3365 if (_deferred_grow_zone(zone, order))
3375 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3377 unsigned int filter = SHOW_MEM_FILTER_NODES;
3378 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3380 if (!__ratelimit(&show_mem_rs))
3384 * This documents exceptions given to allocations in certain
3385 * contexts that are allowed to allocate outside current's set
3388 if (!(gfp_mask & __GFP_NOMEMALLOC))
3389 if (tsk_is_oom_victim(current) ||
3390 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3391 filter &= ~SHOW_MEM_FILTER_NODES;
3392 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3393 filter &= ~SHOW_MEM_FILTER_NODES;
3395 show_mem(filter, nodemask);
3398 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3400 struct va_format vaf;
3402 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3403 DEFAULT_RATELIMIT_BURST);
3405 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3408 va_start(args, fmt);
3411 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3412 current->comm, &vaf, gfp_mask, &gfp_mask,
3413 nodemask_pr_args(nodemask));
3416 cpuset_print_current_mems_allowed();
3419 warn_alloc_show_mem(gfp_mask, nodemask);
3422 static inline struct page *
3423 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3424 unsigned int alloc_flags,
3425 const struct alloc_context *ac)
3429 page = get_page_from_freelist(gfp_mask, order,
3430 alloc_flags|ALLOC_CPUSET, ac);
3432 * fallback to ignore cpuset restriction if our nodes
3436 page = get_page_from_freelist(gfp_mask, order,
3442 static inline struct page *
3443 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3444 const struct alloc_context *ac, unsigned long *did_some_progress)
3446 struct oom_control oc = {
3447 .zonelist = ac->zonelist,
3448 .nodemask = ac->nodemask,
3450 .gfp_mask = gfp_mask,
3455 *did_some_progress = 0;
3458 * Acquire the oom lock. If that fails, somebody else is
3459 * making progress for us.
3461 if (!mutex_trylock(&oom_lock)) {
3462 *did_some_progress = 1;
3463 schedule_timeout_uninterruptible(1);
3468 * Go through the zonelist yet one more time, keep very high watermark
3469 * here, this is only to catch a parallel oom killing, we must fail if
3470 * we're still under heavy pressure. But make sure that this reclaim
3471 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3472 * allocation which will never fail due to oom_lock already held.
3474 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3475 ~__GFP_DIRECT_RECLAIM, order,
3476 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3480 /* Coredumps can quickly deplete all memory reserves */
3481 if (current->flags & PF_DUMPCORE)
3483 /* The OOM killer will not help higher order allocs */
3484 if (order > PAGE_ALLOC_COSTLY_ORDER)
3487 * We have already exhausted all our reclaim opportunities without any
3488 * success so it is time to admit defeat. We will skip the OOM killer
3489 * because it is very likely that the caller has a more reasonable
3490 * fallback than shooting a random task.
3492 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3494 /* The OOM killer does not needlessly kill tasks for lowmem */
3495 if (ac->high_zoneidx < ZONE_NORMAL)
3497 if (pm_suspended_storage())
3500 * XXX: GFP_NOFS allocations should rather fail than rely on
3501 * other request to make a forward progress.
3502 * We are in an unfortunate situation where out_of_memory cannot
3503 * do much for this context but let's try it to at least get
3504 * access to memory reserved if the current task is killed (see
3505 * out_of_memory). Once filesystems are ready to handle allocation
3506 * failures more gracefully we should just bail out here.
3509 /* The OOM killer may not free memory on a specific node */
3510 if (gfp_mask & __GFP_THISNODE)
3513 /* Exhausted what can be done so it's blame time */
3514 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3515 *did_some_progress = 1;
3518 * Help non-failing allocations by giving them access to memory
3521 if (gfp_mask & __GFP_NOFAIL)
3522 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3523 ALLOC_NO_WATERMARKS, ac);
3526 mutex_unlock(&oom_lock);
3531 * Maximum number of compaction retries wit a progress before OOM
3532 * killer is consider as the only way to move forward.
3534 #define MAX_COMPACT_RETRIES 16
3536 #ifdef CONFIG_COMPACTION
3537 /* Try memory compaction for high-order allocations before reclaim */
3538 static struct page *
3539 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3540 unsigned int alloc_flags, const struct alloc_context *ac,
3541 enum compact_priority prio, enum compact_result *compact_result)
3544 unsigned long pflags;
3545 unsigned int noreclaim_flag;
3550 psi_memstall_enter(&pflags);
3551 noreclaim_flag = memalloc_noreclaim_save();
3553 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3556 memalloc_noreclaim_restore(noreclaim_flag);
3557 psi_memstall_leave(&pflags);
3559 if (*compact_result <= COMPACT_INACTIVE)
3563 * At least in one zone compaction wasn't deferred or skipped, so let's
3564 * count a compaction stall
3566 count_vm_event(COMPACTSTALL);
3568 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3571 struct zone *zone = page_zone(page);
3573 zone->compact_blockskip_flush = false;
3574 compaction_defer_reset(zone, order, true);
3575 count_vm_event(COMPACTSUCCESS);
3580 * It's bad if compaction run occurs and fails. The most likely reason
3581 * is that pages exist, but not enough to satisfy watermarks.
3583 count_vm_event(COMPACTFAIL);
3591 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3592 enum compact_result compact_result,
3593 enum compact_priority *compact_priority,
3594 int *compaction_retries)
3596 int max_retries = MAX_COMPACT_RETRIES;
3599 int retries = *compaction_retries;
3600 enum compact_priority priority = *compact_priority;
3605 if (compaction_made_progress(compact_result))
3606 (*compaction_retries)++;
3609 * compaction considers all the zone as desperately out of memory
3610 * so it doesn't really make much sense to retry except when the
3611 * failure could be caused by insufficient priority
3613 if (compaction_failed(compact_result))
3614 goto check_priority;
3617 * make sure the compaction wasn't deferred or didn't bail out early
3618 * due to locks contention before we declare that we should give up.
3619 * But do not retry if the given zonelist is not suitable for
3622 if (compaction_withdrawn(compact_result)) {
3623 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3628 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3629 * costly ones because they are de facto nofail and invoke OOM
3630 * killer to move on while costly can fail and users are ready
3631 * to cope with that. 1/4 retries is rather arbitrary but we
3632 * would need much more detailed feedback from compaction to
3633 * make a better decision.
3635 if (order > PAGE_ALLOC_COSTLY_ORDER)
3637 if (*compaction_retries <= max_retries) {
3643 * Make sure there are attempts at the highest priority if we exhausted
3644 * all retries or failed at the lower priorities.
3647 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3648 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3650 if (*compact_priority > min_priority) {
3651 (*compact_priority)--;
3652 *compaction_retries = 0;
3656 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3660 static inline struct page *
3661 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3662 unsigned int alloc_flags, const struct alloc_context *ac,
3663 enum compact_priority prio, enum compact_result *compact_result)
3665 *compact_result = COMPACT_SKIPPED;
3670 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3671 enum compact_result compact_result,
3672 enum compact_priority *compact_priority,
3673 int *compaction_retries)
3678 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3682 * There are setups with compaction disabled which would prefer to loop
3683 * inside the allocator rather than hit the oom killer prematurely.
3684 * Let's give them a good hope and keep retrying while the order-0
3685 * watermarks are OK.
3687 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3689 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3690 ac_classzone_idx(ac), alloc_flags))
3695 #endif /* CONFIG_COMPACTION */
3697 #ifdef CONFIG_LOCKDEP
3698 static struct lockdep_map __fs_reclaim_map =
3699 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3701 static bool __need_fs_reclaim(gfp_t gfp_mask)
3703 gfp_mask = current_gfp_context(gfp_mask);
3705 /* no reclaim without waiting on it */
3706 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3709 /* this guy won't enter reclaim */
3710 if (current->flags & PF_MEMALLOC)
3713 /* We're only interested __GFP_FS allocations for now */
3714 if (!(gfp_mask & __GFP_FS))
3717 if (gfp_mask & __GFP_NOLOCKDEP)
3723 void __fs_reclaim_acquire(void)
3725 lock_map_acquire(&__fs_reclaim_map);
3728 void __fs_reclaim_release(void)
3730 lock_map_release(&__fs_reclaim_map);
3733 void fs_reclaim_acquire(gfp_t gfp_mask)
3735 if (__need_fs_reclaim(gfp_mask))
3736 __fs_reclaim_acquire();
3738 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3740 void fs_reclaim_release(gfp_t gfp_mask)
3742 if (__need_fs_reclaim(gfp_mask))
3743 __fs_reclaim_release();
3745 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3748 /* Perform direct synchronous page reclaim */
3750 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3751 const struct alloc_context *ac)
3753 struct reclaim_state reclaim_state;
3755 unsigned int noreclaim_flag;
3756 unsigned long pflags;
3760 /* We now go into synchronous reclaim */
3761 cpuset_memory_pressure_bump();
3762 psi_memstall_enter(&pflags);
3763 fs_reclaim_acquire(gfp_mask);
3764 noreclaim_flag = memalloc_noreclaim_save();
3765 reclaim_state.reclaimed_slab = 0;
3766 current->reclaim_state = &reclaim_state;
3768 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3771 current->reclaim_state = NULL;
3772 memalloc_noreclaim_restore(noreclaim_flag);
3773 fs_reclaim_release(gfp_mask);
3774 psi_memstall_leave(&pflags);
3781 /* The really slow allocator path where we enter direct reclaim */
3782 static inline struct page *
3783 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3784 unsigned int alloc_flags, const struct alloc_context *ac,
3785 unsigned long *did_some_progress)
3787 struct page *page = NULL;
3788 bool drained = false;
3790 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3791 if (unlikely(!(*did_some_progress)))
3795 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3798 * If an allocation failed after direct reclaim, it could be because
3799 * pages are pinned on the per-cpu lists or in high alloc reserves.
3800 * Shrink them them and try again
3802 if (!page && !drained) {
3803 unreserve_highatomic_pageblock(ac, false);
3804 drain_all_pages(NULL);
3812 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3813 const struct alloc_context *ac)
3817 pg_data_t *last_pgdat = NULL;
3818 enum zone_type high_zoneidx = ac->high_zoneidx;
3820 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3822 if (last_pgdat != zone->zone_pgdat)
3823 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3824 last_pgdat = zone->zone_pgdat;
3828 static inline unsigned int
3829 gfp_to_alloc_flags(gfp_t gfp_mask)
3831 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3833 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3834 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3837 * The caller may dip into page reserves a bit more if the caller
3838 * cannot run direct reclaim, or if the caller has realtime scheduling
3839 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3840 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3842 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3844 if (gfp_mask & __GFP_ATOMIC) {
3846 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3847 * if it can't schedule.
3849 if (!(gfp_mask & __GFP_NOMEMALLOC))
3850 alloc_flags |= ALLOC_HARDER;
3852 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3853 * comment for __cpuset_node_allowed().
3855 alloc_flags &= ~ALLOC_CPUSET;
3856 } else if (unlikely(rt_task(current)) && !in_interrupt())
3857 alloc_flags |= ALLOC_HARDER;
3860 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3861 alloc_flags |= ALLOC_CMA;
3866 static bool oom_reserves_allowed(struct task_struct *tsk)
3868 if (!tsk_is_oom_victim(tsk))
3872 * !MMU doesn't have oom reaper so give access to memory reserves
3873 * only to the thread with TIF_MEMDIE set
3875 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3882 * Distinguish requests which really need access to full memory
3883 * reserves from oom victims which can live with a portion of it
3885 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3887 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3889 if (gfp_mask & __GFP_MEMALLOC)
3890 return ALLOC_NO_WATERMARKS;
3891 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3892 return ALLOC_NO_WATERMARKS;
3893 if (!in_interrupt()) {
3894 if (current->flags & PF_MEMALLOC)
3895 return ALLOC_NO_WATERMARKS;
3896 else if (oom_reserves_allowed(current))
3903 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3905 return !!__gfp_pfmemalloc_flags(gfp_mask);
3909 * Checks whether it makes sense to retry the reclaim to make a forward progress
3910 * for the given allocation request.
3912 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3913 * without success, or when we couldn't even meet the watermark if we
3914 * reclaimed all remaining pages on the LRU lists.
3916 * Returns true if a retry is viable or false to enter the oom path.
3919 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3920 struct alloc_context *ac, int alloc_flags,
3921 bool did_some_progress, int *no_progress_loops)
3928 * Costly allocations might have made a progress but this doesn't mean
3929 * their order will become available due to high fragmentation so
3930 * always increment the no progress counter for them
3932 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3933 *no_progress_loops = 0;
3935 (*no_progress_loops)++;
3938 * Make sure we converge to OOM if we cannot make any progress
3939 * several times in the row.
3941 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3942 /* Before OOM, exhaust highatomic_reserve */
3943 return unreserve_highatomic_pageblock(ac, true);
3947 * Keep reclaiming pages while there is a chance this will lead
3948 * somewhere. If none of the target zones can satisfy our allocation
3949 * request even if all reclaimable pages are considered then we are
3950 * screwed and have to go OOM.
3952 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3954 unsigned long available;
3955 unsigned long reclaimable;
3956 unsigned long min_wmark = min_wmark_pages(zone);
3959 available = reclaimable = zone_reclaimable_pages(zone);
3960 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3963 * Would the allocation succeed if we reclaimed all
3964 * reclaimable pages?
3966 wmark = __zone_watermark_ok(zone, order, min_wmark,
3967 ac_classzone_idx(ac), alloc_flags, available);
3968 trace_reclaim_retry_zone(z, order, reclaimable,
3969 available, min_wmark, *no_progress_loops, wmark);
3972 * If we didn't make any progress and have a lot of
3973 * dirty + writeback pages then we should wait for
3974 * an IO to complete to slow down the reclaim and
3975 * prevent from pre mature OOM
3977 if (!did_some_progress) {
3978 unsigned long write_pending;
3980 write_pending = zone_page_state_snapshot(zone,
3981 NR_ZONE_WRITE_PENDING);
3983 if (2 * write_pending > reclaimable) {
3984 congestion_wait(BLK_RW_ASYNC, HZ/10);
3996 * Memory allocation/reclaim might be called from a WQ context and the
3997 * current implementation of the WQ concurrency control doesn't
3998 * recognize that a particular WQ is congested if the worker thread is
3999 * looping without ever sleeping. Therefore we have to do a short sleep
4000 * here rather than calling cond_resched().
4002 if (current->flags & PF_WQ_WORKER)
4003 schedule_timeout_uninterruptible(1);
4010 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4013 * It's possible that cpuset's mems_allowed and the nodemask from
4014 * mempolicy don't intersect. This should be normally dealt with by
4015 * policy_nodemask(), but it's possible to race with cpuset update in
4016 * such a way the check therein was true, and then it became false
4017 * before we got our cpuset_mems_cookie here.
4018 * This assumes that for all allocations, ac->nodemask can come only
4019 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4020 * when it does not intersect with the cpuset restrictions) or the
4021 * caller can deal with a violated nodemask.
4023 if (cpusets_enabled() && ac->nodemask &&
4024 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4025 ac->nodemask = NULL;
4030 * When updating a task's mems_allowed or mempolicy nodemask, it is
4031 * possible to race with parallel threads in such a way that our
4032 * allocation can fail while the mask is being updated. If we are about
4033 * to fail, check if the cpuset changed during allocation and if so,
4036 if (read_mems_allowed_retry(cpuset_mems_cookie))
4042 static inline struct page *
4043 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4044 struct alloc_context *ac)
4046 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4047 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4048 struct page *page = NULL;
4049 unsigned int alloc_flags;
4050 unsigned long did_some_progress;
4051 enum compact_priority compact_priority;
4052 enum compact_result compact_result;
4053 int compaction_retries;
4054 int no_progress_loops;
4055 unsigned int cpuset_mems_cookie;
4059 * In the slowpath, we sanity check order to avoid ever trying to
4060 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4061 * be using allocators in order of preference for an area that is
4064 if (order >= MAX_ORDER) {
4065 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4070 * We also sanity check to catch abuse of atomic reserves being used by
4071 * callers that are not in atomic context.
4073 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4074 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4075 gfp_mask &= ~__GFP_ATOMIC;
4078 compaction_retries = 0;
4079 no_progress_loops = 0;
4080 compact_priority = DEF_COMPACT_PRIORITY;
4081 cpuset_mems_cookie = read_mems_allowed_begin();
4084 * The fast path uses conservative alloc_flags to succeed only until
4085 * kswapd needs to be woken up, and to avoid the cost of setting up
4086 * alloc_flags precisely. So we do that now.
4088 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4091 * We need to recalculate the starting point for the zonelist iterator
4092 * because we might have used different nodemask in the fast path, or
4093 * there was a cpuset modification and we are retrying - otherwise we
4094 * could end up iterating over non-eligible zones endlessly.
4096 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4097 ac->high_zoneidx, ac->nodemask);
4098 if (!ac->preferred_zoneref->zone)
4101 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4102 wake_all_kswapds(order, gfp_mask, ac);
4105 * The adjusted alloc_flags might result in immediate success, so try
4108 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4113 * For costly allocations, try direct compaction first, as it's likely
4114 * that we have enough base pages and don't need to reclaim. For non-
4115 * movable high-order allocations, do that as well, as compaction will
4116 * try prevent permanent fragmentation by migrating from blocks of the
4118 * Don't try this for allocations that are allowed to ignore
4119 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4121 if (can_direct_reclaim &&
4123 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4124 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4125 page = __alloc_pages_direct_compact(gfp_mask, order,
4127 INIT_COMPACT_PRIORITY,
4133 * Checks for costly allocations with __GFP_NORETRY, which
4134 * includes THP page fault allocations
4136 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4138 * If compaction is deferred for high-order allocations,
4139 * it is because sync compaction recently failed. If
4140 * this is the case and the caller requested a THP
4141 * allocation, we do not want to heavily disrupt the
4142 * system, so we fail the allocation instead of entering
4145 if (compact_result == COMPACT_DEFERRED)
4149 * Looks like reclaim/compaction is worth trying, but
4150 * sync compaction could be very expensive, so keep
4151 * using async compaction.
4153 compact_priority = INIT_COMPACT_PRIORITY;
4158 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4159 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4160 wake_all_kswapds(order, gfp_mask, ac);
4162 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4164 alloc_flags = reserve_flags;
4167 * Reset the nodemask and zonelist iterators if memory policies can be
4168 * ignored. These allocations are high priority and system rather than
4171 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4172 ac->nodemask = NULL;
4173 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4174 ac->high_zoneidx, ac->nodemask);
4177 /* Attempt with potentially adjusted zonelist and alloc_flags */
4178 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4182 /* Caller is not willing to reclaim, we can't balance anything */
4183 if (!can_direct_reclaim)
4186 /* Avoid recursion of direct reclaim */
4187 if (current->flags & PF_MEMALLOC)
4190 /* Try direct reclaim and then allocating */
4191 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4192 &did_some_progress);
4196 /* Try direct compaction and then allocating */
4197 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4198 compact_priority, &compact_result);
4202 /* Do not loop if specifically requested */
4203 if (gfp_mask & __GFP_NORETRY)
4207 * Do not retry costly high order allocations unless they are
4208 * __GFP_RETRY_MAYFAIL
4210 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4213 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4214 did_some_progress > 0, &no_progress_loops))
4218 * It doesn't make any sense to retry for the compaction if the order-0
4219 * reclaim is not able to make any progress because the current
4220 * implementation of the compaction depends on the sufficient amount
4221 * of free memory (see __compaction_suitable)
4223 if (did_some_progress > 0 &&
4224 should_compact_retry(ac, order, alloc_flags,
4225 compact_result, &compact_priority,
4226 &compaction_retries))
4230 /* Deal with possible cpuset update races before we start OOM killing */
4231 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4234 /* Reclaim has failed us, start killing things */
4235 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4239 /* Avoid allocations with no watermarks from looping endlessly */
4240 if (tsk_is_oom_victim(current) &&
4241 (alloc_flags == ALLOC_OOM ||
4242 (gfp_mask & __GFP_NOMEMALLOC)))
4245 /* Retry as long as the OOM killer is making progress */
4246 if (did_some_progress) {
4247 no_progress_loops = 0;
4252 /* Deal with possible cpuset update races before we fail */
4253 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4257 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4260 if (gfp_mask & __GFP_NOFAIL) {
4262 * All existing users of the __GFP_NOFAIL are blockable, so warn
4263 * of any new users that actually require GFP_NOWAIT
4265 if (WARN_ON_ONCE(!can_direct_reclaim))
4269 * PF_MEMALLOC request from this context is rather bizarre
4270 * because we cannot reclaim anything and only can loop waiting
4271 * for somebody to do a work for us
4273 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4276 * non failing costly orders are a hard requirement which we
4277 * are not prepared for much so let's warn about these users
4278 * so that we can identify them and convert them to something
4281 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4284 * Help non-failing allocations by giving them access to memory
4285 * reserves but do not use ALLOC_NO_WATERMARKS because this
4286 * could deplete whole memory reserves which would just make
4287 * the situation worse
4289 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4297 warn_alloc(gfp_mask, ac->nodemask,
4298 "page allocation failure: order:%u", order);
4303 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4304 int preferred_nid, nodemask_t *nodemask,
4305 struct alloc_context *ac, gfp_t *alloc_mask,
4306 unsigned int *alloc_flags)
4308 ac->high_zoneidx = gfp_zone(gfp_mask);
4309 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4310 ac->nodemask = nodemask;
4311 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4313 if (cpusets_enabled()) {
4314 *alloc_mask |= __GFP_HARDWALL;
4316 ac->nodemask = &cpuset_current_mems_allowed;
4318 *alloc_flags |= ALLOC_CPUSET;
4321 fs_reclaim_acquire(gfp_mask);
4322 fs_reclaim_release(gfp_mask);
4324 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4326 if (should_fail_alloc_page(gfp_mask, order))
4329 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4330 *alloc_flags |= ALLOC_CMA;
4335 /* Determine whether to spread dirty pages and what the first usable zone */
4336 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4338 /* Dirty zone balancing only done in the fast path */
4339 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4342 * The preferred zone is used for statistics but crucially it is
4343 * also used as the starting point for the zonelist iterator. It
4344 * may get reset for allocations that ignore memory policies.
4346 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4347 ac->high_zoneidx, ac->nodemask);
4351 * This is the 'heart' of the zoned buddy allocator.
4354 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4355 nodemask_t *nodemask)
4358 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4359 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4360 struct alloc_context ac = { };
4362 gfp_mask &= gfp_allowed_mask;
4363 alloc_mask = gfp_mask;
4364 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4367 finalise_ac(gfp_mask, &ac);
4369 /* First allocation attempt */
4370 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4375 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4376 * resp. GFP_NOIO which has to be inherited for all allocation requests
4377 * from a particular context which has been marked by
4378 * memalloc_no{fs,io}_{save,restore}.
4380 alloc_mask = current_gfp_context(gfp_mask);
4381 ac.spread_dirty_pages = false;
4384 * Restore the original nodemask if it was potentially replaced with
4385 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4387 if (unlikely(ac.nodemask != nodemask))
4388 ac.nodemask = nodemask;
4390 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4393 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4394 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4395 __free_pages(page, order);
4399 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4403 EXPORT_SYMBOL(__alloc_pages_nodemask);
4406 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4407 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4408 * you need to access high mem.
4410 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4414 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4417 return (unsigned long) page_address(page);
4419 EXPORT_SYMBOL(__get_free_pages);
4421 unsigned long get_zeroed_page(gfp_t gfp_mask)
4423 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4425 EXPORT_SYMBOL(get_zeroed_page);
4427 void __free_pages(struct page *page, unsigned int order)
4429 if (put_page_testzero(page)) {
4431 free_unref_page(page);
4433 __free_pages_ok(page, order);
4437 EXPORT_SYMBOL(__free_pages);
4439 void free_pages(unsigned long addr, unsigned int order)
4442 VM_BUG_ON(!virt_addr_valid((void *)addr));
4443 __free_pages(virt_to_page((void *)addr), order);
4447 EXPORT_SYMBOL(free_pages);
4451 * An arbitrary-length arbitrary-offset area of memory which resides
4452 * within a 0 or higher order page. Multiple fragments within that page
4453 * are individually refcounted, in the page's reference counter.
4455 * The page_frag functions below provide a simple allocation framework for
4456 * page fragments. This is used by the network stack and network device
4457 * drivers to provide a backing region of memory for use as either an
4458 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4460 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4463 struct page *page = NULL;
4464 gfp_t gfp = gfp_mask;
4466 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4467 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4469 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4470 PAGE_FRAG_CACHE_MAX_ORDER);
4471 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4473 if (unlikely(!page))
4474 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4476 nc->va = page ? page_address(page) : NULL;
4481 void __page_frag_cache_drain(struct page *page, unsigned int count)
4483 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4485 if (page_ref_sub_and_test(page, count)) {
4486 unsigned int order = compound_order(page);
4489 free_unref_page(page);
4491 __free_pages_ok(page, order);
4494 EXPORT_SYMBOL(__page_frag_cache_drain);
4496 void *page_frag_alloc(struct page_frag_cache *nc,
4497 unsigned int fragsz, gfp_t gfp_mask)
4499 unsigned int size = PAGE_SIZE;
4503 if (unlikely(!nc->va)) {
4505 page = __page_frag_cache_refill(nc, gfp_mask);
4509 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4510 /* if size can vary use size else just use PAGE_SIZE */
4513 /* Even if we own the page, we do not use atomic_set().
4514 * This would break get_page_unless_zero() users.
4516 page_ref_add(page, size - 1);
4518 /* reset page count bias and offset to start of new frag */
4519 nc->pfmemalloc = page_is_pfmemalloc(page);
4520 nc->pagecnt_bias = size;
4524 offset = nc->offset - fragsz;
4525 if (unlikely(offset < 0)) {
4526 page = virt_to_page(nc->va);
4528 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4531 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4532 /* if size can vary use size else just use PAGE_SIZE */
4535 /* OK, page count is 0, we can safely set it */
4536 set_page_count(page, size);
4538 /* reset page count bias and offset to start of new frag */
4539 nc->pagecnt_bias = size;
4540 offset = size - fragsz;
4544 nc->offset = offset;
4546 return nc->va + offset;
4548 EXPORT_SYMBOL(page_frag_alloc);
4551 * Frees a page fragment allocated out of either a compound or order 0 page.
4553 void page_frag_free(void *addr)
4555 struct page *page = virt_to_head_page(addr);
4557 if (unlikely(put_page_testzero(page)))
4558 __free_pages_ok(page, compound_order(page));
4560 EXPORT_SYMBOL(page_frag_free);
4562 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4566 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4567 unsigned long used = addr + PAGE_ALIGN(size);
4569 split_page(virt_to_page((void *)addr), order);
4570 while (used < alloc_end) {
4575 return (void *)addr;
4579 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4580 * @size: the number of bytes to allocate
4581 * @gfp_mask: GFP flags for the allocation
4583 * This function is similar to alloc_pages(), except that it allocates the
4584 * minimum number of pages to satisfy the request. alloc_pages() can only
4585 * allocate memory in power-of-two pages.
4587 * This function is also limited by MAX_ORDER.
4589 * Memory allocated by this function must be released by free_pages_exact().
4591 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4593 unsigned int order = get_order(size);
4596 addr = __get_free_pages(gfp_mask, order);
4597 return make_alloc_exact(addr, order, size);
4599 EXPORT_SYMBOL(alloc_pages_exact);
4602 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4604 * @nid: the preferred node ID where memory should be allocated
4605 * @size: the number of bytes to allocate
4606 * @gfp_mask: GFP flags for the allocation
4608 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4611 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4613 unsigned int order = get_order(size);
4614 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4617 return make_alloc_exact((unsigned long)page_address(p), order, size);
4621 * free_pages_exact - release memory allocated via alloc_pages_exact()
4622 * @virt: the value returned by alloc_pages_exact.
4623 * @size: size of allocation, same value as passed to alloc_pages_exact().
4625 * Release the memory allocated by a previous call to alloc_pages_exact.
4627 void free_pages_exact(void *virt, size_t size)
4629 unsigned long addr = (unsigned long)virt;
4630 unsigned long end = addr + PAGE_ALIGN(size);
4632 while (addr < end) {
4637 EXPORT_SYMBOL(free_pages_exact);
4640 * nr_free_zone_pages - count number of pages beyond high watermark
4641 * @offset: The zone index of the highest zone
4643 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4644 * high watermark within all zones at or below a given zone index. For each
4645 * zone, the number of pages is calculated as:
4647 * nr_free_zone_pages = managed_pages - high_pages
4649 static unsigned long nr_free_zone_pages(int offset)
4654 /* Just pick one node, since fallback list is circular */
4655 unsigned long sum = 0;
4657 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4659 for_each_zone_zonelist(zone, z, zonelist, offset) {
4660 unsigned long size = zone->managed_pages;
4661 unsigned long high = high_wmark_pages(zone);
4670 * nr_free_buffer_pages - count number of pages beyond high watermark
4672 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4673 * watermark within ZONE_DMA and ZONE_NORMAL.
4675 unsigned long nr_free_buffer_pages(void)
4677 return nr_free_zone_pages(gfp_zone(GFP_USER));
4679 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4682 * nr_free_pagecache_pages - count number of pages beyond high watermark
4684 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4685 * high watermark within all zones.
4687 unsigned long nr_free_pagecache_pages(void)
4689 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4692 static inline void show_node(struct zone *zone)
4694 if (IS_ENABLED(CONFIG_NUMA))
4695 printk("Node %d ", zone_to_nid(zone));
4698 long si_mem_available(void)
4701 unsigned long pagecache;
4702 unsigned long wmark_low = 0;
4703 unsigned long pages[NR_LRU_LISTS];
4704 unsigned long reclaimable;
4708 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4709 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4712 wmark_low += zone->watermark[WMARK_LOW];
4715 * Estimate the amount of memory available for userspace allocations,
4716 * without causing swapping.
4718 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4721 * Not all the page cache can be freed, otherwise the system will
4722 * start swapping. Assume at least half of the page cache, or the
4723 * low watermark worth of cache, needs to stay.
4725 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4726 pagecache -= min(pagecache / 2, wmark_low);
4727 available += pagecache;
4730 * Part of the reclaimable slab and other kernel memory consists of
4731 * items that are in use, and cannot be freed. Cap this estimate at the
4734 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4735 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4736 available += reclaimable - min(reclaimable / 2, wmark_low);
4742 EXPORT_SYMBOL_GPL(si_mem_available);
4744 void si_meminfo(struct sysinfo *val)
4746 val->totalram = totalram_pages;
4747 val->sharedram = global_node_page_state(NR_SHMEM);
4748 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4749 val->bufferram = nr_blockdev_pages();
4750 val->totalhigh = totalhigh_pages;
4751 val->freehigh = nr_free_highpages();
4752 val->mem_unit = PAGE_SIZE;
4755 EXPORT_SYMBOL(si_meminfo);
4758 void si_meminfo_node(struct sysinfo *val, int nid)
4760 int zone_type; /* needs to be signed */
4761 unsigned long managed_pages = 0;
4762 unsigned long managed_highpages = 0;
4763 unsigned long free_highpages = 0;
4764 pg_data_t *pgdat = NODE_DATA(nid);
4766 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4767 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4768 val->totalram = managed_pages;
4769 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4770 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4771 #ifdef CONFIG_HIGHMEM
4772 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4773 struct zone *zone = &pgdat->node_zones[zone_type];
4775 if (is_highmem(zone)) {
4776 managed_highpages += zone->managed_pages;
4777 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4780 val->totalhigh = managed_highpages;
4781 val->freehigh = free_highpages;
4783 val->totalhigh = managed_highpages;
4784 val->freehigh = free_highpages;
4786 val->mem_unit = PAGE_SIZE;
4791 * Determine whether the node should be displayed or not, depending on whether
4792 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4794 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4796 if (!(flags & SHOW_MEM_FILTER_NODES))
4800 * no node mask - aka implicit memory numa policy. Do not bother with
4801 * the synchronization - read_mems_allowed_begin - because we do not
4802 * have to be precise here.
4805 nodemask = &cpuset_current_mems_allowed;
4807 return !node_isset(nid, *nodemask);
4810 #define K(x) ((x) << (PAGE_SHIFT-10))
4812 static void show_migration_types(unsigned char type)
4814 static const char types[MIGRATE_TYPES] = {
4815 [MIGRATE_UNMOVABLE] = 'U',
4816 [MIGRATE_MOVABLE] = 'M',
4817 [MIGRATE_RECLAIMABLE] = 'E',
4818 [MIGRATE_HIGHATOMIC] = 'H',
4820 [MIGRATE_CMA] = 'C',
4822 #ifdef CONFIG_MEMORY_ISOLATION
4823 [MIGRATE_ISOLATE] = 'I',
4826 char tmp[MIGRATE_TYPES + 1];
4830 for (i = 0; i < MIGRATE_TYPES; i++) {
4831 if (type & (1 << i))
4836 printk(KERN_CONT "(%s) ", tmp);
4840 * Show free area list (used inside shift_scroll-lock stuff)
4841 * We also calculate the percentage fragmentation. We do this by counting the
4842 * memory on each free list with the exception of the first item on the list.
4845 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4848 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4850 unsigned long free_pcp = 0;
4855 for_each_populated_zone(zone) {
4856 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4859 for_each_online_cpu(cpu)
4860 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4863 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4864 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4865 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4866 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4867 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4868 " free:%lu free_pcp:%lu free_cma:%lu\n",
4869 global_node_page_state(NR_ACTIVE_ANON),
4870 global_node_page_state(NR_INACTIVE_ANON),
4871 global_node_page_state(NR_ISOLATED_ANON),
4872 global_node_page_state(NR_ACTIVE_FILE),
4873 global_node_page_state(NR_INACTIVE_FILE),
4874 global_node_page_state(NR_ISOLATED_FILE),
4875 global_node_page_state(NR_UNEVICTABLE),
4876 global_node_page_state(NR_FILE_DIRTY),
4877 global_node_page_state(NR_WRITEBACK),
4878 global_node_page_state(NR_UNSTABLE_NFS),
4879 global_node_page_state(NR_SLAB_RECLAIMABLE),
4880 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4881 global_node_page_state(NR_FILE_MAPPED),
4882 global_node_page_state(NR_SHMEM),
4883 global_zone_page_state(NR_PAGETABLE),
4884 global_zone_page_state(NR_BOUNCE),
4885 global_zone_page_state(NR_FREE_PAGES),
4887 global_zone_page_state(NR_FREE_CMA_PAGES));
4889 for_each_online_pgdat(pgdat) {
4890 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4894 " active_anon:%lukB"
4895 " inactive_anon:%lukB"
4896 " active_file:%lukB"
4897 " inactive_file:%lukB"
4898 " unevictable:%lukB"
4899 " isolated(anon):%lukB"
4900 " isolated(file):%lukB"
4905 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4907 " shmem_pmdmapped: %lukB"
4910 " writeback_tmp:%lukB"
4912 " all_unreclaimable? %s"
4915 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4916 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4917 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4918 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4919 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4920 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4921 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4922 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4923 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4924 K(node_page_state(pgdat, NR_WRITEBACK)),
4925 K(node_page_state(pgdat, NR_SHMEM)),
4926 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4927 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4928 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4930 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4932 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4933 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4934 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4938 for_each_populated_zone(zone) {
4941 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4945 for_each_online_cpu(cpu)
4946 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4955 " active_anon:%lukB"
4956 " inactive_anon:%lukB"
4957 " active_file:%lukB"
4958 " inactive_file:%lukB"
4959 " unevictable:%lukB"
4960 " writepending:%lukB"
4964 " kernel_stack:%lukB"
4972 K(zone_page_state(zone, NR_FREE_PAGES)),
4973 K(min_wmark_pages(zone)),
4974 K(low_wmark_pages(zone)),
4975 K(high_wmark_pages(zone)),
4976 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4977 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4978 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4979 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4980 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4981 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4982 K(zone->present_pages),
4983 K(zone->managed_pages),
4984 K(zone_page_state(zone, NR_MLOCK)),
4985 zone_page_state(zone, NR_KERNEL_STACK_KB),
4986 K(zone_page_state(zone, NR_PAGETABLE)),
4987 K(zone_page_state(zone, NR_BOUNCE)),
4989 K(this_cpu_read(zone->pageset->pcp.count)),
4990 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4991 printk("lowmem_reserve[]:");
4992 for (i = 0; i < MAX_NR_ZONES; i++)
4993 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4994 printk(KERN_CONT "\n");
4997 for_each_populated_zone(zone) {
4999 unsigned long nr[MAX_ORDER], flags, total = 0;
5000 unsigned char types[MAX_ORDER];
5002 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5005 printk(KERN_CONT "%s: ", zone->name);
5007 spin_lock_irqsave(&zone->lock, flags);
5008 for (order = 0; order < MAX_ORDER; order++) {
5009 struct free_area *area = &zone->free_area[order];
5012 nr[order] = area->nr_free;
5013 total += nr[order] << order;
5016 for (type = 0; type < MIGRATE_TYPES; type++) {
5017 if (!list_empty(&area->free_list[type]))
5018 types[order] |= 1 << type;
5021 spin_unlock_irqrestore(&zone->lock, flags);
5022 for (order = 0; order < MAX_ORDER; order++) {
5023 printk(KERN_CONT "%lu*%lukB ",
5024 nr[order], K(1UL) << order);
5026 show_migration_types(types[order]);
5028 printk(KERN_CONT "= %lukB\n", K(total));
5031 hugetlb_show_meminfo();
5033 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5035 show_swap_cache_info();
5038 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5040 zoneref->zone = zone;
5041 zoneref->zone_idx = zone_idx(zone);
5045 * Builds allocation fallback zone lists.
5047 * Add all populated zones of a node to the zonelist.
5049 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5052 enum zone_type zone_type = MAX_NR_ZONES;
5057 zone = pgdat->node_zones + zone_type;
5058 if (managed_zone(zone)) {
5059 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5060 check_highest_zone(zone_type);
5062 } while (zone_type);
5069 static int __parse_numa_zonelist_order(char *s)
5072 * We used to support different zonlists modes but they turned
5073 * out to be just not useful. Let's keep the warning in place
5074 * if somebody still use the cmd line parameter so that we do
5075 * not fail it silently
5077 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5078 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5084 static __init int setup_numa_zonelist_order(char *s)
5089 return __parse_numa_zonelist_order(s);
5091 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5093 char numa_zonelist_order[] = "Node";
5096 * sysctl handler for numa_zonelist_order
5098 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5099 void __user *buffer, size_t *length,
5106 return proc_dostring(table, write, buffer, length, ppos);
5107 str = memdup_user_nul(buffer, 16);
5109 return PTR_ERR(str);
5111 ret = __parse_numa_zonelist_order(str);
5117 #define MAX_NODE_LOAD (nr_online_nodes)
5118 static int node_load[MAX_NUMNODES];
5121 * find_next_best_node - find the next node that should appear in a given node's fallback list
5122 * @node: node whose fallback list we're appending
5123 * @used_node_mask: nodemask_t of already used nodes
5125 * We use a number of factors to determine which is the next node that should
5126 * appear on a given node's fallback list. The node should not have appeared
5127 * already in @node's fallback list, and it should be the next closest node
5128 * according to the distance array (which contains arbitrary distance values
5129 * from each node to each node in the system), and should also prefer nodes
5130 * with no CPUs, since presumably they'll have very little allocation pressure
5131 * on them otherwise.
5132 * It returns -1 if no node is found.
5134 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5137 int min_val = INT_MAX;
5138 int best_node = NUMA_NO_NODE;
5139 const struct cpumask *tmp = cpumask_of_node(0);
5141 /* Use the local node if we haven't already */
5142 if (!node_isset(node, *used_node_mask)) {
5143 node_set(node, *used_node_mask);
5147 for_each_node_state(n, N_MEMORY) {
5149 /* Don't want a node to appear more than once */
5150 if (node_isset(n, *used_node_mask))
5153 /* Use the distance array to find the distance */
5154 val = node_distance(node, n);
5156 /* Penalize nodes under us ("prefer the next node") */
5159 /* Give preference to headless and unused nodes */
5160 tmp = cpumask_of_node(n);
5161 if (!cpumask_empty(tmp))
5162 val += PENALTY_FOR_NODE_WITH_CPUS;
5164 /* Slight preference for less loaded node */
5165 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5166 val += node_load[n];
5168 if (val < min_val) {
5175 node_set(best_node, *used_node_mask);
5182 * Build zonelists ordered by node and zones within node.
5183 * This results in maximum locality--normal zone overflows into local
5184 * DMA zone, if any--but risks exhausting DMA zone.
5186 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5189 struct zoneref *zonerefs;
5192 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5194 for (i = 0; i < nr_nodes; i++) {
5197 pg_data_t *node = NODE_DATA(node_order[i]);
5199 nr_zones = build_zonerefs_node(node, zonerefs);
5200 zonerefs += nr_zones;
5202 zonerefs->zone = NULL;
5203 zonerefs->zone_idx = 0;
5207 * Build gfp_thisnode zonelists
5209 static void build_thisnode_zonelists(pg_data_t *pgdat)
5211 struct zoneref *zonerefs;
5214 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5215 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5216 zonerefs += nr_zones;
5217 zonerefs->zone = NULL;
5218 zonerefs->zone_idx = 0;
5222 * Build zonelists ordered by zone and nodes within zones.
5223 * This results in conserving DMA zone[s] until all Normal memory is
5224 * exhausted, but results in overflowing to remote node while memory
5225 * may still exist in local DMA zone.
5228 static void build_zonelists(pg_data_t *pgdat)
5230 static int node_order[MAX_NUMNODES];
5231 int node, load, nr_nodes = 0;
5232 nodemask_t used_mask;
5233 int local_node, prev_node;
5235 /* NUMA-aware ordering of nodes */
5236 local_node = pgdat->node_id;
5237 load = nr_online_nodes;
5238 prev_node = local_node;
5239 nodes_clear(used_mask);
5241 memset(node_order, 0, sizeof(node_order));
5242 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5244 * We don't want to pressure a particular node.
5245 * So adding penalty to the first node in same
5246 * distance group to make it round-robin.
5248 if (node_distance(local_node, node) !=
5249 node_distance(local_node, prev_node))
5250 node_load[node] = load;
5252 node_order[nr_nodes++] = node;
5257 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5258 build_thisnode_zonelists(pgdat);
5261 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5263 * Return node id of node used for "local" allocations.
5264 * I.e., first node id of first zone in arg node's generic zonelist.
5265 * Used for initializing percpu 'numa_mem', which is used primarily
5266 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5268 int local_memory_node(int node)
5272 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5273 gfp_zone(GFP_KERNEL),
5275 return zone_to_nid(z->zone);
5279 static void setup_min_unmapped_ratio(void);
5280 static void setup_min_slab_ratio(void);
5281 #else /* CONFIG_NUMA */
5283 static void build_zonelists(pg_data_t *pgdat)
5285 int node, local_node;
5286 struct zoneref *zonerefs;
5289 local_node = pgdat->node_id;
5291 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5292 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5293 zonerefs += nr_zones;
5296 * Now we build the zonelist so that it contains the zones
5297 * of all the other nodes.
5298 * We don't want to pressure a particular node, so when
5299 * building the zones for node N, we make sure that the
5300 * zones coming right after the local ones are those from
5301 * node N+1 (modulo N)
5303 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5304 if (!node_online(node))
5306 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5307 zonerefs += nr_zones;
5309 for (node = 0; node < local_node; node++) {
5310 if (!node_online(node))
5312 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5313 zonerefs += nr_zones;
5316 zonerefs->zone = NULL;
5317 zonerefs->zone_idx = 0;
5320 #endif /* CONFIG_NUMA */
5323 * Boot pageset table. One per cpu which is going to be used for all
5324 * zones and all nodes. The parameters will be set in such a way
5325 * that an item put on a list will immediately be handed over to
5326 * the buddy list. This is safe since pageset manipulation is done
5327 * with interrupts disabled.
5329 * The boot_pagesets must be kept even after bootup is complete for
5330 * unused processors and/or zones. They do play a role for bootstrapping
5331 * hotplugged processors.
5333 * zoneinfo_show() and maybe other functions do
5334 * not check if the processor is online before following the pageset pointer.
5335 * Other parts of the kernel may not check if the zone is available.
5337 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5338 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5339 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5341 static void __build_all_zonelists(void *data)
5344 int __maybe_unused cpu;
5345 pg_data_t *self = data;
5346 static DEFINE_SPINLOCK(lock);
5351 memset(node_load, 0, sizeof(node_load));
5355 * This node is hotadded and no memory is yet present. So just
5356 * building zonelists is fine - no need to touch other nodes.
5358 if (self && !node_online(self->node_id)) {
5359 build_zonelists(self);
5361 for_each_online_node(nid) {
5362 pg_data_t *pgdat = NODE_DATA(nid);
5364 build_zonelists(pgdat);
5367 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5369 * We now know the "local memory node" for each node--
5370 * i.e., the node of the first zone in the generic zonelist.
5371 * Set up numa_mem percpu variable for on-line cpus. During
5372 * boot, only the boot cpu should be on-line; we'll init the
5373 * secondary cpus' numa_mem as they come on-line. During
5374 * node/memory hotplug, we'll fixup all on-line cpus.
5376 for_each_online_cpu(cpu)
5377 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5384 static noinline void __init
5385 build_all_zonelists_init(void)
5389 __build_all_zonelists(NULL);
5392 * Initialize the boot_pagesets that are going to be used
5393 * for bootstrapping processors. The real pagesets for
5394 * each zone will be allocated later when the per cpu
5395 * allocator is available.
5397 * boot_pagesets are used also for bootstrapping offline
5398 * cpus if the system is already booted because the pagesets
5399 * are needed to initialize allocators on a specific cpu too.
5400 * F.e. the percpu allocator needs the page allocator which
5401 * needs the percpu allocator in order to allocate its pagesets
5402 * (a chicken-egg dilemma).
5404 for_each_possible_cpu(cpu)
5405 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5407 mminit_verify_zonelist();
5408 cpuset_init_current_mems_allowed();
5412 * unless system_state == SYSTEM_BOOTING.
5414 * __ref due to call of __init annotated helper build_all_zonelists_init
5415 * [protected by SYSTEM_BOOTING].
5417 void __ref build_all_zonelists(pg_data_t *pgdat)
5419 if (system_state == SYSTEM_BOOTING) {
5420 build_all_zonelists_init();
5422 __build_all_zonelists(pgdat);
5423 /* cpuset refresh routine should be here */
5425 vm_total_pages = nr_free_pagecache_pages();
5427 * Disable grouping by mobility if the number of pages in the
5428 * system is too low to allow the mechanism to work. It would be
5429 * more accurate, but expensive to check per-zone. This check is
5430 * made on memory-hotadd so a system can start with mobility
5431 * disabled and enable it later
5433 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5434 page_group_by_mobility_disabled = 1;
5436 page_group_by_mobility_disabled = 0;
5438 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5440 page_group_by_mobility_disabled ? "off" : "on",
5443 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5448 * Initially all pages are reserved - free ones are freed
5449 * up by free_all_bootmem() once the early boot process is
5450 * done. Non-atomic initialization, single-pass.
5452 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5453 unsigned long start_pfn, enum memmap_context context,
5454 struct vmem_altmap *altmap)
5456 unsigned long end_pfn = start_pfn + size;
5457 pg_data_t *pgdat = NODE_DATA(nid);
5459 unsigned long nr_initialised = 0;
5461 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5462 struct memblock_region *r = NULL, *tmp;
5465 if (highest_memmap_pfn < end_pfn - 1)
5466 highest_memmap_pfn = end_pfn - 1;
5469 * Honor reservation requested by the driver for this ZONE_DEVICE
5472 if (altmap && start_pfn == altmap->base_pfn)
5473 start_pfn += altmap->reserve;
5475 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5477 * There can be holes in boot-time mem_map[]s handed to this
5478 * function. They do not exist on hotplugged memory.
5480 if (context != MEMMAP_EARLY)
5483 if (!early_pfn_valid(pfn))
5485 if (!early_pfn_in_nid(pfn, nid))
5487 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5490 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5492 * Check given memblock attribute by firmware which can affect
5493 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5494 * mirrored, it's an overlapped memmap init. skip it.
5496 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5497 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5498 for_each_memblock(memory, tmp)
5499 if (pfn < memblock_region_memory_end_pfn(tmp))
5503 if (pfn >= memblock_region_memory_base_pfn(r) &&
5504 memblock_is_mirror(r)) {
5505 /* already initialized as NORMAL */
5506 pfn = memblock_region_memory_end_pfn(r);
5513 page = pfn_to_page(pfn);
5514 __init_single_page(page, pfn, zone, nid);
5515 if (context == MEMMAP_HOTPLUG)
5516 __SetPageReserved(page);
5519 * Mark the block movable so that blocks are reserved for
5520 * movable at startup. This will force kernel allocations
5521 * to reserve their blocks rather than leaking throughout
5522 * the address space during boot when many long-lived
5523 * kernel allocations are made.
5525 * bitmap is created for zone's valid pfn range. but memmap
5526 * can be created for invalid pages (for alignment)
5527 * check here not to call set_pageblock_migratetype() against
5530 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5531 * because this is done early in sparse_add_one_section
5533 if (!(pfn & (pageblock_nr_pages - 1))) {
5534 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5540 static void __meminit zone_init_free_lists(struct zone *zone)
5542 unsigned int order, t;
5543 for_each_migratetype_order(order, t) {
5544 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5545 zone->free_area[order].nr_free = 0;
5549 #ifndef __HAVE_ARCH_MEMMAP_INIT
5550 #define memmap_init(size, nid, zone, start_pfn) \
5551 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5554 static int zone_batchsize(struct zone *zone)
5560 * The per-cpu-pages pools are set to around 1000th of the
5563 batch = zone->managed_pages / 1024;
5564 /* But no more than a meg. */
5565 if (batch * PAGE_SIZE > 1024 * 1024)
5566 batch = (1024 * 1024) / PAGE_SIZE;
5567 batch /= 4; /* We effectively *= 4 below */
5572 * Clamp the batch to a 2^n - 1 value. Having a power
5573 * of 2 value was found to be more likely to have
5574 * suboptimal cache aliasing properties in some cases.
5576 * For example if 2 tasks are alternately allocating
5577 * batches of pages, one task can end up with a lot
5578 * of pages of one half of the possible page colors
5579 * and the other with pages of the other colors.
5581 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5586 /* The deferral and batching of frees should be suppressed under NOMMU
5589 * The problem is that NOMMU needs to be able to allocate large chunks
5590 * of contiguous memory as there's no hardware page translation to
5591 * assemble apparent contiguous memory from discontiguous pages.
5593 * Queueing large contiguous runs of pages for batching, however,
5594 * causes the pages to actually be freed in smaller chunks. As there
5595 * can be a significant delay between the individual batches being
5596 * recycled, this leads to the once large chunks of space being
5597 * fragmented and becoming unavailable for high-order allocations.
5604 * pcp->high and pcp->batch values are related and dependent on one another:
5605 * ->batch must never be higher then ->high.
5606 * The following function updates them in a safe manner without read side
5609 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5610 * those fields changing asynchronously (acording the the above rule).
5612 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5613 * outside of boot time (or some other assurance that no concurrent updaters
5616 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5617 unsigned long batch)
5619 /* start with a fail safe value for batch */
5623 /* Update high, then batch, in order */
5630 /* a companion to pageset_set_high() */
5631 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5633 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5636 static void pageset_init(struct per_cpu_pageset *p)
5638 struct per_cpu_pages *pcp;
5641 memset(p, 0, sizeof(*p));
5645 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5646 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5649 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5652 pageset_set_batch(p, batch);
5656 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5657 * to the value high for the pageset p.
5659 static void pageset_set_high(struct per_cpu_pageset *p,
5662 unsigned long batch = max(1UL, high / 4);
5663 if ((high / 4) > (PAGE_SHIFT * 8))
5664 batch = PAGE_SHIFT * 8;
5666 pageset_update(&p->pcp, high, batch);
5669 static void pageset_set_high_and_batch(struct zone *zone,
5670 struct per_cpu_pageset *pcp)
5672 if (percpu_pagelist_fraction)
5673 pageset_set_high(pcp,
5674 (zone->managed_pages /
5675 percpu_pagelist_fraction));
5677 pageset_set_batch(pcp, zone_batchsize(zone));
5680 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5682 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5685 pageset_set_high_and_batch(zone, pcp);
5688 void __meminit setup_zone_pageset(struct zone *zone)
5691 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5692 for_each_possible_cpu(cpu)
5693 zone_pageset_init(zone, cpu);
5697 * Allocate per cpu pagesets and initialize them.
5698 * Before this call only boot pagesets were available.
5700 void __init setup_per_cpu_pageset(void)
5702 struct pglist_data *pgdat;
5705 for_each_populated_zone(zone)
5706 setup_zone_pageset(zone);
5708 for_each_online_pgdat(pgdat)
5709 pgdat->per_cpu_nodestats =
5710 alloc_percpu(struct per_cpu_nodestat);
5713 static __meminit void zone_pcp_init(struct zone *zone)
5716 * per cpu subsystem is not up at this point. The following code
5717 * relies on the ability of the linker to provide the
5718 * offset of a (static) per cpu variable into the per cpu area.
5720 zone->pageset = &boot_pageset;
5722 if (populated_zone(zone))
5723 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5724 zone->name, zone->present_pages,
5725 zone_batchsize(zone));
5728 void __meminit init_currently_empty_zone(struct zone *zone,
5729 unsigned long zone_start_pfn,
5732 struct pglist_data *pgdat = zone->zone_pgdat;
5734 pgdat->nr_zones = zone_idx(zone) + 1;
5736 zone->zone_start_pfn = zone_start_pfn;
5738 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5739 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5741 (unsigned long)zone_idx(zone),
5742 zone_start_pfn, (zone_start_pfn + size));
5744 zone_init_free_lists(zone);
5745 zone->initialized = 1;
5748 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5749 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5752 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5754 int __meminit __early_pfn_to_nid(unsigned long pfn,
5755 struct mminit_pfnnid_cache *state)
5757 unsigned long start_pfn, end_pfn;
5760 if (state->last_start <= pfn && pfn < state->last_end)
5761 return state->last_nid;
5763 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5765 state->last_start = start_pfn;
5766 state->last_end = end_pfn;
5767 state->last_nid = nid;
5772 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5775 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5776 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5777 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5779 * If an architecture guarantees that all ranges registered contain no holes
5780 * and may be freed, this this function may be used instead of calling
5781 * memblock_free_early_nid() manually.
5783 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5785 unsigned long start_pfn, end_pfn;
5788 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5789 start_pfn = min(start_pfn, max_low_pfn);
5790 end_pfn = min(end_pfn, max_low_pfn);
5792 if (start_pfn < end_pfn)
5793 memblock_free_early_nid(PFN_PHYS(start_pfn),
5794 (end_pfn - start_pfn) << PAGE_SHIFT,
5800 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5801 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5803 * If an architecture guarantees that all ranges registered contain no holes and may
5804 * be freed, this function may be used instead of calling memory_present() manually.
5806 void __init sparse_memory_present_with_active_regions(int nid)
5808 unsigned long start_pfn, end_pfn;
5811 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5812 memory_present(this_nid, start_pfn, end_pfn);
5816 * get_pfn_range_for_nid - Return the start and end page frames for a node
5817 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5818 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5819 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5821 * It returns the start and end page frame of a node based on information
5822 * provided by memblock_set_node(). If called for a node
5823 * with no available memory, a warning is printed and the start and end
5826 void __meminit get_pfn_range_for_nid(unsigned int nid,
5827 unsigned long *start_pfn, unsigned long *end_pfn)
5829 unsigned long this_start_pfn, this_end_pfn;
5835 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5836 *start_pfn = min(*start_pfn, this_start_pfn);
5837 *end_pfn = max(*end_pfn, this_end_pfn);
5840 if (*start_pfn == -1UL)
5845 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5846 * assumption is made that zones within a node are ordered in monotonic
5847 * increasing memory addresses so that the "highest" populated zone is used
5849 static void __init find_usable_zone_for_movable(void)
5852 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5853 if (zone_index == ZONE_MOVABLE)
5856 if (arch_zone_highest_possible_pfn[zone_index] >
5857 arch_zone_lowest_possible_pfn[zone_index])
5861 VM_BUG_ON(zone_index == -1);
5862 movable_zone = zone_index;
5866 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5867 * because it is sized independent of architecture. Unlike the other zones,
5868 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5869 * in each node depending on the size of each node and how evenly kernelcore
5870 * is distributed. This helper function adjusts the zone ranges
5871 * provided by the architecture for a given node by using the end of the
5872 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5873 * zones within a node are in order of monotonic increases memory addresses
5875 static void __meminit adjust_zone_range_for_zone_movable(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)
5882 /* Only adjust if ZONE_MOVABLE is on this node */
5883 if (zone_movable_pfn[nid]) {
5884 /* Size ZONE_MOVABLE */
5885 if (zone_type == ZONE_MOVABLE) {
5886 *zone_start_pfn = zone_movable_pfn[nid];
5887 *zone_end_pfn = min(node_end_pfn,
5888 arch_zone_highest_possible_pfn[movable_zone]);
5890 /* Adjust for ZONE_MOVABLE starting within this range */
5891 } else if (!mirrored_kernelcore &&
5892 *zone_start_pfn < zone_movable_pfn[nid] &&
5893 *zone_end_pfn > zone_movable_pfn[nid]) {
5894 *zone_end_pfn = zone_movable_pfn[nid];
5896 /* Check if this whole range is within ZONE_MOVABLE */
5897 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5898 *zone_start_pfn = *zone_end_pfn;
5903 * Return the number of pages a zone spans in a node, including holes
5904 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5906 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5907 unsigned long zone_type,
5908 unsigned long node_start_pfn,
5909 unsigned long node_end_pfn,
5910 unsigned long *zone_start_pfn,
5911 unsigned long *zone_end_pfn,
5912 unsigned long *ignored)
5914 /* When hotadd a new node from cpu_up(), the node should be empty */
5915 if (!node_start_pfn && !node_end_pfn)
5918 /* Get the start and end of the zone */
5919 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5920 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5921 adjust_zone_range_for_zone_movable(nid, zone_type,
5922 node_start_pfn, node_end_pfn,
5923 zone_start_pfn, zone_end_pfn);
5925 /* Check that this node has pages within the zone's required range */
5926 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5929 /* Move the zone boundaries inside the node if necessary */
5930 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5931 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5933 /* Return the spanned pages */
5934 return *zone_end_pfn - *zone_start_pfn;
5938 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5939 * then all holes in the requested range will be accounted for.
5941 unsigned long __meminit __absent_pages_in_range(int nid,
5942 unsigned long range_start_pfn,
5943 unsigned long range_end_pfn)
5945 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5946 unsigned long start_pfn, end_pfn;
5949 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5950 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5951 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5952 nr_absent -= end_pfn - start_pfn;
5958 * absent_pages_in_range - Return number of page frames in holes within a range
5959 * @start_pfn: The start PFN to start searching for holes
5960 * @end_pfn: The end PFN to stop searching for holes
5962 * It returns the number of pages frames in memory holes within a range.
5964 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5965 unsigned long end_pfn)
5967 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5970 /* Return the number of page frames in holes in a zone on a node */
5971 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5972 unsigned long zone_type,
5973 unsigned long node_start_pfn,
5974 unsigned long node_end_pfn,
5975 unsigned long *ignored)
5977 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5978 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5979 unsigned long zone_start_pfn, zone_end_pfn;
5980 unsigned long nr_absent;
5982 /* When hotadd a new node from cpu_up(), the node should be empty */
5983 if (!node_start_pfn && !node_end_pfn)
5986 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5987 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5989 adjust_zone_range_for_zone_movable(nid, zone_type,
5990 node_start_pfn, node_end_pfn,
5991 &zone_start_pfn, &zone_end_pfn);
5992 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5995 * ZONE_MOVABLE handling.
5996 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5999 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6000 unsigned long start_pfn, end_pfn;
6001 struct memblock_region *r;
6003 for_each_memblock(memory, r) {
6004 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6005 zone_start_pfn, zone_end_pfn);
6006 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6007 zone_start_pfn, zone_end_pfn);
6009 if (zone_type == ZONE_MOVABLE &&
6010 memblock_is_mirror(r))
6011 nr_absent += end_pfn - start_pfn;
6013 if (zone_type == ZONE_NORMAL &&
6014 !memblock_is_mirror(r))
6015 nr_absent += end_pfn - start_pfn;
6022 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6023 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6024 unsigned long zone_type,
6025 unsigned long node_start_pfn,
6026 unsigned long node_end_pfn,
6027 unsigned long *zone_start_pfn,
6028 unsigned long *zone_end_pfn,
6029 unsigned long *zones_size)
6033 *zone_start_pfn = node_start_pfn;
6034 for (zone = 0; zone < zone_type; zone++)
6035 *zone_start_pfn += zones_size[zone];
6037 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6039 return zones_size[zone_type];
6042 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6043 unsigned long zone_type,
6044 unsigned long node_start_pfn,
6045 unsigned long node_end_pfn,
6046 unsigned long *zholes_size)
6051 return zholes_size[zone_type];
6054 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6056 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6057 unsigned long node_start_pfn,
6058 unsigned long node_end_pfn,
6059 unsigned long *zones_size,
6060 unsigned long *zholes_size)
6062 unsigned long realtotalpages = 0, totalpages = 0;
6065 for (i = 0; i < MAX_NR_ZONES; i++) {
6066 struct zone *zone = pgdat->node_zones + i;
6067 unsigned long zone_start_pfn, zone_end_pfn;
6068 unsigned long size, real_size;
6070 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6076 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6077 node_start_pfn, node_end_pfn,
6080 zone->zone_start_pfn = zone_start_pfn;
6082 zone->zone_start_pfn = 0;
6083 zone->spanned_pages = size;
6084 zone->present_pages = real_size;
6087 realtotalpages += real_size;
6090 pgdat->node_spanned_pages = totalpages;
6091 pgdat->node_present_pages = realtotalpages;
6092 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6096 #ifndef CONFIG_SPARSEMEM
6098 * Calculate the size of the zone->blockflags rounded to an unsigned long
6099 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6100 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6101 * round what is now in bits to nearest long in bits, then return it in
6104 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6106 unsigned long usemapsize;
6108 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6109 usemapsize = roundup(zonesize, pageblock_nr_pages);
6110 usemapsize = usemapsize >> pageblock_order;
6111 usemapsize *= NR_PAGEBLOCK_BITS;
6112 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6114 return usemapsize / 8;
6117 static void __ref setup_usemap(struct pglist_data *pgdat,
6119 unsigned long zone_start_pfn,
6120 unsigned long zonesize)
6122 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6123 zone->pageblock_flags = NULL;
6125 zone->pageblock_flags =
6126 memblock_virt_alloc_node_nopanic(usemapsize,
6130 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6131 unsigned long zone_start_pfn, unsigned long zonesize) {}
6132 #endif /* CONFIG_SPARSEMEM */
6134 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6136 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6137 void __init set_pageblock_order(void)
6141 /* Check that pageblock_nr_pages has not already been setup */
6142 if (pageblock_order)
6145 if (HPAGE_SHIFT > PAGE_SHIFT)
6146 order = HUGETLB_PAGE_ORDER;
6148 order = MAX_ORDER - 1;
6151 * Assume the largest contiguous order of interest is a huge page.
6152 * This value may be variable depending on boot parameters on IA64 and
6155 pageblock_order = order;
6157 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6160 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6161 * is unused as pageblock_order is set at compile-time. See
6162 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6165 void __init set_pageblock_order(void)
6169 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6171 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6172 unsigned long present_pages)
6174 unsigned long pages = spanned_pages;
6177 * Provide a more accurate estimation if there are holes within
6178 * the zone and SPARSEMEM is in use. If there are holes within the
6179 * zone, each populated memory region may cost us one or two extra
6180 * memmap pages due to alignment because memmap pages for each
6181 * populated regions may not be naturally aligned on page boundary.
6182 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6184 if (spanned_pages > present_pages + (present_pages >> 4) &&
6185 IS_ENABLED(CONFIG_SPARSEMEM))
6186 pages = present_pages;
6188 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6191 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6192 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6194 spin_lock_init(&pgdat->split_queue_lock);
6195 INIT_LIST_HEAD(&pgdat->split_queue);
6196 pgdat->split_queue_len = 0;
6199 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6202 #ifdef CONFIG_COMPACTION
6203 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6205 init_waitqueue_head(&pgdat->kcompactd_wait);
6208 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6211 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6213 pgdat_resize_init(pgdat);
6215 pgdat_init_split_queue(pgdat);
6216 pgdat_init_kcompactd(pgdat);
6218 init_waitqueue_head(&pgdat->kswapd_wait);
6219 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6221 pgdat_page_ext_init(pgdat);
6222 spin_lock_init(&pgdat->lru_lock);
6223 lruvec_init(node_lruvec(pgdat));
6226 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6227 unsigned long remaining_pages)
6229 zone->managed_pages = remaining_pages;
6230 zone_set_nid(zone, nid);
6231 zone->name = zone_names[idx];
6232 zone->zone_pgdat = NODE_DATA(nid);
6233 spin_lock_init(&zone->lock);
6234 zone_seqlock_init(zone);
6235 zone_pcp_init(zone);
6239 * Set up the zone data structures
6240 * - init pgdat internals
6241 * - init all zones belonging to this node
6243 * NOTE: this function is only called during memory hotplug
6245 #ifdef CONFIG_MEMORY_HOTPLUG
6246 void __ref free_area_init_core_hotplug(int nid)
6249 pg_data_t *pgdat = NODE_DATA(nid);
6251 pgdat_init_internals(pgdat);
6252 for (z = 0; z < MAX_NR_ZONES; z++)
6253 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6258 * Set up the zone data structures:
6259 * - mark all pages reserved
6260 * - mark all memory queues empty
6261 * - clear the memory bitmaps
6263 * NOTE: pgdat should get zeroed by caller.
6264 * NOTE: this function is only called during early init.
6266 static void __init free_area_init_core(struct pglist_data *pgdat)
6269 int nid = pgdat->node_id;
6271 pgdat_init_internals(pgdat);
6272 pgdat->per_cpu_nodestats = &boot_nodestats;
6274 for (j = 0; j < MAX_NR_ZONES; j++) {
6275 struct zone *zone = pgdat->node_zones + j;
6276 unsigned long size, freesize, memmap_pages;
6277 unsigned long zone_start_pfn = zone->zone_start_pfn;
6279 size = zone->spanned_pages;
6280 freesize = zone->present_pages;
6283 * Adjust freesize so that it accounts for how much memory
6284 * is used by this zone for memmap. This affects the watermark
6285 * and per-cpu initialisations
6287 memmap_pages = calc_memmap_size(size, freesize);
6288 if (!is_highmem_idx(j)) {
6289 if (freesize >= memmap_pages) {
6290 freesize -= memmap_pages;
6293 " %s zone: %lu pages used for memmap\n",
6294 zone_names[j], memmap_pages);
6296 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6297 zone_names[j], memmap_pages, freesize);
6300 /* Account for reserved pages */
6301 if (j == 0 && freesize > dma_reserve) {
6302 freesize -= dma_reserve;
6303 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6304 zone_names[0], dma_reserve);
6307 if (!is_highmem_idx(j))
6308 nr_kernel_pages += freesize;
6309 /* Charge for highmem memmap if there are enough kernel pages */
6310 else if (nr_kernel_pages > memmap_pages * 2)
6311 nr_kernel_pages -= memmap_pages;
6312 nr_all_pages += freesize;
6315 * Set an approximate value for lowmem here, it will be adjusted
6316 * when the bootmem allocator frees pages into the buddy system.
6317 * And all highmem pages will be managed by the buddy system.
6319 zone_init_internals(zone, j, nid, freesize);
6324 set_pageblock_order();
6325 setup_usemap(pgdat, zone, zone_start_pfn, size);
6326 init_currently_empty_zone(zone, zone_start_pfn, size);
6327 memmap_init(size, nid, j, zone_start_pfn);
6331 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6332 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6334 unsigned long __maybe_unused start = 0;
6335 unsigned long __maybe_unused offset = 0;
6337 /* Skip empty nodes */
6338 if (!pgdat->node_spanned_pages)
6341 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6342 offset = pgdat->node_start_pfn - start;
6343 /* ia64 gets its own node_mem_map, before this, without bootmem */
6344 if (!pgdat->node_mem_map) {
6345 unsigned long size, end;
6349 * The zone's endpoints aren't required to be MAX_ORDER
6350 * aligned but the node_mem_map endpoints must be in order
6351 * for the buddy allocator to function correctly.
6353 end = pgdat_end_pfn(pgdat);
6354 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6355 size = (end - start) * sizeof(struct page);
6356 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6357 pgdat->node_mem_map = map + offset;
6359 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6360 __func__, pgdat->node_id, (unsigned long)pgdat,
6361 (unsigned long)pgdat->node_mem_map);
6362 #ifndef CONFIG_NEED_MULTIPLE_NODES
6364 * With no DISCONTIG, the global mem_map is just set as node 0's
6366 if (pgdat == NODE_DATA(0)) {
6367 mem_map = NODE_DATA(0)->node_mem_map;
6368 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6369 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6371 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6376 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6377 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6379 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6380 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6383 * We start only with one section of pages, more pages are added as
6384 * needed until the rest of deferred pages are initialized.
6386 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6387 pgdat->node_spanned_pages);
6388 pgdat->first_deferred_pfn = ULONG_MAX;
6391 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6394 void __init free_area_init_node(int nid, unsigned long *zones_size,
6395 unsigned long node_start_pfn,
6396 unsigned long *zholes_size)
6398 pg_data_t *pgdat = NODE_DATA(nid);
6399 unsigned long start_pfn = 0;
6400 unsigned long end_pfn = 0;
6402 /* pg_data_t should be reset to zero when it's allocated */
6403 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6405 pgdat->node_id = nid;
6406 pgdat->node_start_pfn = node_start_pfn;
6407 pgdat->per_cpu_nodestats = NULL;
6408 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6409 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6410 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6411 (u64)start_pfn << PAGE_SHIFT,
6412 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6414 start_pfn = node_start_pfn;
6416 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6417 zones_size, zholes_size);
6419 alloc_node_mem_map(pgdat);
6420 pgdat_set_deferred_range(pgdat);
6422 free_area_init_core(pgdat);
6425 #if defined(CONFIG_HAVE_MEMBLOCK) && !defined(CONFIG_FLAT_NODE_MEM_MAP)
6427 * Only struct pages that are backed by physical memory are zeroed and
6428 * initialized by going through __init_single_page(). But, there are some
6429 * struct pages which are reserved in memblock allocator and their fields
6430 * may be accessed (for example page_to_pfn() on some configuration accesses
6431 * flags). We must explicitly zero those struct pages.
6433 void __init zero_resv_unavail(void)
6435 phys_addr_t start, end;
6440 * Loop through ranges that are reserved, but do not have reported
6441 * physical memory backing.
6444 for_each_resv_unavail_range(i, &start, &end) {
6445 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6446 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6447 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6448 + pageblock_nr_pages - 1;
6451 mm_zero_struct_page(pfn_to_page(pfn));
6457 * Struct pages that do not have backing memory. This could be because
6458 * firmware is using some of this memory, or for some other reasons.
6459 * Once memblock is changed so such behaviour is not allowed: i.e.
6460 * list of "reserved" memory must be a subset of list of "memory", then
6461 * this code can be removed.
6464 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6466 #endif /* CONFIG_HAVE_MEMBLOCK && !CONFIG_FLAT_NODE_MEM_MAP */
6468 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6470 #if MAX_NUMNODES > 1
6472 * Figure out the number of possible node ids.
6474 void __init setup_nr_node_ids(void)
6476 unsigned int highest;
6478 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6479 nr_node_ids = highest + 1;
6484 * node_map_pfn_alignment - determine the maximum internode alignment
6486 * This function should be called after node map is populated and sorted.
6487 * It calculates the maximum power of two alignment which can distinguish
6490 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6491 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6492 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6493 * shifted, 1GiB is enough and this function will indicate so.
6495 * This is used to test whether pfn -> nid mapping of the chosen memory
6496 * model has fine enough granularity to avoid incorrect mapping for the
6497 * populated node map.
6499 * Returns the determined alignment in pfn's. 0 if there is no alignment
6500 * requirement (single node).
6502 unsigned long __init node_map_pfn_alignment(void)
6504 unsigned long accl_mask = 0, last_end = 0;
6505 unsigned long start, end, mask;
6509 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6510 if (!start || last_nid < 0 || last_nid == nid) {
6517 * Start with a mask granular enough to pin-point to the
6518 * start pfn and tick off bits one-by-one until it becomes
6519 * too coarse to separate the current node from the last.
6521 mask = ~((1 << __ffs(start)) - 1);
6522 while (mask && last_end <= (start & (mask << 1)))
6525 /* accumulate all internode masks */
6529 /* convert mask to number of pages */
6530 return ~accl_mask + 1;
6533 /* Find the lowest pfn for a node */
6534 static unsigned long __init find_min_pfn_for_node(int nid)
6536 unsigned long min_pfn = ULONG_MAX;
6537 unsigned long start_pfn;
6540 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6541 min_pfn = min(min_pfn, start_pfn);
6543 if (min_pfn == ULONG_MAX) {
6544 pr_warn("Could not find start_pfn for node %d\n", nid);
6552 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6554 * It returns the minimum PFN based on information provided via
6555 * memblock_set_node().
6557 unsigned long __init find_min_pfn_with_active_regions(void)
6559 return find_min_pfn_for_node(MAX_NUMNODES);
6563 * early_calculate_totalpages()
6564 * Sum pages in active regions for movable zone.
6565 * Populate N_MEMORY for calculating usable_nodes.
6567 static unsigned long __init early_calculate_totalpages(void)
6569 unsigned long totalpages = 0;
6570 unsigned long start_pfn, end_pfn;
6573 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6574 unsigned long pages = end_pfn - start_pfn;
6576 totalpages += pages;
6578 node_set_state(nid, N_MEMORY);
6584 * Find the PFN the Movable zone begins in each node. Kernel memory
6585 * is spread evenly between nodes as long as the nodes have enough
6586 * memory. When they don't, some nodes will have more kernelcore than
6589 static void __init find_zone_movable_pfns_for_nodes(void)
6592 unsigned long usable_startpfn;
6593 unsigned long kernelcore_node, kernelcore_remaining;
6594 /* save the state before borrow the nodemask */
6595 nodemask_t saved_node_state = node_states[N_MEMORY];
6596 unsigned long totalpages = early_calculate_totalpages();
6597 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6598 struct memblock_region *r;
6600 /* Need to find movable_zone earlier when movable_node is specified. */
6601 find_usable_zone_for_movable();
6604 * If movable_node is specified, ignore kernelcore and movablecore
6607 if (movable_node_is_enabled()) {
6608 for_each_memblock(memory, r) {
6609 if (!memblock_is_hotpluggable(r))
6614 usable_startpfn = PFN_DOWN(r->base);
6615 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6616 min(usable_startpfn, zone_movable_pfn[nid]) :
6624 * If kernelcore=mirror is specified, ignore movablecore option
6626 if (mirrored_kernelcore) {
6627 bool mem_below_4gb_not_mirrored = false;
6629 for_each_memblock(memory, r) {
6630 if (memblock_is_mirror(r))
6635 usable_startpfn = memblock_region_memory_base_pfn(r);
6637 if (usable_startpfn < 0x100000) {
6638 mem_below_4gb_not_mirrored = true;
6642 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6643 min(usable_startpfn, zone_movable_pfn[nid]) :
6647 if (mem_below_4gb_not_mirrored)
6648 pr_warn("This configuration results in unmirrored kernel memory.");
6654 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6655 * amount of necessary memory.
6657 if (required_kernelcore_percent)
6658 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6660 if (required_movablecore_percent)
6661 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6665 * If movablecore= was specified, calculate what size of
6666 * kernelcore that corresponds so that memory usable for
6667 * any allocation type is evenly spread. If both kernelcore
6668 * and movablecore are specified, then the value of kernelcore
6669 * will be used for required_kernelcore if it's greater than
6670 * what movablecore would have allowed.
6672 if (required_movablecore) {
6673 unsigned long corepages;
6676 * Round-up so that ZONE_MOVABLE is at least as large as what
6677 * was requested by the user
6679 required_movablecore =
6680 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6681 required_movablecore = min(totalpages, required_movablecore);
6682 corepages = totalpages - required_movablecore;
6684 required_kernelcore = max(required_kernelcore, corepages);
6688 * If kernelcore was not specified or kernelcore size is larger
6689 * than totalpages, there is no ZONE_MOVABLE.
6691 if (!required_kernelcore || required_kernelcore >= totalpages)
6694 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6695 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6698 /* Spread kernelcore memory as evenly as possible throughout nodes */
6699 kernelcore_node = required_kernelcore / usable_nodes;
6700 for_each_node_state(nid, N_MEMORY) {
6701 unsigned long start_pfn, end_pfn;
6704 * Recalculate kernelcore_node if the division per node
6705 * now exceeds what is necessary to satisfy the requested
6706 * amount of memory for the kernel
6708 if (required_kernelcore < kernelcore_node)
6709 kernelcore_node = required_kernelcore / usable_nodes;
6712 * As the map is walked, we track how much memory is usable
6713 * by the kernel using kernelcore_remaining. When it is
6714 * 0, the rest of the node is usable by ZONE_MOVABLE
6716 kernelcore_remaining = kernelcore_node;
6718 /* Go through each range of PFNs within this node */
6719 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6720 unsigned long size_pages;
6722 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6723 if (start_pfn >= end_pfn)
6726 /* Account for what is only usable for kernelcore */
6727 if (start_pfn < usable_startpfn) {
6728 unsigned long kernel_pages;
6729 kernel_pages = min(end_pfn, usable_startpfn)
6732 kernelcore_remaining -= min(kernel_pages,
6733 kernelcore_remaining);
6734 required_kernelcore -= min(kernel_pages,
6735 required_kernelcore);
6737 /* Continue if range is now fully accounted */
6738 if (end_pfn <= usable_startpfn) {
6741 * Push zone_movable_pfn to the end so
6742 * that if we have to rebalance
6743 * kernelcore across nodes, we will
6744 * not double account here
6746 zone_movable_pfn[nid] = end_pfn;
6749 start_pfn = usable_startpfn;
6753 * The usable PFN range for ZONE_MOVABLE is from
6754 * start_pfn->end_pfn. Calculate size_pages as the
6755 * number of pages used as kernelcore
6757 size_pages = end_pfn - start_pfn;
6758 if (size_pages > kernelcore_remaining)
6759 size_pages = kernelcore_remaining;
6760 zone_movable_pfn[nid] = start_pfn + size_pages;
6763 * Some kernelcore has been met, update counts and
6764 * break if the kernelcore for this node has been
6767 required_kernelcore -= min(required_kernelcore,
6769 kernelcore_remaining -= size_pages;
6770 if (!kernelcore_remaining)
6776 * If there is still required_kernelcore, we do another pass with one
6777 * less node in the count. This will push zone_movable_pfn[nid] further
6778 * along on the nodes that still have memory until kernelcore is
6782 if (usable_nodes && required_kernelcore > usable_nodes)
6786 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6787 for (nid = 0; nid < MAX_NUMNODES; nid++)
6788 zone_movable_pfn[nid] =
6789 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6792 /* restore the node_state */
6793 node_states[N_MEMORY] = saved_node_state;
6796 /* Any regular or high memory on that node ? */
6797 static void check_for_memory(pg_data_t *pgdat, int nid)
6799 enum zone_type zone_type;
6801 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6802 struct zone *zone = &pgdat->node_zones[zone_type];
6803 if (populated_zone(zone)) {
6804 if (IS_ENABLED(CONFIG_HIGHMEM))
6805 node_set_state(nid, N_HIGH_MEMORY);
6806 if (zone_type <= ZONE_NORMAL)
6807 node_set_state(nid, N_NORMAL_MEMORY);
6814 * free_area_init_nodes - Initialise all pg_data_t and zone data
6815 * @max_zone_pfn: an array of max PFNs for each zone
6817 * This will call free_area_init_node() for each active node in the system.
6818 * Using the page ranges provided by memblock_set_node(), the size of each
6819 * zone in each node and their holes is calculated. If the maximum PFN
6820 * between two adjacent zones match, it is assumed that the zone is empty.
6821 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6822 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6823 * starts where the previous one ended. For example, ZONE_DMA32 starts
6824 * at arch_max_dma_pfn.
6826 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6828 unsigned long start_pfn, end_pfn;
6831 /* Record where the zone boundaries are */
6832 memset(arch_zone_lowest_possible_pfn, 0,
6833 sizeof(arch_zone_lowest_possible_pfn));
6834 memset(arch_zone_highest_possible_pfn, 0,
6835 sizeof(arch_zone_highest_possible_pfn));
6837 start_pfn = find_min_pfn_with_active_regions();
6839 for (i = 0; i < MAX_NR_ZONES; i++) {
6840 if (i == ZONE_MOVABLE)
6843 end_pfn = max(max_zone_pfn[i], start_pfn);
6844 arch_zone_lowest_possible_pfn[i] = start_pfn;
6845 arch_zone_highest_possible_pfn[i] = end_pfn;
6847 start_pfn = end_pfn;
6850 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6851 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6852 find_zone_movable_pfns_for_nodes();
6854 /* Print out the zone ranges */
6855 pr_info("Zone ranges:\n");
6856 for (i = 0; i < MAX_NR_ZONES; i++) {
6857 if (i == ZONE_MOVABLE)
6859 pr_info(" %-8s ", zone_names[i]);
6860 if (arch_zone_lowest_possible_pfn[i] ==
6861 arch_zone_highest_possible_pfn[i])
6864 pr_cont("[mem %#018Lx-%#018Lx]\n",
6865 (u64)arch_zone_lowest_possible_pfn[i]
6867 ((u64)arch_zone_highest_possible_pfn[i]
6868 << PAGE_SHIFT) - 1);
6871 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6872 pr_info("Movable zone start for each node\n");
6873 for (i = 0; i < MAX_NUMNODES; i++) {
6874 if (zone_movable_pfn[i])
6875 pr_info(" Node %d: %#018Lx\n", i,
6876 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6879 /* Print out the early node map */
6880 pr_info("Early memory node ranges\n");
6881 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6882 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6883 (u64)start_pfn << PAGE_SHIFT,
6884 ((u64)end_pfn << PAGE_SHIFT) - 1);
6886 /* Initialise every node */
6887 mminit_verify_pageflags_layout();
6888 setup_nr_node_ids();
6889 zero_resv_unavail();
6890 for_each_online_node(nid) {
6891 pg_data_t *pgdat = NODE_DATA(nid);
6892 free_area_init_node(nid, NULL,
6893 find_min_pfn_for_node(nid), NULL);
6895 /* Any memory on that node */
6896 if (pgdat->node_present_pages)
6897 node_set_state(nid, N_MEMORY);
6898 check_for_memory(pgdat, nid);
6902 static int __init cmdline_parse_core(char *p, unsigned long *core,
6903 unsigned long *percent)
6905 unsigned long long coremem;
6911 /* Value may be a percentage of total memory, otherwise bytes */
6912 coremem = simple_strtoull(p, &endptr, 0);
6913 if (*endptr == '%') {
6914 /* Paranoid check for percent values greater than 100 */
6915 WARN_ON(coremem > 100);
6919 coremem = memparse(p, &p);
6920 /* Paranoid check that UL is enough for the coremem value */
6921 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6923 *core = coremem >> PAGE_SHIFT;
6930 * kernelcore=size sets the amount of memory for use for allocations that
6931 * cannot be reclaimed or migrated.
6933 static int __init cmdline_parse_kernelcore(char *p)
6935 /* parse kernelcore=mirror */
6936 if (parse_option_str(p, "mirror")) {
6937 mirrored_kernelcore = true;
6941 return cmdline_parse_core(p, &required_kernelcore,
6942 &required_kernelcore_percent);
6946 * movablecore=size sets the amount of memory for use for allocations that
6947 * can be reclaimed or migrated.
6949 static int __init cmdline_parse_movablecore(char *p)
6951 return cmdline_parse_core(p, &required_movablecore,
6952 &required_movablecore_percent);
6955 early_param("kernelcore", cmdline_parse_kernelcore);
6956 early_param("movablecore", cmdline_parse_movablecore);
6958 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6960 void adjust_managed_page_count(struct page *page, long count)
6962 spin_lock(&managed_page_count_lock);
6963 page_zone(page)->managed_pages += count;
6964 totalram_pages += count;
6965 #ifdef CONFIG_HIGHMEM
6966 if (PageHighMem(page))
6967 totalhigh_pages += count;
6969 spin_unlock(&managed_page_count_lock);
6971 EXPORT_SYMBOL(adjust_managed_page_count);
6973 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6976 unsigned long pages = 0;
6978 start = (void *)PAGE_ALIGN((unsigned long)start);
6979 end = (void *)((unsigned long)end & PAGE_MASK);
6980 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6981 struct page *page = virt_to_page(pos);
6982 void *direct_map_addr;
6985 * 'direct_map_addr' might be different from 'pos'
6986 * because some architectures' virt_to_page()
6987 * work with aliases. Getting the direct map
6988 * address ensures that we get a _writeable_
6989 * alias for the memset().
6991 direct_map_addr = page_address(page);
6992 if ((unsigned int)poison <= 0xFF)
6993 memset(direct_map_addr, poison, PAGE_SIZE);
6995 free_reserved_page(page);
6999 pr_info("Freeing %s memory: %ldK\n",
7000 s, pages << (PAGE_SHIFT - 10));
7004 EXPORT_SYMBOL(free_reserved_area);
7006 #ifdef CONFIG_HIGHMEM
7007 void free_highmem_page(struct page *page)
7009 __free_reserved_page(page);
7011 page_zone(page)->managed_pages++;
7017 void __init mem_init_print_info(const char *str)
7019 unsigned long physpages, codesize, datasize, rosize, bss_size;
7020 unsigned long init_code_size, init_data_size;
7022 physpages = get_num_physpages();
7023 codesize = _etext - _stext;
7024 datasize = _edata - _sdata;
7025 rosize = __end_rodata - __start_rodata;
7026 bss_size = __bss_stop - __bss_start;
7027 init_data_size = __init_end - __init_begin;
7028 init_code_size = _einittext - _sinittext;
7031 * Detect special cases and adjust section sizes accordingly:
7032 * 1) .init.* may be embedded into .data sections
7033 * 2) .init.text.* may be out of [__init_begin, __init_end],
7034 * please refer to arch/tile/kernel/vmlinux.lds.S.
7035 * 3) .rodata.* may be embedded into .text or .data sections.
7037 #define adj_init_size(start, end, size, pos, adj) \
7039 if (start <= pos && pos < end && size > adj) \
7043 adj_init_size(__init_begin, __init_end, init_data_size,
7044 _sinittext, init_code_size);
7045 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7046 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7047 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7048 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7050 #undef adj_init_size
7052 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7053 #ifdef CONFIG_HIGHMEM
7057 nr_free_pages() << (PAGE_SHIFT - 10),
7058 physpages << (PAGE_SHIFT - 10),
7059 codesize >> 10, datasize >> 10, rosize >> 10,
7060 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7061 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7062 totalcma_pages << (PAGE_SHIFT - 10),
7063 #ifdef CONFIG_HIGHMEM
7064 totalhigh_pages << (PAGE_SHIFT - 10),
7066 str ? ", " : "", str ? str : "");
7070 * set_dma_reserve - set the specified number of pages reserved in the first zone
7071 * @new_dma_reserve: The number of pages to mark reserved
7073 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7074 * In the DMA zone, a significant percentage may be consumed by kernel image
7075 * and other unfreeable allocations which can skew the watermarks badly. This
7076 * function may optionally be used to account for unfreeable pages in the
7077 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7078 * smaller per-cpu batchsize.
7080 void __init set_dma_reserve(unsigned long new_dma_reserve)
7082 dma_reserve = new_dma_reserve;
7085 void __init free_area_init(unsigned long *zones_size)
7087 zero_resv_unavail();
7088 free_area_init_node(0, zones_size,
7089 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7092 static int page_alloc_cpu_dead(unsigned int cpu)
7095 lru_add_drain_cpu(cpu);
7099 * Spill the event counters of the dead processor
7100 * into the current processors event counters.
7101 * This artificially elevates the count of the current
7104 vm_events_fold_cpu(cpu);
7107 * Zero the differential counters of the dead processor
7108 * so that the vm statistics are consistent.
7110 * This is only okay since the processor is dead and cannot
7111 * race with what we are doing.
7113 cpu_vm_stats_fold(cpu);
7117 void __init page_alloc_init(void)
7121 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7122 "mm/page_alloc:dead", NULL,
7123 page_alloc_cpu_dead);
7128 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7129 * or min_free_kbytes changes.
7131 static void calculate_totalreserve_pages(void)
7133 struct pglist_data *pgdat;
7134 unsigned long reserve_pages = 0;
7135 enum zone_type i, j;
7137 for_each_online_pgdat(pgdat) {
7139 pgdat->totalreserve_pages = 0;
7141 for (i = 0; i < MAX_NR_ZONES; i++) {
7142 struct zone *zone = pgdat->node_zones + i;
7145 /* Find valid and maximum lowmem_reserve in the zone */
7146 for (j = i; j < MAX_NR_ZONES; j++) {
7147 if (zone->lowmem_reserve[j] > max)
7148 max = zone->lowmem_reserve[j];
7151 /* we treat the high watermark as reserved pages. */
7152 max += high_wmark_pages(zone);
7154 if (max > zone->managed_pages)
7155 max = zone->managed_pages;
7157 pgdat->totalreserve_pages += max;
7159 reserve_pages += max;
7162 totalreserve_pages = reserve_pages;
7166 * setup_per_zone_lowmem_reserve - called whenever
7167 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7168 * has a correct pages reserved value, so an adequate number of
7169 * pages are left in the zone after a successful __alloc_pages().
7171 static void setup_per_zone_lowmem_reserve(void)
7173 struct pglist_data *pgdat;
7174 enum zone_type j, idx;
7176 for_each_online_pgdat(pgdat) {
7177 for (j = 0; j < MAX_NR_ZONES; j++) {
7178 struct zone *zone = pgdat->node_zones + j;
7179 unsigned long managed_pages = zone->managed_pages;
7181 zone->lowmem_reserve[j] = 0;
7185 struct zone *lower_zone;
7188 lower_zone = pgdat->node_zones + idx;
7190 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7191 sysctl_lowmem_reserve_ratio[idx] = 0;
7192 lower_zone->lowmem_reserve[j] = 0;
7194 lower_zone->lowmem_reserve[j] =
7195 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7197 managed_pages += lower_zone->managed_pages;
7202 /* update totalreserve_pages */
7203 calculate_totalreserve_pages();
7206 static void __setup_per_zone_wmarks(void)
7208 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7209 unsigned long lowmem_pages = 0;
7211 unsigned long flags;
7213 /* Calculate total number of !ZONE_HIGHMEM pages */
7214 for_each_zone(zone) {
7215 if (!is_highmem(zone))
7216 lowmem_pages += zone->managed_pages;
7219 for_each_zone(zone) {
7222 spin_lock_irqsave(&zone->lock, flags);
7223 tmp = (u64)pages_min * zone->managed_pages;
7224 do_div(tmp, lowmem_pages);
7225 if (is_highmem(zone)) {
7227 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7228 * need highmem pages, so cap pages_min to a small
7231 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7232 * deltas control asynch page reclaim, and so should
7233 * not be capped for highmem.
7235 unsigned long min_pages;
7237 min_pages = zone->managed_pages / 1024;
7238 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7239 zone->watermark[WMARK_MIN] = min_pages;
7242 * If it's a lowmem zone, reserve a number of pages
7243 * proportionate to the zone's size.
7245 zone->watermark[WMARK_MIN] = tmp;
7249 * Set the kswapd watermarks distance according to the
7250 * scale factor in proportion to available memory, but
7251 * ensure a minimum size on small systems.
7253 tmp = max_t(u64, tmp >> 2,
7254 mult_frac(zone->managed_pages,
7255 watermark_scale_factor, 10000));
7257 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7258 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7260 spin_unlock_irqrestore(&zone->lock, flags);
7263 /* update totalreserve_pages */
7264 calculate_totalreserve_pages();
7268 * setup_per_zone_wmarks - called when min_free_kbytes changes
7269 * or when memory is hot-{added|removed}
7271 * Ensures that the watermark[min,low,high] values for each zone are set
7272 * correctly with respect to min_free_kbytes.
7274 void setup_per_zone_wmarks(void)
7276 static DEFINE_SPINLOCK(lock);
7279 __setup_per_zone_wmarks();
7284 * Initialise min_free_kbytes.
7286 * For small machines we want it small (128k min). For large machines
7287 * we want it large (64MB max). But it is not linear, because network
7288 * bandwidth does not increase linearly with machine size. We use
7290 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7291 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7307 int __meminit init_per_zone_wmark_min(void)
7309 unsigned long lowmem_kbytes;
7310 int new_min_free_kbytes;
7312 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7313 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7315 if (new_min_free_kbytes > user_min_free_kbytes) {
7316 min_free_kbytes = new_min_free_kbytes;
7317 if (min_free_kbytes < 128)
7318 min_free_kbytes = 128;
7319 if (min_free_kbytes > 65536)
7320 min_free_kbytes = 65536;
7322 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7323 new_min_free_kbytes, user_min_free_kbytes);
7325 setup_per_zone_wmarks();
7326 refresh_zone_stat_thresholds();
7327 setup_per_zone_lowmem_reserve();
7330 setup_min_unmapped_ratio();
7331 setup_min_slab_ratio();
7336 core_initcall(init_per_zone_wmark_min)
7339 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7340 * that we can call two helper functions whenever min_free_kbytes
7343 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7344 void __user *buffer, size_t *length, loff_t *ppos)
7348 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7353 user_min_free_kbytes = min_free_kbytes;
7354 setup_per_zone_wmarks();
7359 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7360 void __user *buffer, size_t *length, loff_t *ppos)
7364 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7369 setup_per_zone_wmarks();
7375 static void setup_min_unmapped_ratio(void)
7380 for_each_online_pgdat(pgdat)
7381 pgdat->min_unmapped_pages = 0;
7384 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7385 sysctl_min_unmapped_ratio) / 100;
7389 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7390 void __user *buffer, size_t *length, loff_t *ppos)
7394 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7398 setup_min_unmapped_ratio();
7403 static void setup_min_slab_ratio(void)
7408 for_each_online_pgdat(pgdat)
7409 pgdat->min_slab_pages = 0;
7412 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7413 sysctl_min_slab_ratio) / 100;
7416 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7417 void __user *buffer, size_t *length, loff_t *ppos)
7421 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7425 setup_min_slab_ratio();
7432 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7433 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7434 * whenever sysctl_lowmem_reserve_ratio changes.
7436 * The reserve ratio obviously has absolutely no relation with the
7437 * minimum watermarks. The lowmem reserve ratio can only make sense
7438 * if in function of the boot time zone sizes.
7440 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7441 void __user *buffer, size_t *length, loff_t *ppos)
7443 proc_dointvec_minmax(table, write, buffer, length, ppos);
7444 setup_per_zone_lowmem_reserve();
7449 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7450 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7451 * pagelist can have before it gets flushed back to buddy allocator.
7453 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7454 void __user *buffer, size_t *length, loff_t *ppos)
7457 int old_percpu_pagelist_fraction;
7460 mutex_lock(&pcp_batch_high_lock);
7461 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7463 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7464 if (!write || ret < 0)
7467 /* Sanity checking to avoid pcp imbalance */
7468 if (percpu_pagelist_fraction &&
7469 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7470 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7476 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7479 for_each_populated_zone(zone) {
7482 for_each_possible_cpu(cpu)
7483 pageset_set_high_and_batch(zone,
7484 per_cpu_ptr(zone->pageset, cpu));
7487 mutex_unlock(&pcp_batch_high_lock);
7492 int hashdist = HASHDIST_DEFAULT;
7494 static int __init set_hashdist(char *str)
7498 hashdist = simple_strtoul(str, &str, 0);
7501 __setup("hashdist=", set_hashdist);
7504 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7506 * Returns the number of pages that arch has reserved but
7507 * is not known to alloc_large_system_hash().
7509 static unsigned long __init arch_reserved_kernel_pages(void)
7516 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7517 * machines. As memory size is increased the scale is also increased but at
7518 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7519 * quadruples the scale is increased by one, which means the size of hash table
7520 * only doubles, instead of quadrupling as well.
7521 * Because 32-bit systems cannot have large physical memory, where this scaling
7522 * makes sense, it is disabled on such platforms.
7524 #if __BITS_PER_LONG > 32
7525 #define ADAPT_SCALE_BASE (64ul << 30)
7526 #define ADAPT_SCALE_SHIFT 2
7527 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7531 * allocate a large system hash table from bootmem
7532 * - it is assumed that the hash table must contain an exact power-of-2
7533 * quantity of entries
7534 * - limit is the number of hash buckets, not the total allocation size
7536 void *__init alloc_large_system_hash(const char *tablename,
7537 unsigned long bucketsize,
7538 unsigned long numentries,
7541 unsigned int *_hash_shift,
7542 unsigned int *_hash_mask,
7543 unsigned long low_limit,
7544 unsigned long high_limit)
7546 unsigned long long max = high_limit;
7547 unsigned long log2qty, size;
7551 /* allow the kernel cmdline to have a say */
7553 /* round applicable memory size up to nearest megabyte */
7554 numentries = nr_kernel_pages;
7555 numentries -= arch_reserved_kernel_pages();
7557 /* It isn't necessary when PAGE_SIZE >= 1MB */
7558 if (PAGE_SHIFT < 20)
7559 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7561 #if __BITS_PER_LONG > 32
7563 unsigned long adapt;
7565 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7566 adapt <<= ADAPT_SCALE_SHIFT)
7571 /* limit to 1 bucket per 2^scale bytes of low memory */
7572 if (scale > PAGE_SHIFT)
7573 numentries >>= (scale - PAGE_SHIFT);
7575 numentries <<= (PAGE_SHIFT - scale);
7577 /* Make sure we've got at least a 0-order allocation.. */
7578 if (unlikely(flags & HASH_SMALL)) {
7579 /* Makes no sense without HASH_EARLY */
7580 WARN_ON(!(flags & HASH_EARLY));
7581 if (!(numentries >> *_hash_shift)) {
7582 numentries = 1UL << *_hash_shift;
7583 BUG_ON(!numentries);
7585 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7586 numentries = PAGE_SIZE / bucketsize;
7588 numentries = roundup_pow_of_two(numentries);
7590 /* limit allocation size to 1/16 total memory by default */
7592 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7593 do_div(max, bucketsize);
7595 max = min(max, 0x80000000ULL);
7597 if (numentries < low_limit)
7598 numentries = low_limit;
7599 if (numentries > max)
7602 log2qty = ilog2(numentries);
7604 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7606 size = bucketsize << log2qty;
7607 if (flags & HASH_EARLY) {
7608 if (flags & HASH_ZERO)
7609 table = memblock_virt_alloc_nopanic(size, 0);
7611 table = memblock_virt_alloc_raw(size, 0);
7612 } else if (hashdist) {
7613 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7616 * If bucketsize is not a power-of-two, we may free
7617 * some pages at the end of hash table which
7618 * alloc_pages_exact() automatically does
7620 if (get_order(size) < MAX_ORDER) {
7621 table = alloc_pages_exact(size, gfp_flags);
7622 kmemleak_alloc(table, size, 1, gfp_flags);
7625 } while (!table && size > PAGE_SIZE && --log2qty);
7628 panic("Failed to allocate %s hash table\n", tablename);
7630 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7631 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7634 *_hash_shift = log2qty;
7636 *_hash_mask = (1 << log2qty) - 1;
7642 * This function checks whether pageblock includes unmovable pages or not.
7643 * If @count is not zero, it is okay to include less @count unmovable pages
7645 * PageLRU check without isolation or lru_lock could race so that
7646 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7647 * check without lock_page also may miss some movable non-lru pages at
7648 * race condition. So you can't expect this function should be exact.
7650 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7652 bool skip_hwpoisoned_pages)
7654 unsigned long pfn, iter, found;
7657 * TODO we could make this much more efficient by not checking every
7658 * page in the range if we know all of them are in MOVABLE_ZONE and
7659 * that the movable zone guarantees that pages are migratable but
7660 * the later is not the case right now unfortunatelly. E.g. movablecore
7661 * can still lead to having bootmem allocations in zone_movable.
7665 * CMA allocations (alloc_contig_range) really need to mark isolate
7666 * CMA pageblocks even when they are not movable in fact so consider
7667 * them movable here.
7669 if (is_migrate_cma(migratetype) &&
7670 is_migrate_cma(get_pageblock_migratetype(page)))
7673 pfn = page_to_pfn(page);
7674 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7675 unsigned long check = pfn + iter;
7677 if (!pfn_valid_within(check))
7680 page = pfn_to_page(check);
7682 if (PageReserved(page))
7686 * Hugepages are not in LRU lists, but they're movable.
7687 * We need not scan over tail pages bacause we don't
7688 * handle each tail page individually in migration.
7690 if (PageHuge(page)) {
7692 if (!hugepage_migration_supported(page_hstate(page)))
7695 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7700 * We can't use page_count without pin a page
7701 * because another CPU can free compound page.
7702 * This check already skips compound tails of THP
7703 * because their page->_refcount is zero at all time.
7705 if (!page_ref_count(page)) {
7706 if (PageBuddy(page))
7707 iter += (1 << page_order(page)) - 1;
7712 * The HWPoisoned page may be not in buddy system, and
7713 * page_count() is not 0.
7715 if (skip_hwpoisoned_pages && PageHWPoison(page))
7718 if (__PageMovable(page))
7724 * If there are RECLAIMABLE pages, we need to check
7725 * it. But now, memory offline itself doesn't call
7726 * shrink_node_slabs() and it still to be fixed.
7729 * If the page is not RAM, page_count()should be 0.
7730 * we don't need more check. This is an _used_ not-movable page.
7732 * The problematic thing here is PG_reserved pages. PG_reserved
7733 * is set to both of a memory hole page and a _used_ kernel
7741 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7745 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7747 static unsigned long pfn_max_align_down(unsigned long pfn)
7749 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7750 pageblock_nr_pages) - 1);
7753 static unsigned long pfn_max_align_up(unsigned long pfn)
7755 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7756 pageblock_nr_pages));
7759 /* [start, end) must belong to a single zone. */
7760 static int __alloc_contig_migrate_range(struct compact_control *cc,
7761 unsigned long start, unsigned long end)
7763 /* This function is based on compact_zone() from compaction.c. */
7764 unsigned long nr_reclaimed;
7765 unsigned long pfn = start;
7766 unsigned int tries = 0;
7771 while (pfn < end || !list_empty(&cc->migratepages)) {
7772 if (fatal_signal_pending(current)) {
7777 if (list_empty(&cc->migratepages)) {
7778 cc->nr_migratepages = 0;
7779 pfn = isolate_migratepages_range(cc, pfn, end);
7785 } else if (++tries == 5) {
7786 ret = ret < 0 ? ret : -EBUSY;
7790 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7792 cc->nr_migratepages -= nr_reclaimed;
7794 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7795 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7798 putback_movable_pages(&cc->migratepages);
7805 * alloc_contig_range() -- tries to allocate given range of pages
7806 * @start: start PFN to allocate
7807 * @end: one-past-the-last PFN to allocate
7808 * @migratetype: migratetype of the underlaying pageblocks (either
7809 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7810 * in range must have the same migratetype and it must
7811 * be either of the two.
7812 * @gfp_mask: GFP mask to use during compaction
7814 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7815 * aligned. The PFN range must belong to a single zone.
7817 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7818 * pageblocks in the range. Once isolated, the pageblocks should not
7819 * be modified by others.
7821 * Returns zero on success or negative error code. On success all
7822 * pages which PFN is in [start, end) are allocated for the caller and
7823 * need to be freed with free_contig_range().
7825 int alloc_contig_range(unsigned long start, unsigned long end,
7826 unsigned migratetype, gfp_t gfp_mask)
7828 unsigned long outer_start, outer_end;
7832 struct compact_control cc = {
7833 .nr_migratepages = 0,
7835 .zone = page_zone(pfn_to_page(start)),
7836 .mode = MIGRATE_SYNC,
7837 .ignore_skip_hint = true,
7838 .no_set_skip_hint = true,
7839 .gfp_mask = current_gfp_context(gfp_mask),
7841 INIT_LIST_HEAD(&cc.migratepages);
7844 * What we do here is we mark all pageblocks in range as
7845 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7846 * have different sizes, and due to the way page allocator
7847 * work, we align the range to biggest of the two pages so
7848 * that page allocator won't try to merge buddies from
7849 * different pageblocks and change MIGRATE_ISOLATE to some
7850 * other migration type.
7852 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7853 * migrate the pages from an unaligned range (ie. pages that
7854 * we are interested in). This will put all the pages in
7855 * range back to page allocator as MIGRATE_ISOLATE.
7857 * When this is done, we take the pages in range from page
7858 * allocator removing them from the buddy system. This way
7859 * page allocator will never consider using them.
7861 * This lets us mark the pageblocks back as
7862 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7863 * aligned range but not in the unaligned, original range are
7864 * put back to page allocator so that buddy can use them.
7867 ret = start_isolate_page_range(pfn_max_align_down(start),
7868 pfn_max_align_up(end), migratetype,
7874 * In case of -EBUSY, we'd like to know which page causes problem.
7875 * So, just fall through. test_pages_isolated() has a tracepoint
7876 * which will report the busy page.
7878 * It is possible that busy pages could become available before
7879 * the call to test_pages_isolated, and the range will actually be
7880 * allocated. So, if we fall through be sure to clear ret so that
7881 * -EBUSY is not accidentally used or returned to caller.
7883 ret = __alloc_contig_migrate_range(&cc, start, end);
7884 if (ret && ret != -EBUSY)
7889 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7890 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7891 * more, all pages in [start, end) are free in page allocator.
7892 * What we are going to do is to allocate all pages from
7893 * [start, end) (that is remove them from page allocator).
7895 * The only problem is that pages at the beginning and at the
7896 * end of interesting range may be not aligned with pages that
7897 * page allocator holds, ie. they can be part of higher order
7898 * pages. Because of this, we reserve the bigger range and
7899 * once this is done free the pages we are not interested in.
7901 * We don't have to hold zone->lock here because the pages are
7902 * isolated thus they won't get removed from buddy.
7905 lru_add_drain_all();
7906 drain_all_pages(cc.zone);
7909 outer_start = start;
7910 while (!PageBuddy(pfn_to_page(outer_start))) {
7911 if (++order >= MAX_ORDER) {
7912 outer_start = start;
7915 outer_start &= ~0UL << order;
7918 if (outer_start != start) {
7919 order = page_order(pfn_to_page(outer_start));
7922 * outer_start page could be small order buddy page and
7923 * it doesn't include start page. Adjust outer_start
7924 * in this case to report failed page properly
7925 * on tracepoint in test_pages_isolated()
7927 if (outer_start + (1UL << order) <= start)
7928 outer_start = start;
7931 /* Make sure the range is really isolated. */
7932 if (test_pages_isolated(outer_start, end, false)) {
7933 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7934 __func__, outer_start, end);
7939 /* Grab isolated pages from freelists. */
7940 outer_end = isolate_freepages_range(&cc, outer_start, end);
7946 /* Free head and tail (if any) */
7947 if (start != outer_start)
7948 free_contig_range(outer_start, start - outer_start);
7949 if (end != outer_end)
7950 free_contig_range(end, outer_end - end);
7953 undo_isolate_page_range(pfn_max_align_down(start),
7954 pfn_max_align_up(end), migratetype);
7958 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7960 unsigned int count = 0;
7962 for (; nr_pages--; pfn++) {
7963 struct page *page = pfn_to_page(pfn);
7965 count += page_count(page) != 1;
7968 WARN(count != 0, "%d pages are still in use!\n", count);
7972 #ifdef CONFIG_MEMORY_HOTPLUG
7974 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7975 * page high values need to be recalulated.
7977 void __meminit zone_pcp_update(struct zone *zone)
7980 mutex_lock(&pcp_batch_high_lock);
7981 for_each_possible_cpu(cpu)
7982 pageset_set_high_and_batch(zone,
7983 per_cpu_ptr(zone->pageset, cpu));
7984 mutex_unlock(&pcp_batch_high_lock);
7988 void zone_pcp_reset(struct zone *zone)
7990 unsigned long flags;
7992 struct per_cpu_pageset *pset;
7994 /* avoid races with drain_pages() */
7995 local_irq_save(flags);
7996 if (zone->pageset != &boot_pageset) {
7997 for_each_online_cpu(cpu) {
7998 pset = per_cpu_ptr(zone->pageset, cpu);
7999 drain_zonestat(zone, pset);
8001 free_percpu(zone->pageset);
8002 zone->pageset = &boot_pageset;
8004 local_irq_restore(flags);
8007 #ifdef CONFIG_MEMORY_HOTREMOVE
8009 * All pages in the range must be in a single zone and isolated
8010 * before calling this.
8013 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8017 unsigned int order, i;
8019 unsigned long flags;
8020 /* find the first valid pfn */
8021 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8026 offline_mem_sections(pfn, end_pfn);
8027 zone = page_zone(pfn_to_page(pfn));
8028 spin_lock_irqsave(&zone->lock, flags);
8030 while (pfn < end_pfn) {
8031 if (!pfn_valid(pfn)) {
8035 page = pfn_to_page(pfn);
8037 * The HWPoisoned page may be not in buddy system, and
8038 * page_count() is not 0.
8040 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8042 SetPageReserved(page);
8046 BUG_ON(page_count(page));
8047 BUG_ON(!PageBuddy(page));
8048 order = page_order(page);
8049 #ifdef CONFIG_DEBUG_VM
8050 pr_info("remove from free list %lx %d %lx\n",
8051 pfn, 1 << order, end_pfn);
8053 list_del(&page->lru);
8054 rmv_page_order(page);
8055 zone->free_area[order].nr_free--;
8056 for (i = 0; i < (1 << order); i++)
8057 SetPageReserved((page+i));
8058 pfn += (1 << order);
8060 spin_unlock_irqrestore(&zone->lock, flags);
8064 bool is_free_buddy_page(struct page *page)
8066 struct zone *zone = page_zone(page);
8067 unsigned long pfn = page_to_pfn(page);
8068 unsigned long flags;
8071 spin_lock_irqsave(&zone->lock, flags);
8072 for (order = 0; order < MAX_ORDER; order++) {
8073 struct page *page_head = page - (pfn & ((1 << order) - 1));
8075 if (PageBuddy(page_head) && page_order(page_head) >= order)
8078 spin_unlock_irqrestore(&zone->lock, flags);
8080 return order < MAX_ORDER;
8083 #ifdef CONFIG_MEMORY_FAILURE
8085 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8086 * test is performed under the zone lock to prevent a race against page
8089 bool set_hwpoison_free_buddy_page(struct page *page)
8091 struct zone *zone = page_zone(page);
8092 unsigned long pfn = page_to_pfn(page);
8093 unsigned long flags;
8095 bool hwpoisoned = false;
8097 spin_lock_irqsave(&zone->lock, flags);
8098 for (order = 0; order < MAX_ORDER; order++) {
8099 struct page *page_head = page - (pfn & ((1 << order) - 1));
8101 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8102 if (!TestSetPageHWPoison(page))
8107 spin_unlock_irqrestore(&zone->lock, flags);