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>
70 #include <asm/sections.h>
71 #include <asm/tlbflush.h>
72 #include <asm/div64.h>
75 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
76 static DEFINE_MUTEX(pcp_batch_high_lock);
77 #define MIN_PERCPU_PAGELIST_FRACTION (8)
79 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
80 DEFINE_PER_CPU(int, numa_node);
81 EXPORT_PER_CPU_SYMBOL(numa_node);
84 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
86 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
95 int _node_numa_mem_[MAX_NUMNODES];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex);
100 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy;
104 EXPORT_SYMBOL(latent_entropy);
108 * Array of node states.
110 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
111 [N_POSSIBLE] = NODE_MASK_ALL,
112 [N_ONLINE] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118 [N_MEMORY] = { { [0] = 1UL } },
119 [N_CPU] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock);
127 unsigned long totalram_pages __read_mostly;
128 unsigned long totalreserve_pages __read_mostly;
129 unsigned long totalcma_pages __read_mostly;
131 int percpu_pagelist_fraction;
132 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page *page)
147 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149 page->index = migratetype;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with system_transition_mutex held
158 * (gfp_allowed_mask also should only be modified with system_transition_mutex
159 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
160 * with that modification).
163 static gfp_t saved_gfp_mask;
165 void pm_restore_gfp_mask(void)
167 WARN_ON(!mutex_is_locked(&system_transition_mutex));
168 if (saved_gfp_mask) {
169 gfp_allowed_mask = saved_gfp_mask;
174 void pm_restrict_gfp_mask(void)
176 WARN_ON(!mutex_is_locked(&system_transition_mutex));
177 WARN_ON(saved_gfp_mask);
178 saved_gfp_mask = gfp_allowed_mask;
179 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
182 bool pm_suspended_storage(void)
184 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
188 #endif /* CONFIG_PM_SLEEP */
190 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191 unsigned int pageblock_order __read_mostly;
194 static void __free_pages_ok(struct page *page, unsigned int order);
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
207 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
208 #ifdef CONFIG_ZONE_DMA
211 #ifdef CONFIG_ZONE_DMA32
215 #ifdef CONFIG_HIGHMEM
221 EXPORT_SYMBOL(totalram_pages);
223 static char * const zone_names[MAX_NR_ZONES] = {
224 #ifdef CONFIG_ZONE_DMA
227 #ifdef CONFIG_ZONE_DMA32
231 #ifdef CONFIG_HIGHMEM
235 #ifdef CONFIG_ZONE_DEVICE
240 char * const migratetype_names[MIGRATE_TYPES] = {
248 #ifdef CONFIG_MEMORY_ISOLATION
253 compound_page_dtor * const compound_page_dtors[] = {
256 #ifdef CONFIG_HUGETLB_PAGE
259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
264 int min_free_kbytes = 1024;
265 int user_min_free_kbytes = -1;
266 int watermark_scale_factor = 10;
268 static unsigned long nr_kernel_pages __meminitdata;
269 static unsigned long nr_all_pages __meminitdata;
270 static unsigned long dma_reserve __meminitdata;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275 static unsigned long required_kernelcore __initdata;
276 static unsigned long required_kernelcore_percent __initdata;
277 static unsigned long required_movablecore __initdata;
278 static unsigned long required_movablecore_percent __initdata;
279 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
280 static bool mirrored_kernelcore __meminitdata;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly = MAX_NUMNODES;
289 int nr_online_nodes __read_mostly = 1;
290 EXPORT_SYMBOL(nr_node_ids);
291 EXPORT_SYMBOL(nr_online_nodes);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 * During boot we initialize deferred pages on-demand, as needed, but once
299 * page_alloc_init_late() has finished, the deferred pages are all initialized,
300 * and we can permanently disable that path.
302 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
305 * Calling kasan_free_pages() only after deferred memory initialization
306 * has completed. Poisoning pages during deferred memory init will greatly
307 * lengthen the process and cause problem in large memory systems as the
308 * deferred pages initialization is done with interrupt disabled.
310 * Assuming that there will be no reference to those newly initialized
311 * pages before they are ever allocated, this should have no effect on
312 * KASAN memory tracking as the poison will be properly inserted at page
313 * allocation time. The only corner case is when pages are allocated by
314 * on-demand allocation and then freed again before the deferred pages
315 * initialization is done, but this is not likely to happen.
317 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
319 if (!static_branch_unlikely(&deferred_pages))
320 kasan_free_pages(page, order);
323 /* Returns true if the struct page for the pfn is uninitialised */
324 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
326 int nid = early_pfn_to_nid(pfn);
328 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
335 * Returns false when the remaining initialisation should be deferred until
336 * later in the boot cycle when it can be parallelised.
338 static inline bool update_defer_init(pg_data_t *pgdat,
339 unsigned long pfn, unsigned long zone_end,
340 unsigned long *nr_initialised)
342 /* Always populate low zones for address-constrained allocations */
343 if (zone_end < pgdat_end_pfn(pgdat))
346 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
347 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
348 pgdat->first_deferred_pfn = pfn;
355 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
357 static inline bool early_page_uninitialised(unsigned long pfn)
362 static inline bool update_defer_init(pg_data_t *pgdat,
363 unsigned long pfn, unsigned long zone_end,
364 unsigned long *nr_initialised)
370 /* Return a pointer to the bitmap storing bits affecting a block of pages */
371 static inline unsigned long *get_pageblock_bitmap(struct page *page,
374 #ifdef CONFIG_SPARSEMEM
375 return __pfn_to_section(pfn)->pageblock_flags;
377 return page_zone(page)->pageblock_flags;
378 #endif /* CONFIG_SPARSEMEM */
381 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
383 #ifdef CONFIG_SPARSEMEM
384 pfn &= (PAGES_PER_SECTION-1);
385 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
387 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
388 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
389 #endif /* CONFIG_SPARSEMEM */
393 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
394 * @page: The page within the block of interest
395 * @pfn: The target page frame number
396 * @end_bitidx: The last bit of interest to retrieve
397 * @mask: mask of bits that the caller is interested in
399 * Return: pageblock_bits flags
401 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
403 unsigned long end_bitidx,
406 unsigned long *bitmap;
407 unsigned long bitidx, word_bitidx;
410 bitmap = get_pageblock_bitmap(page, pfn);
411 bitidx = pfn_to_bitidx(page, pfn);
412 word_bitidx = bitidx / BITS_PER_LONG;
413 bitidx &= (BITS_PER_LONG-1);
415 word = bitmap[word_bitidx];
416 bitidx += end_bitidx;
417 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
420 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
421 unsigned long end_bitidx,
424 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
427 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
429 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
433 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
434 * @page: The page within the block of interest
435 * @flags: The flags to set
436 * @pfn: The target page frame number
437 * @end_bitidx: The last bit of interest
438 * @mask: mask of bits that the caller is interested in
440 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
442 unsigned long end_bitidx,
445 unsigned long *bitmap;
446 unsigned long bitidx, word_bitidx;
447 unsigned long old_word, word;
449 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
451 bitmap = get_pageblock_bitmap(page, pfn);
452 bitidx = pfn_to_bitidx(page, pfn);
453 word_bitidx = bitidx / BITS_PER_LONG;
454 bitidx &= (BITS_PER_LONG-1);
456 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
458 bitidx += end_bitidx;
459 mask <<= (BITS_PER_LONG - bitidx - 1);
460 flags <<= (BITS_PER_LONG - bitidx - 1);
462 word = READ_ONCE(bitmap[word_bitidx]);
464 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
465 if (word == old_word)
471 void set_pageblock_migratetype(struct page *page, int migratetype)
473 if (unlikely(page_group_by_mobility_disabled &&
474 migratetype < MIGRATE_PCPTYPES))
475 migratetype = MIGRATE_UNMOVABLE;
477 set_pageblock_flags_group(page, (unsigned long)migratetype,
478 PB_migrate, PB_migrate_end);
481 #ifdef CONFIG_DEBUG_VM
482 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
486 unsigned long pfn = page_to_pfn(page);
487 unsigned long sp, start_pfn;
490 seq = zone_span_seqbegin(zone);
491 start_pfn = zone->zone_start_pfn;
492 sp = zone->spanned_pages;
493 if (!zone_spans_pfn(zone, pfn))
495 } while (zone_span_seqretry(zone, seq));
498 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
499 pfn, zone_to_nid(zone), zone->name,
500 start_pfn, start_pfn + sp);
505 static int page_is_consistent(struct zone *zone, struct page *page)
507 if (!pfn_valid_within(page_to_pfn(page)))
509 if (zone != page_zone(page))
515 * Temporary debugging check for pages not lying within a given zone.
517 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
519 if (page_outside_zone_boundaries(zone, page))
521 if (!page_is_consistent(zone, page))
527 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
533 static void bad_page(struct page *page, const char *reason,
534 unsigned long bad_flags)
536 static unsigned long resume;
537 static unsigned long nr_shown;
538 static unsigned long nr_unshown;
541 * Allow a burst of 60 reports, then keep quiet for that minute;
542 * or allow a steady drip of one report per second.
544 if (nr_shown == 60) {
545 if (time_before(jiffies, resume)) {
551 "BUG: Bad page state: %lu messages suppressed\n",
558 resume = jiffies + 60 * HZ;
560 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
561 current->comm, page_to_pfn(page));
562 __dump_page(page, reason);
563 bad_flags &= page->flags;
565 pr_alert("bad because of flags: %#lx(%pGp)\n",
566 bad_flags, &bad_flags);
567 dump_page_owner(page);
572 /* Leave bad fields for debug, except PageBuddy could make trouble */
573 page_mapcount_reset(page); /* remove PageBuddy */
574 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
578 * Higher-order pages are called "compound pages". They are structured thusly:
580 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
582 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
583 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
585 * The first tail page's ->compound_dtor holds the offset in array of compound
586 * page destructors. See compound_page_dtors.
588 * The first tail page's ->compound_order holds the order of allocation.
589 * This usage means that zero-order pages may not be compound.
592 void free_compound_page(struct page *page)
594 __free_pages_ok(page, compound_order(page));
597 void prep_compound_page(struct page *page, unsigned int order)
600 int nr_pages = 1 << order;
602 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
603 set_compound_order(page, order);
605 for (i = 1; i < nr_pages; i++) {
606 struct page *p = page + i;
607 set_page_count(p, 0);
608 p->mapping = TAIL_MAPPING;
609 set_compound_head(p, page);
611 atomic_set(compound_mapcount_ptr(page), -1);
614 #ifdef CONFIG_DEBUG_PAGEALLOC
615 unsigned int _debug_guardpage_minorder;
616 bool _debug_pagealloc_enabled __read_mostly
617 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
618 EXPORT_SYMBOL(_debug_pagealloc_enabled);
619 bool _debug_guardpage_enabled __read_mostly;
621 static int __init early_debug_pagealloc(char *buf)
625 return kstrtobool(buf, &_debug_pagealloc_enabled);
627 early_param("debug_pagealloc", early_debug_pagealloc);
629 static bool need_debug_guardpage(void)
631 /* If we don't use debug_pagealloc, we don't need guard page */
632 if (!debug_pagealloc_enabled())
635 if (!debug_guardpage_minorder())
641 static void init_debug_guardpage(void)
643 if (!debug_pagealloc_enabled())
646 if (!debug_guardpage_minorder())
649 _debug_guardpage_enabled = true;
652 struct page_ext_operations debug_guardpage_ops = {
653 .need = need_debug_guardpage,
654 .init = init_debug_guardpage,
657 static int __init debug_guardpage_minorder_setup(char *buf)
661 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
662 pr_err("Bad debug_guardpage_minorder value\n");
665 _debug_guardpage_minorder = res;
666 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
669 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
671 static inline bool set_page_guard(struct zone *zone, struct page *page,
672 unsigned int order, int migratetype)
674 struct page_ext *page_ext;
676 if (!debug_guardpage_enabled())
679 if (order >= debug_guardpage_minorder())
682 page_ext = lookup_page_ext(page);
683 if (unlikely(!page_ext))
686 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
688 INIT_LIST_HEAD(&page->lru);
689 set_page_private(page, order);
690 /* Guard pages are not available for any usage */
691 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
696 static inline void clear_page_guard(struct zone *zone, struct page *page,
697 unsigned int order, int migratetype)
699 struct page_ext *page_ext;
701 if (!debug_guardpage_enabled())
704 page_ext = lookup_page_ext(page);
705 if (unlikely(!page_ext))
708 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
710 set_page_private(page, 0);
711 if (!is_migrate_isolate(migratetype))
712 __mod_zone_freepage_state(zone, (1 << order), migratetype);
715 struct page_ext_operations debug_guardpage_ops;
716 static inline bool set_page_guard(struct zone *zone, struct page *page,
717 unsigned int order, int migratetype) { return false; }
718 static inline void clear_page_guard(struct zone *zone, struct page *page,
719 unsigned int order, int migratetype) {}
722 static inline void set_page_order(struct page *page, unsigned int order)
724 set_page_private(page, order);
725 __SetPageBuddy(page);
728 static inline void rmv_page_order(struct page *page)
730 __ClearPageBuddy(page);
731 set_page_private(page, 0);
735 * This function checks whether a page is free && is the buddy
736 * we can coalesce a page and its buddy if
737 * (a) the buddy is not in a hole (check before calling!) &&
738 * (b) the buddy is in the buddy system &&
739 * (c) a page and its buddy have the same order &&
740 * (d) a page and its buddy are in the same zone.
742 * For recording whether a page is in the buddy system, we set PageBuddy.
743 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
745 * For recording page's order, we use page_private(page).
747 static inline int page_is_buddy(struct page *page, struct page *buddy,
750 if (page_is_guard(buddy) && page_order(buddy) == order) {
751 if (page_zone_id(page) != page_zone_id(buddy))
754 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
759 if (PageBuddy(buddy) && page_order(buddy) == order) {
761 * zone check is done late to avoid uselessly
762 * calculating zone/node ids for pages that could
765 if (page_zone_id(page) != page_zone_id(buddy))
768 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
776 * Freeing function for a buddy system allocator.
778 * The concept of a buddy system is to maintain direct-mapped table
779 * (containing bit values) for memory blocks of various "orders".
780 * The bottom level table contains the map for the smallest allocatable
781 * units of memory (here, pages), and each level above it describes
782 * pairs of units from the levels below, hence, "buddies".
783 * At a high level, all that happens here is marking the table entry
784 * at the bottom level available, and propagating the changes upward
785 * as necessary, plus some accounting needed to play nicely with other
786 * parts of the VM system.
787 * At each level, we keep a list of pages, which are heads of continuous
788 * free pages of length of (1 << order) and marked with PageBuddy.
789 * Page's order is recorded in page_private(page) field.
790 * So when we are allocating or freeing one, we can derive the state of the
791 * other. That is, if we allocate a small block, and both were
792 * free, the remainder of the region must be split into blocks.
793 * If a block is freed, and its buddy is also free, then this
794 * triggers coalescing into a block of larger size.
799 static inline void __free_one_page(struct page *page,
801 struct zone *zone, unsigned int order,
804 unsigned long combined_pfn;
805 unsigned long uninitialized_var(buddy_pfn);
807 unsigned int max_order;
809 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
811 VM_BUG_ON(!zone_is_initialized(zone));
812 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
814 VM_BUG_ON(migratetype == -1);
815 if (likely(!is_migrate_isolate(migratetype)))
816 __mod_zone_freepage_state(zone, 1 << order, migratetype);
818 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
819 VM_BUG_ON_PAGE(bad_range(zone, page), page);
822 while (order < max_order - 1) {
823 buddy_pfn = __find_buddy_pfn(pfn, order);
824 buddy = page + (buddy_pfn - pfn);
826 if (!pfn_valid_within(buddy_pfn))
828 if (!page_is_buddy(page, buddy, order))
831 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
832 * merge with it and move up one order.
834 if (page_is_guard(buddy)) {
835 clear_page_guard(zone, buddy, order, migratetype);
837 list_del(&buddy->lru);
838 zone->free_area[order].nr_free--;
839 rmv_page_order(buddy);
841 combined_pfn = buddy_pfn & pfn;
842 page = page + (combined_pfn - pfn);
846 if (max_order < MAX_ORDER) {
847 /* If we are here, it means order is >= pageblock_order.
848 * We want to prevent merge between freepages on isolate
849 * pageblock and normal pageblock. Without this, pageblock
850 * isolation could cause incorrect freepage or CMA accounting.
852 * We don't want to hit this code for the more frequent
855 if (unlikely(has_isolate_pageblock(zone))) {
858 buddy_pfn = __find_buddy_pfn(pfn, order);
859 buddy = page + (buddy_pfn - pfn);
860 buddy_mt = get_pageblock_migratetype(buddy);
862 if (migratetype != buddy_mt
863 && (is_migrate_isolate(migratetype) ||
864 is_migrate_isolate(buddy_mt)))
868 goto continue_merging;
872 set_page_order(page, order);
875 * If this is not the largest possible page, check if the buddy
876 * of the next-highest order is free. If it is, it's possible
877 * that pages are being freed that will coalesce soon. In case,
878 * that is happening, add the free page to the tail of the list
879 * so it's less likely to be used soon and more likely to be merged
880 * as a higher order page
882 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
883 struct page *higher_page, *higher_buddy;
884 combined_pfn = buddy_pfn & pfn;
885 higher_page = page + (combined_pfn - pfn);
886 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
887 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
888 if (pfn_valid_within(buddy_pfn) &&
889 page_is_buddy(higher_page, higher_buddy, order + 1)) {
890 list_add_tail(&page->lru,
891 &zone->free_area[order].free_list[migratetype]);
896 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
898 zone->free_area[order].nr_free++;
902 * A bad page could be due to a number of fields. Instead of multiple branches,
903 * try and check multiple fields with one check. The caller must do a detailed
904 * check if necessary.
906 static inline bool page_expected_state(struct page *page,
907 unsigned long check_flags)
909 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 if (unlikely((unsigned long)page->mapping |
913 page_ref_count(page) |
915 (unsigned long)page->mem_cgroup |
917 (page->flags & check_flags)))
923 static void free_pages_check_bad(struct page *page)
925 const char *bad_reason;
926 unsigned long bad_flags;
931 if (unlikely(atomic_read(&page->_mapcount) != -1))
932 bad_reason = "nonzero mapcount";
933 if (unlikely(page->mapping != NULL))
934 bad_reason = "non-NULL mapping";
935 if (unlikely(page_ref_count(page) != 0))
936 bad_reason = "nonzero _refcount";
937 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
938 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
939 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
942 if (unlikely(page->mem_cgroup))
943 bad_reason = "page still charged to cgroup";
945 bad_page(page, bad_reason, bad_flags);
948 static inline int free_pages_check(struct page *page)
950 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
953 /* Something has gone sideways, find it */
954 free_pages_check_bad(page);
958 static int free_tail_pages_check(struct page *head_page, struct page *page)
963 * We rely page->lru.next never has bit 0 set, unless the page
964 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
966 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
968 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
972 switch (page - head_page) {
974 /* the first tail page: ->mapping may be compound_mapcount() */
975 if (unlikely(compound_mapcount(page))) {
976 bad_page(page, "nonzero compound_mapcount", 0);
982 * the second tail page: ->mapping is
983 * deferred_list.next -- ignore value.
987 if (page->mapping != TAIL_MAPPING) {
988 bad_page(page, "corrupted mapping in tail page", 0);
993 if (unlikely(!PageTail(page))) {
994 bad_page(page, "PageTail not set", 0);
997 if (unlikely(compound_head(page) != head_page)) {
998 bad_page(page, "compound_head not consistent", 0);
1003 page->mapping = NULL;
1004 clear_compound_head(page);
1008 static __always_inline bool free_pages_prepare(struct page *page,
1009 unsigned int order, bool check_free)
1013 VM_BUG_ON_PAGE(PageTail(page), page);
1015 trace_mm_page_free(page, order);
1018 * Check tail pages before head page information is cleared to
1019 * avoid checking PageCompound for order-0 pages.
1021 if (unlikely(order)) {
1022 bool compound = PageCompound(page);
1025 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1028 ClearPageDoubleMap(page);
1029 for (i = 1; i < (1 << order); i++) {
1031 bad += free_tail_pages_check(page, page + i);
1032 if (unlikely(free_pages_check(page + i))) {
1036 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1039 if (PageMappingFlags(page))
1040 page->mapping = NULL;
1041 if (memcg_kmem_enabled() && PageKmemcg(page))
1042 memcg_kmem_uncharge(page, order);
1044 bad += free_pages_check(page);
1048 page_cpupid_reset_last(page);
1049 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1050 reset_page_owner(page, order);
1052 if (!PageHighMem(page)) {
1053 debug_check_no_locks_freed(page_address(page),
1054 PAGE_SIZE << order);
1055 debug_check_no_obj_freed(page_address(page),
1056 PAGE_SIZE << order);
1058 arch_free_page(page, order);
1059 kernel_poison_pages(page, 1 << order, 0);
1060 kernel_map_pages(page, 1 << order, 0);
1061 kasan_free_nondeferred_pages(page, order);
1066 #ifdef CONFIG_DEBUG_VM
1067 static inline bool free_pcp_prepare(struct page *page)
1069 return free_pages_prepare(page, 0, true);
1072 static inline bool bulkfree_pcp_prepare(struct page *page)
1077 static bool free_pcp_prepare(struct page *page)
1079 return free_pages_prepare(page, 0, false);
1082 static bool bulkfree_pcp_prepare(struct page *page)
1084 return free_pages_check(page);
1086 #endif /* CONFIG_DEBUG_VM */
1088 static inline void prefetch_buddy(struct page *page)
1090 unsigned long pfn = page_to_pfn(page);
1091 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1092 struct page *buddy = page + (buddy_pfn - pfn);
1098 * Frees a number of pages from the PCP lists
1099 * Assumes all pages on list are in same zone, and of same order.
1100 * count is the number of pages to free.
1102 * If the zone was previously in an "all pages pinned" state then look to
1103 * see if this freeing clears that state.
1105 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1106 * pinned" detection logic.
1108 static void free_pcppages_bulk(struct zone *zone, int count,
1109 struct per_cpu_pages *pcp)
1111 int migratetype = 0;
1113 int prefetch_nr = 0;
1114 bool isolated_pageblocks;
1115 struct page *page, *tmp;
1119 struct list_head *list;
1122 * Remove pages from lists in a round-robin fashion. A
1123 * batch_free count is maintained that is incremented when an
1124 * empty list is encountered. This is so more pages are freed
1125 * off fuller lists instead of spinning excessively around empty
1130 if (++migratetype == MIGRATE_PCPTYPES)
1132 list = &pcp->lists[migratetype];
1133 } while (list_empty(list));
1135 /* This is the only non-empty list. Free them all. */
1136 if (batch_free == MIGRATE_PCPTYPES)
1140 page = list_last_entry(list, struct page, lru);
1141 /* must delete to avoid corrupting pcp list */
1142 list_del(&page->lru);
1145 if (bulkfree_pcp_prepare(page))
1148 list_add_tail(&page->lru, &head);
1151 * We are going to put the page back to the global
1152 * pool, prefetch its buddy to speed up later access
1153 * under zone->lock. It is believed the overhead of
1154 * an additional test and calculating buddy_pfn here
1155 * can be offset by reduced memory latency later. To
1156 * avoid excessive prefetching due to large count, only
1157 * prefetch buddy for the first pcp->batch nr of pages.
1159 if (prefetch_nr++ < pcp->batch)
1160 prefetch_buddy(page);
1161 } while (--count && --batch_free && !list_empty(list));
1164 spin_lock(&zone->lock);
1165 isolated_pageblocks = has_isolate_pageblock(zone);
1168 * Use safe version since after __free_one_page(),
1169 * page->lru.next will not point to original list.
1171 list_for_each_entry_safe(page, tmp, &head, lru) {
1172 int mt = get_pcppage_migratetype(page);
1173 /* MIGRATE_ISOLATE page should not go to pcplists */
1174 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1175 /* Pageblock could have been isolated meanwhile */
1176 if (unlikely(isolated_pageblocks))
1177 mt = get_pageblock_migratetype(page);
1179 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1180 trace_mm_page_pcpu_drain(page, 0, mt);
1182 spin_unlock(&zone->lock);
1185 static void free_one_page(struct zone *zone,
1186 struct page *page, unsigned long pfn,
1190 spin_lock(&zone->lock);
1191 if (unlikely(has_isolate_pageblock(zone) ||
1192 is_migrate_isolate(migratetype))) {
1193 migratetype = get_pfnblock_migratetype(page, pfn);
1195 __free_one_page(page, pfn, zone, order, migratetype);
1196 spin_unlock(&zone->lock);
1199 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1200 unsigned long zone, int nid)
1202 mm_zero_struct_page(page);
1203 set_page_links(page, zone, nid, pfn);
1204 init_page_count(page);
1205 page_mapcount_reset(page);
1206 page_cpupid_reset_last(page);
1208 INIT_LIST_HEAD(&page->lru);
1209 #ifdef WANT_PAGE_VIRTUAL
1210 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1211 if (!is_highmem_idx(zone))
1212 set_page_address(page, __va(pfn << PAGE_SHIFT));
1216 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1217 static void __meminit init_reserved_page(unsigned long pfn)
1222 if (!early_page_uninitialised(pfn))
1225 nid = early_pfn_to_nid(pfn);
1226 pgdat = NODE_DATA(nid);
1228 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1229 struct zone *zone = &pgdat->node_zones[zid];
1231 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1234 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1237 static inline void init_reserved_page(unsigned long pfn)
1240 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1243 * Initialised pages do not have PageReserved set. This function is
1244 * called for each range allocated by the bootmem allocator and
1245 * marks the pages PageReserved. The remaining valid pages are later
1246 * sent to the buddy page allocator.
1248 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1250 unsigned long start_pfn = PFN_DOWN(start);
1251 unsigned long end_pfn = PFN_UP(end);
1253 for (; start_pfn < end_pfn; start_pfn++) {
1254 if (pfn_valid(start_pfn)) {
1255 struct page *page = pfn_to_page(start_pfn);
1257 init_reserved_page(start_pfn);
1259 /* Avoid false-positive PageTail() */
1260 INIT_LIST_HEAD(&page->lru);
1262 SetPageReserved(page);
1267 static void __free_pages_ok(struct page *page, unsigned int order)
1269 unsigned long flags;
1271 unsigned long pfn = page_to_pfn(page);
1273 if (!free_pages_prepare(page, order, true))
1276 migratetype = get_pfnblock_migratetype(page, pfn);
1277 local_irq_save(flags);
1278 __count_vm_events(PGFREE, 1 << order);
1279 free_one_page(page_zone(page), page, pfn, order, migratetype);
1280 local_irq_restore(flags);
1283 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1285 unsigned int nr_pages = 1 << order;
1286 struct page *p = page;
1290 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1292 __ClearPageReserved(p);
1293 set_page_count(p, 0);
1295 __ClearPageReserved(p);
1296 set_page_count(p, 0);
1298 page_zone(page)->managed_pages += nr_pages;
1299 set_page_refcounted(page);
1300 __free_pages(page, order);
1303 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1304 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1306 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1308 int __meminit early_pfn_to_nid(unsigned long pfn)
1310 static DEFINE_SPINLOCK(early_pfn_lock);
1313 spin_lock(&early_pfn_lock);
1314 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1316 nid = first_online_node;
1317 spin_unlock(&early_pfn_lock);
1323 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1324 static inline bool __meminit __maybe_unused
1325 meminit_pfn_in_nid(unsigned long pfn, int node,
1326 struct mminit_pfnnid_cache *state)
1330 nid = __early_pfn_to_nid(pfn, state);
1331 if (nid >= 0 && nid != node)
1336 /* Only safe to use early in boot when initialisation is single-threaded */
1337 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1339 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1344 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1348 static inline bool __meminit __maybe_unused
1349 meminit_pfn_in_nid(unsigned long pfn, int node,
1350 struct mminit_pfnnid_cache *state)
1357 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1360 if (early_page_uninitialised(pfn))
1362 return __free_pages_boot_core(page, order);
1366 * Check that the whole (or subset of) a pageblock given by the interval of
1367 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1368 * with the migration of free compaction scanner. The scanners then need to
1369 * use only pfn_valid_within() check for arches that allow holes within
1372 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1374 * It's possible on some configurations to have a setup like node0 node1 node0
1375 * i.e. it's possible that all pages within a zones range of pages do not
1376 * belong to a single zone. We assume that a border between node0 and node1
1377 * can occur within a single pageblock, but not a node0 node1 node0
1378 * interleaving within a single pageblock. It is therefore sufficient to check
1379 * the first and last page of a pageblock and avoid checking each individual
1380 * page in a pageblock.
1382 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1383 unsigned long end_pfn, struct zone *zone)
1385 struct page *start_page;
1386 struct page *end_page;
1388 /* end_pfn is one past the range we are checking */
1391 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1394 start_page = pfn_to_online_page(start_pfn);
1398 if (page_zone(start_page) != zone)
1401 end_page = pfn_to_page(end_pfn);
1403 /* This gives a shorter code than deriving page_zone(end_page) */
1404 if (page_zone_id(start_page) != page_zone_id(end_page))
1410 void set_zone_contiguous(struct zone *zone)
1412 unsigned long block_start_pfn = zone->zone_start_pfn;
1413 unsigned long block_end_pfn;
1415 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1416 for (; block_start_pfn < zone_end_pfn(zone);
1417 block_start_pfn = block_end_pfn,
1418 block_end_pfn += pageblock_nr_pages) {
1420 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1422 if (!__pageblock_pfn_to_page(block_start_pfn,
1423 block_end_pfn, zone))
1427 /* We confirm that there is no hole */
1428 zone->contiguous = true;
1431 void clear_zone_contiguous(struct zone *zone)
1433 zone->contiguous = false;
1436 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1437 static void __init deferred_free_range(unsigned long pfn,
1438 unsigned long nr_pages)
1446 page = pfn_to_page(pfn);
1448 /* Free a large naturally-aligned chunk if possible */
1449 if (nr_pages == pageblock_nr_pages &&
1450 (pfn & (pageblock_nr_pages - 1)) == 0) {
1451 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1452 __free_pages_boot_core(page, pageblock_order);
1456 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1457 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1458 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1459 __free_pages_boot_core(page, 0);
1463 /* Completion tracking for deferred_init_memmap() threads */
1464 static atomic_t pgdat_init_n_undone __initdata;
1465 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1467 static inline void __init pgdat_init_report_one_done(void)
1469 if (atomic_dec_and_test(&pgdat_init_n_undone))
1470 complete(&pgdat_init_all_done_comp);
1474 * Returns true if page needs to be initialized or freed to buddy allocator.
1476 * First we check if pfn is valid on architectures where it is possible to have
1477 * holes within pageblock_nr_pages. On systems where it is not possible, this
1478 * function is optimized out.
1480 * Then, we check if a current large page is valid by only checking the validity
1483 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1484 * within a node: a pfn is between start and end of a node, but does not belong
1485 * to this memory node.
1487 static inline bool __init
1488 deferred_pfn_valid(int nid, unsigned long pfn,
1489 struct mminit_pfnnid_cache *nid_init_state)
1491 if (!pfn_valid_within(pfn))
1493 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1495 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1501 * Free pages to buddy allocator. Try to free aligned pages in
1502 * pageblock_nr_pages sizes.
1504 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1505 unsigned long end_pfn)
1507 struct mminit_pfnnid_cache nid_init_state = { };
1508 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1509 unsigned long nr_free = 0;
1511 for (; pfn < end_pfn; pfn++) {
1512 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1513 deferred_free_range(pfn - nr_free, nr_free);
1515 } else if (!(pfn & nr_pgmask)) {
1516 deferred_free_range(pfn - nr_free, nr_free);
1518 touch_nmi_watchdog();
1523 /* Free the last block of pages to allocator */
1524 deferred_free_range(pfn - nr_free, nr_free);
1528 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1529 * by performing it only once every pageblock_nr_pages.
1530 * Return number of pages initialized.
1532 static unsigned long __init deferred_init_pages(int nid, int zid,
1534 unsigned long end_pfn)
1536 struct mminit_pfnnid_cache nid_init_state = { };
1537 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1538 unsigned long nr_pages = 0;
1539 struct page *page = NULL;
1541 for (; pfn < end_pfn; pfn++) {
1542 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1545 } else if (!page || !(pfn & nr_pgmask)) {
1546 page = pfn_to_page(pfn);
1547 touch_nmi_watchdog();
1551 __init_single_page(page, pfn, zid, nid);
1557 /* Initialise remaining memory on a node */
1558 static int __init deferred_init_memmap(void *data)
1560 pg_data_t *pgdat = data;
1561 int nid = pgdat->node_id;
1562 unsigned long start = jiffies;
1563 unsigned long nr_pages = 0;
1564 unsigned long spfn, epfn, first_init_pfn, flags;
1565 phys_addr_t spa, epa;
1568 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1571 /* Bind memory initialisation thread to a local node if possible */
1572 if (!cpumask_empty(cpumask))
1573 set_cpus_allowed_ptr(current, cpumask);
1575 pgdat_resize_lock(pgdat, &flags);
1576 first_init_pfn = pgdat->first_deferred_pfn;
1577 if (first_init_pfn == ULONG_MAX) {
1578 pgdat_resize_unlock(pgdat, &flags);
1579 pgdat_init_report_one_done();
1583 /* Sanity check boundaries */
1584 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1585 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1586 pgdat->first_deferred_pfn = ULONG_MAX;
1588 /* Only the highest zone is deferred so find it */
1589 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1590 zone = pgdat->node_zones + zid;
1591 if (first_init_pfn < zone_end_pfn(zone))
1594 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1597 * Initialize and free pages. We do it in two loops: first we initialize
1598 * struct page, than free to buddy allocator, because while we are
1599 * freeing pages we can access pages that are ahead (computing buddy
1600 * page in __free_one_page()).
1602 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1603 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1604 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1605 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1607 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1608 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1609 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1610 deferred_free_pages(nid, zid, spfn, epfn);
1612 pgdat_resize_unlock(pgdat, &flags);
1614 /* Sanity check that the next zone really is unpopulated */
1615 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1617 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1618 jiffies_to_msecs(jiffies - start));
1620 pgdat_init_report_one_done();
1625 * If this zone has deferred pages, try to grow it by initializing enough
1626 * deferred pages to satisfy the allocation specified by order, rounded up to
1627 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1628 * of SECTION_SIZE bytes by initializing struct pages in increments of
1629 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1631 * Return true when zone was grown, otherwise return false. We return true even
1632 * when we grow less than requested, to let the caller decide if there are
1633 * enough pages to satisfy the allocation.
1635 * Note: We use noinline because this function is needed only during boot, and
1636 * it is called from a __ref function _deferred_grow_zone. This way we are
1637 * making sure that it is not inlined into permanent text section.
1639 static noinline bool __init
1640 deferred_grow_zone(struct zone *zone, unsigned int order)
1642 int zid = zone_idx(zone);
1643 int nid = zone_to_nid(zone);
1644 pg_data_t *pgdat = NODE_DATA(nid);
1645 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1646 unsigned long nr_pages = 0;
1647 unsigned long first_init_pfn, spfn, epfn, t, flags;
1648 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1649 phys_addr_t spa, epa;
1652 /* Only the last zone may have deferred pages */
1653 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1656 pgdat_resize_lock(pgdat, &flags);
1659 * If deferred pages have been initialized while we were waiting for
1660 * the lock, return true, as the zone was grown. The caller will retry
1661 * this zone. We won't return to this function since the caller also
1662 * has this static branch.
1664 if (!static_branch_unlikely(&deferred_pages)) {
1665 pgdat_resize_unlock(pgdat, &flags);
1670 * If someone grew this zone while we were waiting for spinlock, return
1671 * true, as there might be enough pages already.
1673 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1674 pgdat_resize_unlock(pgdat, &flags);
1678 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1680 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1681 pgdat_resize_unlock(pgdat, &flags);
1685 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1686 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1687 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1689 while (spfn < epfn && nr_pages < nr_pages_needed) {
1690 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1691 first_deferred_pfn = min(t, epfn);
1692 nr_pages += deferred_init_pages(nid, zid, spfn,
1693 first_deferred_pfn);
1694 spfn = first_deferred_pfn;
1697 if (nr_pages >= nr_pages_needed)
1701 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1702 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1703 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1704 deferred_free_pages(nid, zid, spfn, epfn);
1706 if (first_deferred_pfn == epfn)
1709 pgdat->first_deferred_pfn = first_deferred_pfn;
1710 pgdat_resize_unlock(pgdat, &flags);
1712 return nr_pages > 0;
1716 * deferred_grow_zone() is __init, but it is called from
1717 * get_page_from_freelist() during early boot until deferred_pages permanently
1718 * disables this call. This is why we have refdata wrapper to avoid warning,
1719 * and to ensure that the function body gets unloaded.
1722 _deferred_grow_zone(struct zone *zone, unsigned int order)
1724 return deferred_grow_zone(zone, order);
1727 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1729 void __init page_alloc_init_late(void)
1733 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1736 /* There will be num_node_state(N_MEMORY) threads */
1737 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1738 for_each_node_state(nid, N_MEMORY) {
1739 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1742 /* Block until all are initialised */
1743 wait_for_completion(&pgdat_init_all_done_comp);
1746 * The number of managed pages has changed due to the initialisation
1747 * so the pcpu batch and high limits needs to be updated or the limits
1748 * will be artificially small.
1750 for_each_populated_zone(zone)
1751 zone_pcp_update(zone);
1754 * We initialized the rest of the deferred pages. Permanently disable
1755 * on-demand struct page initialization.
1757 static_branch_disable(&deferred_pages);
1759 /* Reinit limits that are based on free pages after the kernel is up */
1760 files_maxfiles_init();
1762 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1763 /* Discard memblock private memory */
1767 for_each_populated_zone(zone)
1768 set_zone_contiguous(zone);
1772 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1773 void __init init_cma_reserved_pageblock(struct page *page)
1775 unsigned i = pageblock_nr_pages;
1776 struct page *p = page;
1779 __ClearPageReserved(p);
1780 set_page_count(p, 0);
1783 set_pageblock_migratetype(page, MIGRATE_CMA);
1785 if (pageblock_order >= MAX_ORDER) {
1786 i = pageblock_nr_pages;
1789 set_page_refcounted(p);
1790 __free_pages(p, MAX_ORDER - 1);
1791 p += MAX_ORDER_NR_PAGES;
1792 } while (i -= MAX_ORDER_NR_PAGES);
1794 set_page_refcounted(page);
1795 __free_pages(page, pageblock_order);
1798 adjust_managed_page_count(page, pageblock_nr_pages);
1803 * The order of subdivision here is critical for the IO subsystem.
1804 * Please do not alter this order without good reasons and regression
1805 * testing. Specifically, as large blocks of memory are subdivided,
1806 * the order in which smaller blocks are delivered depends on the order
1807 * they're subdivided in this function. This is the primary factor
1808 * influencing the order in which pages are delivered to the IO
1809 * subsystem according to empirical testing, and this is also justified
1810 * by considering the behavior of a buddy system containing a single
1811 * large block of memory acted on by a series of small allocations.
1812 * This behavior is a critical factor in sglist merging's success.
1816 static inline void expand(struct zone *zone, struct page *page,
1817 int low, int high, struct free_area *area,
1820 unsigned long size = 1 << high;
1822 while (high > low) {
1826 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1829 * Mark as guard pages (or page), that will allow to
1830 * merge back to allocator when buddy will be freed.
1831 * Corresponding page table entries will not be touched,
1832 * pages will stay not present in virtual address space
1834 if (set_page_guard(zone, &page[size], high, migratetype))
1837 list_add(&page[size].lru, &area->free_list[migratetype]);
1839 set_page_order(&page[size], high);
1843 static void check_new_page_bad(struct page *page)
1845 const char *bad_reason = NULL;
1846 unsigned long bad_flags = 0;
1848 if (unlikely(atomic_read(&page->_mapcount) != -1))
1849 bad_reason = "nonzero mapcount";
1850 if (unlikely(page->mapping != NULL))
1851 bad_reason = "non-NULL mapping";
1852 if (unlikely(page_ref_count(page) != 0))
1853 bad_reason = "nonzero _count";
1854 if (unlikely(page->flags & __PG_HWPOISON)) {
1855 bad_reason = "HWPoisoned (hardware-corrupted)";
1856 bad_flags = __PG_HWPOISON;
1857 /* Don't complain about hwpoisoned pages */
1858 page_mapcount_reset(page); /* remove PageBuddy */
1861 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1862 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1863 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1866 if (unlikely(page->mem_cgroup))
1867 bad_reason = "page still charged to cgroup";
1869 bad_page(page, bad_reason, bad_flags);
1873 * This page is about to be returned from the page allocator
1875 static inline int check_new_page(struct page *page)
1877 if (likely(page_expected_state(page,
1878 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1881 check_new_page_bad(page);
1885 static inline bool free_pages_prezeroed(void)
1887 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1888 page_poisoning_enabled();
1891 #ifdef CONFIG_DEBUG_VM
1892 static bool check_pcp_refill(struct page *page)
1897 static bool check_new_pcp(struct page *page)
1899 return check_new_page(page);
1902 static bool check_pcp_refill(struct page *page)
1904 return check_new_page(page);
1906 static bool check_new_pcp(struct page *page)
1910 #endif /* CONFIG_DEBUG_VM */
1912 static bool check_new_pages(struct page *page, unsigned int order)
1915 for (i = 0; i < (1 << order); i++) {
1916 struct page *p = page + i;
1918 if (unlikely(check_new_page(p)))
1925 inline void post_alloc_hook(struct page *page, unsigned int order,
1928 set_page_private(page, 0);
1929 set_page_refcounted(page);
1931 arch_alloc_page(page, order);
1932 kernel_map_pages(page, 1 << order, 1);
1933 kasan_alloc_pages(page, order);
1934 kernel_poison_pages(page, 1 << order, 1);
1935 set_page_owner(page, order, gfp_flags);
1938 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1939 unsigned int alloc_flags)
1943 post_alloc_hook(page, order, gfp_flags);
1945 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1946 for (i = 0; i < (1 << order); i++)
1947 clear_highpage(page + i);
1949 if (order && (gfp_flags & __GFP_COMP))
1950 prep_compound_page(page, order);
1953 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1954 * allocate the page. The expectation is that the caller is taking
1955 * steps that will free more memory. The caller should avoid the page
1956 * being used for !PFMEMALLOC purposes.
1958 if (alloc_flags & ALLOC_NO_WATERMARKS)
1959 set_page_pfmemalloc(page);
1961 clear_page_pfmemalloc(page);
1965 * Go through the free lists for the given migratetype and remove
1966 * the smallest available page from the freelists
1968 static __always_inline
1969 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1972 unsigned int current_order;
1973 struct free_area *area;
1976 /* Find a page of the appropriate size in the preferred list */
1977 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1978 area = &(zone->free_area[current_order]);
1979 page = list_first_entry_or_null(&area->free_list[migratetype],
1983 list_del(&page->lru);
1984 rmv_page_order(page);
1986 expand(zone, page, order, current_order, area, migratetype);
1987 set_pcppage_migratetype(page, migratetype);
1996 * This array describes the order lists are fallen back to when
1997 * the free lists for the desirable migrate type are depleted
1999 static int fallbacks[MIGRATE_TYPES][4] = {
2000 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2001 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2002 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2004 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2006 #ifdef CONFIG_MEMORY_ISOLATION
2007 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2012 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2015 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2018 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2019 unsigned int order) { return NULL; }
2023 * Move the free pages in a range to the free lists of the requested type.
2024 * Note that start_page and end_pages are not aligned on a pageblock
2025 * boundary. If alignment is required, use move_freepages_block()
2027 static int move_freepages(struct zone *zone,
2028 struct page *start_page, struct page *end_page,
2029 int migratetype, int *num_movable)
2033 int pages_moved = 0;
2035 #ifndef CONFIG_HOLES_IN_ZONE
2037 * page_zone is not safe to call in this context when
2038 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2039 * anyway as we check zone boundaries in move_freepages_block().
2040 * Remove at a later date when no bug reports exist related to
2041 * grouping pages by mobility
2043 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2044 pfn_valid(page_to_pfn(end_page)) &&
2045 page_zone(start_page) != page_zone(end_page));
2051 for (page = start_page; page <= end_page;) {
2052 if (!pfn_valid_within(page_to_pfn(page))) {
2057 /* Make sure we are not inadvertently changing nodes */
2058 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2060 if (!PageBuddy(page)) {
2062 * We assume that pages that could be isolated for
2063 * migration are movable. But we don't actually try
2064 * isolating, as that would be expensive.
2067 (PageLRU(page) || __PageMovable(page)))
2074 order = page_order(page);
2075 list_move(&page->lru,
2076 &zone->free_area[order].free_list[migratetype]);
2078 pages_moved += 1 << order;
2084 int move_freepages_block(struct zone *zone, struct page *page,
2085 int migratetype, int *num_movable)
2087 unsigned long start_pfn, end_pfn;
2088 struct page *start_page, *end_page;
2090 start_pfn = page_to_pfn(page);
2091 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2092 start_page = pfn_to_page(start_pfn);
2093 end_page = start_page + pageblock_nr_pages - 1;
2094 end_pfn = start_pfn + pageblock_nr_pages - 1;
2096 /* Do not cross zone boundaries */
2097 if (!zone_spans_pfn(zone, start_pfn))
2099 if (!zone_spans_pfn(zone, end_pfn))
2102 return move_freepages(zone, start_page, end_page, migratetype,
2106 static void change_pageblock_range(struct page *pageblock_page,
2107 int start_order, int migratetype)
2109 int nr_pageblocks = 1 << (start_order - pageblock_order);
2111 while (nr_pageblocks--) {
2112 set_pageblock_migratetype(pageblock_page, migratetype);
2113 pageblock_page += pageblock_nr_pages;
2118 * When we are falling back to another migratetype during allocation, try to
2119 * steal extra free pages from the same pageblocks to satisfy further
2120 * allocations, instead of polluting multiple pageblocks.
2122 * If we are stealing a relatively large buddy page, it is likely there will
2123 * be more free pages in the pageblock, so try to steal them all. For
2124 * reclaimable and unmovable allocations, we steal regardless of page size,
2125 * as fragmentation caused by those allocations polluting movable pageblocks
2126 * is worse than movable allocations stealing from unmovable and reclaimable
2129 static bool can_steal_fallback(unsigned int order, int start_mt)
2132 * Leaving this order check is intended, although there is
2133 * relaxed order check in next check. The reason is that
2134 * we can actually steal whole pageblock if this condition met,
2135 * but, below check doesn't guarantee it and that is just heuristic
2136 * so could be changed anytime.
2138 if (order >= pageblock_order)
2141 if (order >= pageblock_order / 2 ||
2142 start_mt == MIGRATE_RECLAIMABLE ||
2143 start_mt == MIGRATE_UNMOVABLE ||
2144 page_group_by_mobility_disabled)
2151 * This function implements actual steal behaviour. If order is large enough,
2152 * we can steal whole pageblock. If not, we first move freepages in this
2153 * pageblock to our migratetype and determine how many already-allocated pages
2154 * are there in the pageblock with a compatible migratetype. If at least half
2155 * of pages are free or compatible, we can change migratetype of the pageblock
2156 * itself, so pages freed in the future will be put on the correct free list.
2158 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2159 int start_type, bool whole_block)
2161 unsigned int current_order = page_order(page);
2162 struct free_area *area;
2163 int free_pages, movable_pages, alike_pages;
2166 old_block_type = get_pageblock_migratetype(page);
2169 * This can happen due to races and we want to prevent broken
2170 * highatomic accounting.
2172 if (is_migrate_highatomic(old_block_type))
2175 /* Take ownership for orders >= pageblock_order */
2176 if (current_order >= pageblock_order) {
2177 change_pageblock_range(page, current_order, start_type);
2181 /* We are not allowed to try stealing from the whole block */
2185 free_pages = move_freepages_block(zone, page, start_type,
2188 * Determine how many pages are compatible with our allocation.
2189 * For movable allocation, it's the number of movable pages which
2190 * we just obtained. For other types it's a bit more tricky.
2192 if (start_type == MIGRATE_MOVABLE) {
2193 alike_pages = movable_pages;
2196 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2197 * to MOVABLE pageblock, consider all non-movable pages as
2198 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2199 * vice versa, be conservative since we can't distinguish the
2200 * exact migratetype of non-movable pages.
2202 if (old_block_type == MIGRATE_MOVABLE)
2203 alike_pages = pageblock_nr_pages
2204 - (free_pages + movable_pages);
2209 /* moving whole block can fail due to zone boundary conditions */
2214 * If a sufficient number of pages in the block are either free or of
2215 * comparable migratability as our allocation, claim the whole block.
2217 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2218 page_group_by_mobility_disabled)
2219 set_pageblock_migratetype(page, start_type);
2224 area = &zone->free_area[current_order];
2225 list_move(&page->lru, &area->free_list[start_type]);
2229 * Check whether there is a suitable fallback freepage with requested order.
2230 * If only_stealable is true, this function returns fallback_mt only if
2231 * we can steal other freepages all together. This would help to reduce
2232 * fragmentation due to mixed migratetype pages in one pageblock.
2234 int find_suitable_fallback(struct free_area *area, unsigned int order,
2235 int migratetype, bool only_stealable, bool *can_steal)
2240 if (area->nr_free == 0)
2245 fallback_mt = fallbacks[migratetype][i];
2246 if (fallback_mt == MIGRATE_TYPES)
2249 if (list_empty(&area->free_list[fallback_mt]))
2252 if (can_steal_fallback(order, migratetype))
2255 if (!only_stealable)
2266 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2267 * there are no empty page blocks that contain a page with a suitable order
2269 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2270 unsigned int alloc_order)
2273 unsigned long max_managed, flags;
2276 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2277 * Check is race-prone but harmless.
2279 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2280 if (zone->nr_reserved_highatomic >= max_managed)
2283 spin_lock_irqsave(&zone->lock, flags);
2285 /* Recheck the nr_reserved_highatomic limit under the lock */
2286 if (zone->nr_reserved_highatomic >= max_managed)
2290 mt = get_pageblock_migratetype(page);
2291 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2292 && !is_migrate_cma(mt)) {
2293 zone->nr_reserved_highatomic += pageblock_nr_pages;
2294 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2295 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2299 spin_unlock_irqrestore(&zone->lock, flags);
2303 * Used when an allocation is about to fail under memory pressure. This
2304 * potentially hurts the reliability of high-order allocations when under
2305 * intense memory pressure but failed atomic allocations should be easier
2306 * to recover from than an OOM.
2308 * If @force is true, try to unreserve a pageblock even though highatomic
2309 * pageblock is exhausted.
2311 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2314 struct zonelist *zonelist = ac->zonelist;
2315 unsigned long flags;
2322 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2325 * Preserve at least one pageblock unless memory pressure
2328 if (!force && zone->nr_reserved_highatomic <=
2332 spin_lock_irqsave(&zone->lock, flags);
2333 for (order = 0; order < MAX_ORDER; order++) {
2334 struct free_area *area = &(zone->free_area[order]);
2336 page = list_first_entry_or_null(
2337 &area->free_list[MIGRATE_HIGHATOMIC],
2343 * In page freeing path, migratetype change is racy so
2344 * we can counter several free pages in a pageblock
2345 * in this loop althoug we changed the pageblock type
2346 * from highatomic to ac->migratetype. So we should
2347 * adjust the count once.
2349 if (is_migrate_highatomic_page(page)) {
2351 * It should never happen but changes to
2352 * locking could inadvertently allow a per-cpu
2353 * drain to add pages to MIGRATE_HIGHATOMIC
2354 * while unreserving so be safe and watch for
2357 zone->nr_reserved_highatomic -= min(
2359 zone->nr_reserved_highatomic);
2363 * Convert to ac->migratetype and avoid the normal
2364 * pageblock stealing heuristics. Minimally, the caller
2365 * is doing the work and needs the pages. More
2366 * importantly, if the block was always converted to
2367 * MIGRATE_UNMOVABLE or another type then the number
2368 * of pageblocks that cannot be completely freed
2371 set_pageblock_migratetype(page, ac->migratetype);
2372 ret = move_freepages_block(zone, page, ac->migratetype,
2375 spin_unlock_irqrestore(&zone->lock, flags);
2379 spin_unlock_irqrestore(&zone->lock, flags);
2386 * Try finding a free buddy page on the fallback list and put it on the free
2387 * list of requested migratetype, possibly along with other pages from the same
2388 * block, depending on fragmentation avoidance heuristics. Returns true if
2389 * fallback was found so that __rmqueue_smallest() can grab it.
2391 * The use of signed ints for order and current_order is a deliberate
2392 * deviation from the rest of this file, to make the for loop
2393 * condition simpler.
2395 static __always_inline bool
2396 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2398 struct free_area *area;
2405 * Find the largest available free page in the other list. This roughly
2406 * approximates finding the pageblock with the most free pages, which
2407 * would be too costly to do exactly.
2409 for (current_order = MAX_ORDER - 1; current_order >= order;
2411 area = &(zone->free_area[current_order]);
2412 fallback_mt = find_suitable_fallback(area, current_order,
2413 start_migratetype, false, &can_steal);
2414 if (fallback_mt == -1)
2418 * We cannot steal all free pages from the pageblock and the
2419 * requested migratetype is movable. In that case it's better to
2420 * steal and split the smallest available page instead of the
2421 * largest available page, because even if the next movable
2422 * allocation falls back into a different pageblock than this
2423 * one, it won't cause permanent fragmentation.
2425 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2426 && current_order > order)
2435 for (current_order = order; current_order < MAX_ORDER;
2437 area = &(zone->free_area[current_order]);
2438 fallback_mt = find_suitable_fallback(area, current_order,
2439 start_migratetype, false, &can_steal);
2440 if (fallback_mt != -1)
2445 * This should not happen - we already found a suitable fallback
2446 * when looking for the largest page.
2448 VM_BUG_ON(current_order == MAX_ORDER);
2451 page = list_first_entry(&area->free_list[fallback_mt],
2454 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2456 trace_mm_page_alloc_extfrag(page, order, current_order,
2457 start_migratetype, fallback_mt);
2464 * Do the hard work of removing an element from the buddy allocator.
2465 * Call me with the zone->lock already held.
2467 static __always_inline struct page *
2468 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2473 page = __rmqueue_smallest(zone, order, migratetype);
2474 if (unlikely(!page)) {
2475 if (migratetype == MIGRATE_MOVABLE)
2476 page = __rmqueue_cma_fallback(zone, order);
2478 if (!page && __rmqueue_fallback(zone, order, migratetype))
2482 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2487 * Obtain a specified number of elements from the buddy allocator, all under
2488 * a single hold of the lock, for efficiency. Add them to the supplied list.
2489 * Returns the number of new pages which were placed at *list.
2491 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2492 unsigned long count, struct list_head *list,
2497 spin_lock(&zone->lock);
2498 for (i = 0; i < count; ++i) {
2499 struct page *page = __rmqueue(zone, order, migratetype);
2500 if (unlikely(page == NULL))
2503 if (unlikely(check_pcp_refill(page)))
2507 * Split buddy pages returned by expand() are received here in
2508 * physical page order. The page is added to the tail of
2509 * caller's list. From the callers perspective, the linked list
2510 * is ordered by page number under some conditions. This is
2511 * useful for IO devices that can forward direction from the
2512 * head, thus also in the physical page order. This is useful
2513 * for IO devices that can merge IO requests if the physical
2514 * pages are ordered properly.
2516 list_add_tail(&page->lru, list);
2518 if (is_migrate_cma(get_pcppage_migratetype(page)))
2519 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2524 * i pages were removed from the buddy list even if some leak due
2525 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2526 * on i. Do not confuse with 'alloced' which is the number of
2527 * pages added to the pcp list.
2529 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2530 spin_unlock(&zone->lock);
2536 * Called from the vmstat counter updater to drain pagesets of this
2537 * currently executing processor on remote nodes after they have
2540 * Note that this function must be called with the thread pinned to
2541 * a single processor.
2543 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2545 unsigned long flags;
2546 int to_drain, batch;
2548 local_irq_save(flags);
2549 batch = READ_ONCE(pcp->batch);
2550 to_drain = min(pcp->count, batch);
2552 free_pcppages_bulk(zone, to_drain, pcp);
2553 local_irq_restore(flags);
2558 * Drain pcplists of the indicated processor and zone.
2560 * The processor must either be the current processor and the
2561 * thread pinned to the current processor or a processor that
2564 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2566 unsigned long flags;
2567 struct per_cpu_pageset *pset;
2568 struct per_cpu_pages *pcp;
2570 local_irq_save(flags);
2571 pset = per_cpu_ptr(zone->pageset, cpu);
2575 free_pcppages_bulk(zone, pcp->count, pcp);
2576 local_irq_restore(flags);
2580 * Drain pcplists of all zones on the indicated processor.
2582 * The processor must either be the current processor and the
2583 * thread pinned to the current processor or a processor that
2586 static void drain_pages(unsigned int cpu)
2590 for_each_populated_zone(zone) {
2591 drain_pages_zone(cpu, zone);
2596 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2598 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2599 * the single zone's pages.
2601 void drain_local_pages(struct zone *zone)
2603 int cpu = smp_processor_id();
2606 drain_pages_zone(cpu, zone);
2611 static void drain_local_pages_wq(struct work_struct *work)
2614 * drain_all_pages doesn't use proper cpu hotplug protection so
2615 * we can race with cpu offline when the WQ can move this from
2616 * a cpu pinned worker to an unbound one. We can operate on a different
2617 * cpu which is allright but we also have to make sure to not move to
2621 drain_local_pages(NULL);
2626 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2628 * When zone parameter is non-NULL, spill just the single zone's pages.
2630 * Note that this can be extremely slow as the draining happens in a workqueue.
2632 void drain_all_pages(struct zone *zone)
2637 * Allocate in the BSS so we wont require allocation in
2638 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2640 static cpumask_t cpus_with_pcps;
2643 * Make sure nobody triggers this path before mm_percpu_wq is fully
2646 if (WARN_ON_ONCE(!mm_percpu_wq))
2650 * Do not drain if one is already in progress unless it's specific to
2651 * a zone. Such callers are primarily CMA and memory hotplug and need
2652 * the drain to be complete when the call returns.
2654 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2657 mutex_lock(&pcpu_drain_mutex);
2661 * We don't care about racing with CPU hotplug event
2662 * as offline notification will cause the notified
2663 * cpu to drain that CPU pcps and on_each_cpu_mask
2664 * disables preemption as part of its processing
2666 for_each_online_cpu(cpu) {
2667 struct per_cpu_pageset *pcp;
2669 bool has_pcps = false;
2672 pcp = per_cpu_ptr(zone->pageset, cpu);
2676 for_each_populated_zone(z) {
2677 pcp = per_cpu_ptr(z->pageset, cpu);
2678 if (pcp->pcp.count) {
2686 cpumask_set_cpu(cpu, &cpus_with_pcps);
2688 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2691 for_each_cpu(cpu, &cpus_with_pcps) {
2692 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2693 INIT_WORK(work, drain_local_pages_wq);
2694 queue_work_on(cpu, mm_percpu_wq, work);
2696 for_each_cpu(cpu, &cpus_with_pcps)
2697 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2699 mutex_unlock(&pcpu_drain_mutex);
2702 #ifdef CONFIG_HIBERNATION
2705 * Touch the watchdog for every WD_PAGE_COUNT pages.
2707 #define WD_PAGE_COUNT (128*1024)
2709 void mark_free_pages(struct zone *zone)
2711 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2712 unsigned long flags;
2713 unsigned int order, t;
2716 if (zone_is_empty(zone))
2719 spin_lock_irqsave(&zone->lock, flags);
2721 max_zone_pfn = zone_end_pfn(zone);
2722 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2723 if (pfn_valid(pfn)) {
2724 page = pfn_to_page(pfn);
2726 if (!--page_count) {
2727 touch_nmi_watchdog();
2728 page_count = WD_PAGE_COUNT;
2731 if (page_zone(page) != zone)
2734 if (!swsusp_page_is_forbidden(page))
2735 swsusp_unset_page_free(page);
2738 for_each_migratetype_order(order, t) {
2739 list_for_each_entry(page,
2740 &zone->free_area[order].free_list[t], lru) {
2743 pfn = page_to_pfn(page);
2744 for (i = 0; i < (1UL << order); i++) {
2745 if (!--page_count) {
2746 touch_nmi_watchdog();
2747 page_count = WD_PAGE_COUNT;
2749 swsusp_set_page_free(pfn_to_page(pfn + i));
2753 spin_unlock_irqrestore(&zone->lock, flags);
2755 #endif /* CONFIG_PM */
2757 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2761 if (!free_pcp_prepare(page))
2764 migratetype = get_pfnblock_migratetype(page, pfn);
2765 set_pcppage_migratetype(page, migratetype);
2769 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2771 struct zone *zone = page_zone(page);
2772 struct per_cpu_pages *pcp;
2775 migratetype = get_pcppage_migratetype(page);
2776 __count_vm_event(PGFREE);
2779 * We only track unmovable, reclaimable and movable on pcp lists.
2780 * Free ISOLATE pages back to the allocator because they are being
2781 * offlined but treat HIGHATOMIC as movable pages so we can get those
2782 * areas back if necessary. Otherwise, we may have to free
2783 * excessively into the page allocator
2785 if (migratetype >= MIGRATE_PCPTYPES) {
2786 if (unlikely(is_migrate_isolate(migratetype))) {
2787 free_one_page(zone, page, pfn, 0, migratetype);
2790 migratetype = MIGRATE_MOVABLE;
2793 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2794 list_add(&page->lru, &pcp->lists[migratetype]);
2796 if (pcp->count >= pcp->high) {
2797 unsigned long batch = READ_ONCE(pcp->batch);
2798 free_pcppages_bulk(zone, batch, pcp);
2803 * Free a 0-order page
2805 void free_unref_page(struct page *page)
2807 unsigned long flags;
2808 unsigned long pfn = page_to_pfn(page);
2810 if (!free_unref_page_prepare(page, pfn))
2813 local_irq_save(flags);
2814 free_unref_page_commit(page, pfn);
2815 local_irq_restore(flags);
2819 * Free a list of 0-order pages
2821 void free_unref_page_list(struct list_head *list)
2823 struct page *page, *next;
2824 unsigned long flags, pfn;
2825 int batch_count = 0;
2827 /* Prepare pages for freeing */
2828 list_for_each_entry_safe(page, next, list, lru) {
2829 pfn = page_to_pfn(page);
2830 if (!free_unref_page_prepare(page, pfn))
2831 list_del(&page->lru);
2832 set_page_private(page, pfn);
2835 local_irq_save(flags);
2836 list_for_each_entry_safe(page, next, list, lru) {
2837 unsigned long pfn = page_private(page);
2839 set_page_private(page, 0);
2840 trace_mm_page_free_batched(page);
2841 free_unref_page_commit(page, pfn);
2844 * Guard against excessive IRQ disabled times when we get
2845 * a large list of pages to free.
2847 if (++batch_count == SWAP_CLUSTER_MAX) {
2848 local_irq_restore(flags);
2850 local_irq_save(flags);
2853 local_irq_restore(flags);
2857 * split_page takes a non-compound higher-order page, and splits it into
2858 * n (1<<order) sub-pages: page[0..n]
2859 * Each sub-page must be freed individually.
2861 * Note: this is probably too low level an operation for use in drivers.
2862 * Please consult with lkml before using this in your driver.
2864 void split_page(struct page *page, unsigned int order)
2868 VM_BUG_ON_PAGE(PageCompound(page), page);
2869 VM_BUG_ON_PAGE(!page_count(page), page);
2871 for (i = 1; i < (1 << order); i++)
2872 set_page_refcounted(page + i);
2873 split_page_owner(page, order);
2875 EXPORT_SYMBOL_GPL(split_page);
2877 int __isolate_free_page(struct page *page, unsigned int order)
2879 unsigned long watermark;
2883 BUG_ON(!PageBuddy(page));
2885 zone = page_zone(page);
2886 mt = get_pageblock_migratetype(page);
2888 if (!is_migrate_isolate(mt)) {
2890 * Obey watermarks as if the page was being allocated. We can
2891 * emulate a high-order watermark check with a raised order-0
2892 * watermark, because we already know our high-order page
2895 watermark = min_wmark_pages(zone) + (1UL << order);
2896 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2899 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2902 /* Remove page from free list */
2903 list_del(&page->lru);
2904 zone->free_area[order].nr_free--;
2905 rmv_page_order(page);
2908 * Set the pageblock if the isolated page is at least half of a
2911 if (order >= pageblock_order - 1) {
2912 struct page *endpage = page + (1 << order) - 1;
2913 for (; page < endpage; page += pageblock_nr_pages) {
2914 int mt = get_pageblock_migratetype(page);
2915 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2916 && !is_migrate_highatomic(mt))
2917 set_pageblock_migratetype(page,
2923 return 1UL << order;
2927 * Update NUMA hit/miss statistics
2929 * Must be called with interrupts disabled.
2931 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2934 enum numa_stat_item local_stat = NUMA_LOCAL;
2936 /* skip numa counters update if numa stats is disabled */
2937 if (!static_branch_likely(&vm_numa_stat_key))
2940 if (zone_to_nid(z) != numa_node_id())
2941 local_stat = NUMA_OTHER;
2943 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2944 __inc_numa_state(z, NUMA_HIT);
2946 __inc_numa_state(z, NUMA_MISS);
2947 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2949 __inc_numa_state(z, local_stat);
2953 /* Remove page from the per-cpu list, caller must protect the list */
2954 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2955 struct per_cpu_pages *pcp,
2956 struct list_head *list)
2961 if (list_empty(list)) {
2962 pcp->count += rmqueue_bulk(zone, 0,
2965 if (unlikely(list_empty(list)))
2969 page = list_first_entry(list, struct page, lru);
2970 list_del(&page->lru);
2972 } while (check_new_pcp(page));
2977 /* Lock and remove page from the per-cpu list */
2978 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2979 struct zone *zone, unsigned int order,
2980 gfp_t gfp_flags, int migratetype)
2982 struct per_cpu_pages *pcp;
2983 struct list_head *list;
2985 unsigned long flags;
2987 local_irq_save(flags);
2988 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2989 list = &pcp->lists[migratetype];
2990 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2992 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2993 zone_statistics(preferred_zone, zone);
2995 local_irq_restore(flags);
3000 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3003 struct page *rmqueue(struct zone *preferred_zone,
3004 struct zone *zone, unsigned int order,
3005 gfp_t gfp_flags, unsigned int alloc_flags,
3008 unsigned long flags;
3011 if (likely(order == 0)) {
3012 page = rmqueue_pcplist(preferred_zone, zone, order,
3013 gfp_flags, migratetype);
3018 * We most definitely don't want callers attempting to
3019 * allocate greater than order-1 page units with __GFP_NOFAIL.
3021 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3022 spin_lock_irqsave(&zone->lock, flags);
3026 if (alloc_flags & ALLOC_HARDER) {
3027 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3029 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3032 page = __rmqueue(zone, order, migratetype);
3033 } while (page && check_new_pages(page, order));
3034 spin_unlock(&zone->lock);
3037 __mod_zone_freepage_state(zone, -(1 << order),
3038 get_pcppage_migratetype(page));
3040 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3041 zone_statistics(preferred_zone, zone);
3042 local_irq_restore(flags);
3045 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3049 local_irq_restore(flags);
3053 #ifdef CONFIG_FAIL_PAGE_ALLOC
3056 struct fault_attr attr;
3058 bool ignore_gfp_highmem;
3059 bool ignore_gfp_reclaim;
3061 } fail_page_alloc = {
3062 .attr = FAULT_ATTR_INITIALIZER,
3063 .ignore_gfp_reclaim = true,
3064 .ignore_gfp_highmem = true,
3068 static int __init setup_fail_page_alloc(char *str)
3070 return setup_fault_attr(&fail_page_alloc.attr, str);
3072 __setup("fail_page_alloc=", setup_fail_page_alloc);
3074 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3076 if (order < fail_page_alloc.min_order)
3078 if (gfp_mask & __GFP_NOFAIL)
3080 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3082 if (fail_page_alloc.ignore_gfp_reclaim &&
3083 (gfp_mask & __GFP_DIRECT_RECLAIM))
3086 return should_fail(&fail_page_alloc.attr, 1 << order);
3089 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3091 static int __init fail_page_alloc_debugfs(void)
3093 umode_t mode = S_IFREG | 0600;
3096 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3097 &fail_page_alloc.attr);
3099 return PTR_ERR(dir);
3101 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3102 &fail_page_alloc.ignore_gfp_reclaim))
3104 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3105 &fail_page_alloc.ignore_gfp_highmem))
3107 if (!debugfs_create_u32("min-order", mode, dir,
3108 &fail_page_alloc.min_order))
3113 debugfs_remove_recursive(dir);
3118 late_initcall(fail_page_alloc_debugfs);
3120 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3122 #else /* CONFIG_FAIL_PAGE_ALLOC */
3124 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3129 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3132 * Return true if free base pages are above 'mark'. For high-order checks it
3133 * will return true of the order-0 watermark is reached and there is at least
3134 * one free page of a suitable size. Checking now avoids taking the zone lock
3135 * to check in the allocation paths if no pages are free.
3137 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3138 int classzone_idx, unsigned int alloc_flags,
3143 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3145 /* free_pages may go negative - that's OK */
3146 free_pages -= (1 << order) - 1;
3148 if (alloc_flags & ALLOC_HIGH)
3152 * If the caller does not have rights to ALLOC_HARDER then subtract
3153 * the high-atomic reserves. This will over-estimate the size of the
3154 * atomic reserve but it avoids a search.
3156 if (likely(!alloc_harder)) {
3157 free_pages -= z->nr_reserved_highatomic;
3160 * OOM victims can try even harder than normal ALLOC_HARDER
3161 * users on the grounds that it's definitely going to be in
3162 * the exit path shortly and free memory. Any allocation it
3163 * makes during the free path will be small and short-lived.
3165 if (alloc_flags & ALLOC_OOM)
3173 /* If allocation can't use CMA areas don't use free CMA pages */
3174 if (!(alloc_flags & ALLOC_CMA))
3175 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3179 * Check watermarks for an order-0 allocation request. If these
3180 * are not met, then a high-order request also cannot go ahead
3181 * even if a suitable page happened to be free.
3183 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3186 /* If this is an order-0 request then the watermark is fine */
3190 /* For a high-order request, check at least one suitable page is free */
3191 for (o = order; o < MAX_ORDER; o++) {
3192 struct free_area *area = &z->free_area[o];
3198 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3199 if (!list_empty(&area->free_list[mt]))
3204 if ((alloc_flags & ALLOC_CMA) &&
3205 !list_empty(&area->free_list[MIGRATE_CMA])) {
3210 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3216 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3217 int classzone_idx, unsigned int alloc_flags)
3219 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3220 zone_page_state(z, NR_FREE_PAGES));
3223 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3224 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3226 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3230 /* If allocation can't use CMA areas don't use free CMA pages */
3231 if (!(alloc_flags & ALLOC_CMA))
3232 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3236 * Fast check for order-0 only. If this fails then the reserves
3237 * need to be calculated. There is a corner case where the check
3238 * passes but only the high-order atomic reserve are free. If
3239 * the caller is !atomic then it'll uselessly search the free
3240 * list. That corner case is then slower but it is harmless.
3242 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3245 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3249 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3250 unsigned long mark, int classzone_idx)
3252 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3254 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3255 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3257 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3262 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3264 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3267 #else /* CONFIG_NUMA */
3268 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3272 #endif /* CONFIG_NUMA */
3275 * get_page_from_freelist goes through the zonelist trying to allocate
3278 static struct page *
3279 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3280 const struct alloc_context *ac)
3282 struct zoneref *z = ac->preferred_zoneref;
3284 struct pglist_data *last_pgdat_dirty_limit = NULL;
3287 * Scan zonelist, looking for a zone with enough free.
3288 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3290 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3295 if (cpusets_enabled() &&
3296 (alloc_flags & ALLOC_CPUSET) &&
3297 !__cpuset_zone_allowed(zone, gfp_mask))
3300 * When allocating a page cache page for writing, we
3301 * want to get it from a node that is within its dirty
3302 * limit, such that no single node holds more than its
3303 * proportional share of globally allowed dirty pages.
3304 * The dirty limits take into account the node's
3305 * lowmem reserves and high watermark so that kswapd
3306 * should be able to balance it without having to
3307 * write pages from its LRU list.
3309 * XXX: For now, allow allocations to potentially
3310 * exceed the per-node dirty limit in the slowpath
3311 * (spread_dirty_pages unset) before going into reclaim,
3312 * which is important when on a NUMA setup the allowed
3313 * nodes are together not big enough to reach the
3314 * global limit. The proper fix for these situations
3315 * will require awareness of nodes in the
3316 * dirty-throttling and the flusher threads.
3318 if (ac->spread_dirty_pages) {
3319 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3322 if (!node_dirty_ok(zone->zone_pgdat)) {
3323 last_pgdat_dirty_limit = zone->zone_pgdat;
3328 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3329 if (!zone_watermark_fast(zone, order, mark,
3330 ac_classzone_idx(ac), alloc_flags)) {
3333 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3335 * Watermark failed for this zone, but see if we can
3336 * grow this zone if it contains deferred pages.
3338 if (static_branch_unlikely(&deferred_pages)) {
3339 if (_deferred_grow_zone(zone, order))
3343 /* Checked here to keep the fast path fast */
3344 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3345 if (alloc_flags & ALLOC_NO_WATERMARKS)
3348 if (node_reclaim_mode == 0 ||
3349 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3352 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3354 case NODE_RECLAIM_NOSCAN:
3357 case NODE_RECLAIM_FULL:
3358 /* scanned but unreclaimable */
3361 /* did we reclaim enough */
3362 if (zone_watermark_ok(zone, order, mark,
3363 ac_classzone_idx(ac), alloc_flags))
3371 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3372 gfp_mask, alloc_flags, ac->migratetype);
3374 prep_new_page(page, order, gfp_mask, alloc_flags);
3377 * If this is a high-order atomic allocation then check
3378 * if the pageblock should be reserved for the future
3380 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3381 reserve_highatomic_pageblock(page, zone, order);
3385 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3386 /* Try again if zone has deferred pages */
3387 if (static_branch_unlikely(&deferred_pages)) {
3388 if (_deferred_grow_zone(zone, order))
3399 * Large machines with many possible nodes should not always dump per-node
3400 * meminfo in irq context.
3402 static inline bool should_suppress_show_mem(void)
3407 ret = in_interrupt();
3412 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3414 unsigned int filter = SHOW_MEM_FILTER_NODES;
3415 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3417 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3421 * This documents exceptions given to allocations in certain
3422 * contexts that are allowed to allocate outside current's set
3425 if (!(gfp_mask & __GFP_NOMEMALLOC))
3426 if (tsk_is_oom_victim(current) ||
3427 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3428 filter &= ~SHOW_MEM_FILTER_NODES;
3429 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3430 filter &= ~SHOW_MEM_FILTER_NODES;
3432 show_mem(filter, nodemask);
3435 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3437 struct va_format vaf;
3439 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3440 DEFAULT_RATELIMIT_BURST);
3442 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3445 va_start(args, fmt);
3448 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3449 current->comm, &vaf, gfp_mask, &gfp_mask,
3450 nodemask_pr_args(nodemask));
3453 cpuset_print_current_mems_allowed();
3456 warn_alloc_show_mem(gfp_mask, nodemask);
3459 static inline struct page *
3460 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3461 unsigned int alloc_flags,
3462 const struct alloc_context *ac)
3466 page = get_page_from_freelist(gfp_mask, order,
3467 alloc_flags|ALLOC_CPUSET, ac);
3469 * fallback to ignore cpuset restriction if our nodes
3473 page = get_page_from_freelist(gfp_mask, order,
3479 static inline struct page *
3480 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3481 const struct alloc_context *ac, unsigned long *did_some_progress)
3483 struct oom_control oc = {
3484 .zonelist = ac->zonelist,
3485 .nodemask = ac->nodemask,
3487 .gfp_mask = gfp_mask,
3492 *did_some_progress = 0;
3495 * Acquire the oom lock. If that fails, somebody else is
3496 * making progress for us.
3498 if (!mutex_trylock(&oom_lock)) {
3499 *did_some_progress = 1;
3500 schedule_timeout_uninterruptible(1);
3505 * Go through the zonelist yet one more time, keep very high watermark
3506 * here, this is only to catch a parallel oom killing, we must fail if
3507 * we're still under heavy pressure. But make sure that this reclaim
3508 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3509 * allocation which will never fail due to oom_lock already held.
3511 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3512 ~__GFP_DIRECT_RECLAIM, order,
3513 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3517 /* Coredumps can quickly deplete all memory reserves */
3518 if (current->flags & PF_DUMPCORE)
3520 /* The OOM killer will not help higher order allocs */
3521 if (order > PAGE_ALLOC_COSTLY_ORDER)
3524 * We have already exhausted all our reclaim opportunities without any
3525 * success so it is time to admit defeat. We will skip the OOM killer
3526 * because it is very likely that the caller has a more reasonable
3527 * fallback than shooting a random task.
3529 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3531 /* The OOM killer does not needlessly kill tasks for lowmem */
3532 if (ac->high_zoneidx < ZONE_NORMAL)
3534 if (pm_suspended_storage())
3537 * XXX: GFP_NOFS allocations should rather fail than rely on
3538 * other request to make a forward progress.
3539 * We are in an unfortunate situation where out_of_memory cannot
3540 * do much for this context but let's try it to at least get
3541 * access to memory reserved if the current task is killed (see
3542 * out_of_memory). Once filesystems are ready to handle allocation
3543 * failures more gracefully we should just bail out here.
3546 /* The OOM killer may not free memory on a specific node */
3547 if (gfp_mask & __GFP_THISNODE)
3550 /* Exhausted what can be done so it's blame time */
3551 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3552 *did_some_progress = 1;
3555 * Help non-failing allocations by giving them access to memory
3558 if (gfp_mask & __GFP_NOFAIL)
3559 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3560 ALLOC_NO_WATERMARKS, ac);
3563 mutex_unlock(&oom_lock);
3568 * Maximum number of compaction retries wit a progress before OOM
3569 * killer is consider as the only way to move forward.
3571 #define MAX_COMPACT_RETRIES 16
3573 #ifdef CONFIG_COMPACTION
3574 /* Try memory compaction for high-order allocations before reclaim */
3575 static struct page *
3576 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3577 unsigned int alloc_flags, const struct alloc_context *ac,
3578 enum compact_priority prio, enum compact_result *compact_result)
3581 unsigned int noreclaim_flag;
3586 noreclaim_flag = memalloc_noreclaim_save();
3587 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3589 memalloc_noreclaim_restore(noreclaim_flag);
3591 if (*compact_result <= COMPACT_INACTIVE)
3595 * At least in one zone compaction wasn't deferred or skipped, so let's
3596 * count a compaction stall
3598 count_vm_event(COMPACTSTALL);
3600 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3603 struct zone *zone = page_zone(page);
3605 zone->compact_blockskip_flush = false;
3606 compaction_defer_reset(zone, order, true);
3607 count_vm_event(COMPACTSUCCESS);
3612 * It's bad if compaction run occurs and fails. The most likely reason
3613 * is that pages exist, but not enough to satisfy watermarks.
3615 count_vm_event(COMPACTFAIL);
3623 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3624 enum compact_result compact_result,
3625 enum compact_priority *compact_priority,
3626 int *compaction_retries)
3628 int max_retries = MAX_COMPACT_RETRIES;
3631 int retries = *compaction_retries;
3632 enum compact_priority priority = *compact_priority;
3637 if (compaction_made_progress(compact_result))
3638 (*compaction_retries)++;
3641 * compaction considers all the zone as desperately out of memory
3642 * so it doesn't really make much sense to retry except when the
3643 * failure could be caused by insufficient priority
3645 if (compaction_failed(compact_result))
3646 goto check_priority;
3649 * make sure the compaction wasn't deferred or didn't bail out early
3650 * due to locks contention before we declare that we should give up.
3651 * But do not retry if the given zonelist is not suitable for
3654 if (compaction_withdrawn(compact_result)) {
3655 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3660 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3661 * costly ones because they are de facto nofail and invoke OOM
3662 * killer to move on while costly can fail and users are ready
3663 * to cope with that. 1/4 retries is rather arbitrary but we
3664 * would need much more detailed feedback from compaction to
3665 * make a better decision.
3667 if (order > PAGE_ALLOC_COSTLY_ORDER)
3669 if (*compaction_retries <= max_retries) {
3675 * Make sure there are attempts at the highest priority if we exhausted
3676 * all retries or failed at the lower priorities.
3679 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3680 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3682 if (*compact_priority > min_priority) {
3683 (*compact_priority)--;
3684 *compaction_retries = 0;
3688 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3692 static inline struct page *
3693 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3694 unsigned int alloc_flags, const struct alloc_context *ac,
3695 enum compact_priority prio, enum compact_result *compact_result)
3697 *compact_result = COMPACT_SKIPPED;
3702 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3703 enum compact_result compact_result,
3704 enum compact_priority *compact_priority,
3705 int *compaction_retries)
3710 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3714 * There are setups with compaction disabled which would prefer to loop
3715 * inside the allocator rather than hit the oom killer prematurely.
3716 * Let's give them a good hope and keep retrying while the order-0
3717 * watermarks are OK.
3719 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3721 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3722 ac_classzone_idx(ac), alloc_flags))
3727 #endif /* CONFIG_COMPACTION */
3729 #ifdef CONFIG_LOCKDEP
3730 static struct lockdep_map __fs_reclaim_map =
3731 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3733 static bool __need_fs_reclaim(gfp_t gfp_mask)
3735 gfp_mask = current_gfp_context(gfp_mask);
3737 /* no reclaim without waiting on it */
3738 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3741 /* this guy won't enter reclaim */
3742 if (current->flags & PF_MEMALLOC)
3745 /* We're only interested __GFP_FS allocations for now */
3746 if (!(gfp_mask & __GFP_FS))
3749 if (gfp_mask & __GFP_NOLOCKDEP)
3755 void __fs_reclaim_acquire(void)
3757 lock_map_acquire(&__fs_reclaim_map);
3760 void __fs_reclaim_release(void)
3762 lock_map_release(&__fs_reclaim_map);
3765 void fs_reclaim_acquire(gfp_t gfp_mask)
3767 if (__need_fs_reclaim(gfp_mask))
3768 __fs_reclaim_acquire();
3770 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3772 void fs_reclaim_release(gfp_t gfp_mask)
3774 if (__need_fs_reclaim(gfp_mask))
3775 __fs_reclaim_release();
3777 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3780 /* Perform direct synchronous page reclaim */
3782 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3783 const struct alloc_context *ac)
3785 struct reclaim_state reclaim_state;
3787 unsigned int noreclaim_flag;
3791 /* We now go into synchronous reclaim */
3792 cpuset_memory_pressure_bump();
3793 fs_reclaim_acquire(gfp_mask);
3794 noreclaim_flag = memalloc_noreclaim_save();
3795 reclaim_state.reclaimed_slab = 0;
3796 current->reclaim_state = &reclaim_state;
3798 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3801 current->reclaim_state = NULL;
3802 memalloc_noreclaim_restore(noreclaim_flag);
3803 fs_reclaim_release(gfp_mask);
3810 /* The really slow allocator path where we enter direct reclaim */
3811 static inline struct page *
3812 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3813 unsigned int alloc_flags, const struct alloc_context *ac,
3814 unsigned long *did_some_progress)
3816 struct page *page = NULL;
3817 bool drained = false;
3819 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3820 if (unlikely(!(*did_some_progress)))
3824 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3827 * If an allocation failed after direct reclaim, it could be because
3828 * pages are pinned on the per-cpu lists or in high alloc reserves.
3829 * Shrink them them and try again
3831 if (!page && !drained) {
3832 unreserve_highatomic_pageblock(ac, false);
3833 drain_all_pages(NULL);
3841 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3842 const struct alloc_context *ac)
3846 pg_data_t *last_pgdat = NULL;
3847 enum zone_type high_zoneidx = ac->high_zoneidx;
3849 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3851 if (last_pgdat != zone->zone_pgdat)
3852 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3853 last_pgdat = zone->zone_pgdat;
3857 static inline unsigned int
3858 gfp_to_alloc_flags(gfp_t gfp_mask)
3860 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3862 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3863 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3866 * The caller may dip into page reserves a bit more if the caller
3867 * cannot run direct reclaim, or if the caller has realtime scheduling
3868 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3869 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3871 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3873 if (gfp_mask & __GFP_ATOMIC) {
3875 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3876 * if it can't schedule.
3878 if (!(gfp_mask & __GFP_NOMEMALLOC))
3879 alloc_flags |= ALLOC_HARDER;
3881 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3882 * comment for __cpuset_node_allowed().
3884 alloc_flags &= ~ALLOC_CPUSET;
3885 } else if (unlikely(rt_task(current)) && !in_interrupt())
3886 alloc_flags |= ALLOC_HARDER;
3889 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3890 alloc_flags |= ALLOC_CMA;
3895 static bool oom_reserves_allowed(struct task_struct *tsk)
3897 if (!tsk_is_oom_victim(tsk))
3901 * !MMU doesn't have oom reaper so give access to memory reserves
3902 * only to the thread with TIF_MEMDIE set
3904 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3911 * Distinguish requests which really need access to full memory
3912 * reserves from oom victims which can live with a portion of it
3914 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3916 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3918 if (gfp_mask & __GFP_MEMALLOC)
3919 return ALLOC_NO_WATERMARKS;
3920 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3921 return ALLOC_NO_WATERMARKS;
3922 if (!in_interrupt()) {
3923 if (current->flags & PF_MEMALLOC)
3924 return ALLOC_NO_WATERMARKS;
3925 else if (oom_reserves_allowed(current))
3932 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3934 return !!__gfp_pfmemalloc_flags(gfp_mask);
3938 * Checks whether it makes sense to retry the reclaim to make a forward progress
3939 * for the given allocation request.
3941 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3942 * without success, or when we couldn't even meet the watermark if we
3943 * reclaimed all remaining pages on the LRU lists.
3945 * Returns true if a retry is viable or false to enter the oom path.
3948 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3949 struct alloc_context *ac, int alloc_flags,
3950 bool did_some_progress, int *no_progress_loops)
3956 * Costly allocations might have made a progress but this doesn't mean
3957 * their order will become available due to high fragmentation so
3958 * always increment the no progress counter for them
3960 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3961 *no_progress_loops = 0;
3963 (*no_progress_loops)++;
3966 * Make sure we converge to OOM if we cannot make any progress
3967 * several times in the row.
3969 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3970 /* Before OOM, exhaust highatomic_reserve */
3971 return unreserve_highatomic_pageblock(ac, true);
3975 * Keep reclaiming pages while there is a chance this will lead
3976 * somewhere. If none of the target zones can satisfy our allocation
3977 * request even if all reclaimable pages are considered then we are
3978 * screwed and have to go OOM.
3980 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3982 unsigned long available;
3983 unsigned long reclaimable;
3984 unsigned long min_wmark = min_wmark_pages(zone);
3987 available = reclaimable = zone_reclaimable_pages(zone);
3988 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3991 * Would the allocation succeed if we reclaimed all
3992 * reclaimable pages?
3994 wmark = __zone_watermark_ok(zone, order, min_wmark,
3995 ac_classzone_idx(ac), alloc_flags, available);
3996 trace_reclaim_retry_zone(z, order, reclaimable,
3997 available, min_wmark, *no_progress_loops, wmark);
4000 * If we didn't make any progress and have a lot of
4001 * dirty + writeback pages then we should wait for
4002 * an IO to complete to slow down the reclaim and
4003 * prevent from pre mature OOM
4005 if (!did_some_progress) {
4006 unsigned long write_pending;
4008 write_pending = zone_page_state_snapshot(zone,
4009 NR_ZONE_WRITE_PENDING);
4011 if (2 * write_pending > reclaimable) {
4012 congestion_wait(BLK_RW_ASYNC, HZ/10);
4018 * Memory allocation/reclaim might be called from a WQ
4019 * context and the current implementation of the WQ
4020 * concurrency control doesn't recognize that
4021 * a particular WQ is congested if the worker thread is
4022 * looping without ever sleeping. Therefore we have to
4023 * do a short sleep here rather than calling
4026 if (current->flags & PF_WQ_WORKER)
4027 schedule_timeout_uninterruptible(1);
4039 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4042 * It's possible that cpuset's mems_allowed and the nodemask from
4043 * mempolicy don't intersect. This should be normally dealt with by
4044 * policy_nodemask(), but it's possible to race with cpuset update in
4045 * such a way the check therein was true, and then it became false
4046 * before we got our cpuset_mems_cookie here.
4047 * This assumes that for all allocations, ac->nodemask can come only
4048 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4049 * when it does not intersect with the cpuset restrictions) or the
4050 * caller can deal with a violated nodemask.
4052 if (cpusets_enabled() && ac->nodemask &&
4053 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4054 ac->nodemask = NULL;
4059 * When updating a task's mems_allowed or mempolicy nodemask, it is
4060 * possible to race with parallel threads in such a way that our
4061 * allocation can fail while the mask is being updated. If we are about
4062 * to fail, check if the cpuset changed during allocation and if so,
4065 if (read_mems_allowed_retry(cpuset_mems_cookie))
4071 static inline struct page *
4072 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4073 struct alloc_context *ac)
4075 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4076 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4077 struct page *page = NULL;
4078 unsigned int alloc_flags;
4079 unsigned long did_some_progress;
4080 enum compact_priority compact_priority;
4081 enum compact_result compact_result;
4082 int compaction_retries;
4083 int no_progress_loops;
4084 unsigned int cpuset_mems_cookie;
4088 * We also sanity check to catch abuse of atomic reserves being used by
4089 * callers that are not in atomic context.
4091 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4092 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4093 gfp_mask &= ~__GFP_ATOMIC;
4096 compaction_retries = 0;
4097 no_progress_loops = 0;
4098 compact_priority = DEF_COMPACT_PRIORITY;
4099 cpuset_mems_cookie = read_mems_allowed_begin();
4102 * The fast path uses conservative alloc_flags to succeed only until
4103 * kswapd needs to be woken up, and to avoid the cost of setting up
4104 * alloc_flags precisely. So we do that now.
4106 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4109 * We need to recalculate the starting point for the zonelist iterator
4110 * because we might have used different nodemask in the fast path, or
4111 * there was a cpuset modification and we are retrying - otherwise we
4112 * could end up iterating over non-eligible zones endlessly.
4114 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4115 ac->high_zoneidx, ac->nodemask);
4116 if (!ac->preferred_zoneref->zone)
4119 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4120 wake_all_kswapds(order, gfp_mask, ac);
4123 * The adjusted alloc_flags might result in immediate success, so try
4126 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4131 * For costly allocations, try direct compaction first, as it's likely
4132 * that we have enough base pages and don't need to reclaim. For non-
4133 * movable high-order allocations, do that as well, as compaction will
4134 * try prevent permanent fragmentation by migrating from blocks of the
4136 * Don't try this for allocations that are allowed to ignore
4137 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4139 if (can_direct_reclaim &&
4141 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4142 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4143 page = __alloc_pages_direct_compact(gfp_mask, order,
4145 INIT_COMPACT_PRIORITY,
4151 * Checks for costly allocations with __GFP_NORETRY, which
4152 * includes THP page fault allocations
4154 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4156 * If compaction is deferred for high-order allocations,
4157 * it is because sync compaction recently failed. If
4158 * this is the case and the caller requested a THP
4159 * allocation, we do not want to heavily disrupt the
4160 * system, so we fail the allocation instead of entering
4163 if (compact_result == COMPACT_DEFERRED)
4167 * Looks like reclaim/compaction is worth trying, but
4168 * sync compaction could be very expensive, so keep
4169 * using async compaction.
4171 compact_priority = INIT_COMPACT_PRIORITY;
4176 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4177 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4178 wake_all_kswapds(order, gfp_mask, ac);
4180 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4182 alloc_flags = reserve_flags;
4185 * Reset the nodemask and zonelist iterators if memory policies can be
4186 * ignored. These allocations are high priority and system rather than
4189 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4190 ac->nodemask = NULL;
4191 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4192 ac->high_zoneidx, ac->nodemask);
4195 /* Attempt with potentially adjusted zonelist and alloc_flags */
4196 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4200 /* Caller is not willing to reclaim, we can't balance anything */
4201 if (!can_direct_reclaim)
4204 /* Avoid recursion of direct reclaim */
4205 if (current->flags & PF_MEMALLOC)
4208 /* Try direct reclaim and then allocating */
4209 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4210 &did_some_progress);
4214 /* Try direct compaction and then allocating */
4215 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4216 compact_priority, &compact_result);
4220 /* Do not loop if specifically requested */
4221 if (gfp_mask & __GFP_NORETRY)
4225 * Do not retry costly high order allocations unless they are
4226 * __GFP_RETRY_MAYFAIL
4228 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4231 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4232 did_some_progress > 0, &no_progress_loops))
4236 * It doesn't make any sense to retry for the compaction if the order-0
4237 * reclaim is not able to make any progress because the current
4238 * implementation of the compaction depends on the sufficient amount
4239 * of free memory (see __compaction_suitable)
4241 if (did_some_progress > 0 &&
4242 should_compact_retry(ac, order, alloc_flags,
4243 compact_result, &compact_priority,
4244 &compaction_retries))
4248 /* Deal with possible cpuset update races before we start OOM killing */
4249 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4252 /* Reclaim has failed us, start killing things */
4253 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4257 /* Avoid allocations with no watermarks from looping endlessly */
4258 if (tsk_is_oom_victim(current) &&
4259 (alloc_flags == ALLOC_OOM ||
4260 (gfp_mask & __GFP_NOMEMALLOC)))
4263 /* Retry as long as the OOM killer is making progress */
4264 if (did_some_progress) {
4265 no_progress_loops = 0;
4270 /* Deal with possible cpuset update races before we fail */
4271 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4275 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4278 if (gfp_mask & __GFP_NOFAIL) {
4280 * All existing users of the __GFP_NOFAIL are blockable, so warn
4281 * of any new users that actually require GFP_NOWAIT
4283 if (WARN_ON_ONCE(!can_direct_reclaim))
4287 * PF_MEMALLOC request from this context is rather bizarre
4288 * because we cannot reclaim anything and only can loop waiting
4289 * for somebody to do a work for us
4291 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4294 * non failing costly orders are a hard requirement which we
4295 * are not prepared for much so let's warn about these users
4296 * so that we can identify them and convert them to something
4299 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4302 * Help non-failing allocations by giving them access to memory
4303 * reserves but do not use ALLOC_NO_WATERMARKS because this
4304 * could deplete whole memory reserves which would just make
4305 * the situation worse
4307 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4315 warn_alloc(gfp_mask, ac->nodemask,
4316 "page allocation failure: order:%u", order);
4321 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4322 int preferred_nid, nodemask_t *nodemask,
4323 struct alloc_context *ac, gfp_t *alloc_mask,
4324 unsigned int *alloc_flags)
4326 ac->high_zoneidx = gfp_zone(gfp_mask);
4327 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4328 ac->nodemask = nodemask;
4329 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4331 if (cpusets_enabled()) {
4332 *alloc_mask |= __GFP_HARDWALL;
4334 ac->nodemask = &cpuset_current_mems_allowed;
4336 *alloc_flags |= ALLOC_CPUSET;
4339 fs_reclaim_acquire(gfp_mask);
4340 fs_reclaim_release(gfp_mask);
4342 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4344 if (should_fail_alloc_page(gfp_mask, order))
4347 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4348 *alloc_flags |= ALLOC_CMA;
4353 /* Determine whether to spread dirty pages and what the first usable zone */
4354 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4356 /* Dirty zone balancing only done in the fast path */
4357 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4360 * The preferred zone is used for statistics but crucially it is
4361 * also used as the starting point for the zonelist iterator. It
4362 * may get reset for allocations that ignore memory policies.
4364 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4365 ac->high_zoneidx, ac->nodemask);
4369 * This is the 'heart' of the zoned buddy allocator.
4372 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4373 nodemask_t *nodemask)
4376 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4377 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4378 struct alloc_context ac = { };
4381 * There are several places where we assume that the order value is sane
4382 * so bail out early if the request is out of bound.
4384 if (unlikely(order >= MAX_ORDER)) {
4385 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4389 gfp_mask &= gfp_allowed_mask;
4390 alloc_mask = gfp_mask;
4391 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4394 finalise_ac(gfp_mask, &ac);
4396 /* First allocation attempt */
4397 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4402 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4403 * resp. GFP_NOIO which has to be inherited for all allocation requests
4404 * from a particular context which has been marked by
4405 * memalloc_no{fs,io}_{save,restore}.
4407 alloc_mask = current_gfp_context(gfp_mask);
4408 ac.spread_dirty_pages = false;
4411 * Restore the original nodemask if it was potentially replaced with
4412 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4414 if (unlikely(ac.nodemask != nodemask))
4415 ac.nodemask = nodemask;
4417 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4420 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4421 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4422 __free_pages(page, order);
4426 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4430 EXPORT_SYMBOL(__alloc_pages_nodemask);
4433 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4434 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4435 * you need to access high mem.
4437 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4441 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4444 return (unsigned long) page_address(page);
4446 EXPORT_SYMBOL(__get_free_pages);
4448 unsigned long get_zeroed_page(gfp_t gfp_mask)
4450 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4452 EXPORT_SYMBOL(get_zeroed_page);
4454 static inline void free_the_page(struct page *page, unsigned int order)
4456 if (order == 0) /* Via pcp? */
4457 free_unref_page(page);
4459 __free_pages_ok(page, order);
4462 void __free_pages(struct page *page, unsigned int order)
4464 if (put_page_testzero(page))
4465 free_the_page(page, order);
4467 EXPORT_SYMBOL(__free_pages);
4469 void free_pages(unsigned long addr, unsigned int order)
4472 VM_BUG_ON(!virt_addr_valid((void *)addr));
4473 __free_pages(virt_to_page((void *)addr), order);
4477 EXPORT_SYMBOL(free_pages);
4481 * An arbitrary-length arbitrary-offset area of memory which resides
4482 * within a 0 or higher order page. Multiple fragments within that page
4483 * are individually refcounted, in the page's reference counter.
4485 * The page_frag functions below provide a simple allocation framework for
4486 * page fragments. This is used by the network stack and network device
4487 * drivers to provide a backing region of memory for use as either an
4488 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4490 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4493 struct page *page = NULL;
4494 gfp_t gfp = gfp_mask;
4496 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4497 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4499 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4500 PAGE_FRAG_CACHE_MAX_ORDER);
4501 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4503 if (unlikely(!page))
4504 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4506 nc->va = page ? page_address(page) : NULL;
4511 void __page_frag_cache_drain(struct page *page, unsigned int count)
4513 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4515 if (page_ref_sub_and_test(page, count))
4516 free_the_page(page, compound_order(page));
4518 EXPORT_SYMBOL(__page_frag_cache_drain);
4520 void *page_frag_alloc(struct page_frag_cache *nc,
4521 unsigned int fragsz, gfp_t gfp_mask)
4523 unsigned int size = PAGE_SIZE;
4527 if (unlikely(!nc->va)) {
4529 page = __page_frag_cache_refill(nc, gfp_mask);
4533 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4534 /* if size can vary use size else just use PAGE_SIZE */
4537 /* Even if we own the page, we do not use atomic_set().
4538 * This would break get_page_unless_zero() users.
4540 page_ref_add(page, size);
4542 /* reset page count bias and offset to start of new frag */
4543 nc->pfmemalloc = page_is_pfmemalloc(page);
4544 nc->pagecnt_bias = size + 1;
4548 offset = nc->offset - fragsz;
4549 if (unlikely(offset < 0)) {
4550 page = virt_to_page(nc->va);
4552 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4555 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4556 /* if size can vary use size else just use PAGE_SIZE */
4559 /* OK, page count is 0, we can safely set it */
4560 set_page_count(page, size + 1);
4562 /* reset page count bias and offset to start of new frag */
4563 nc->pagecnt_bias = size + 1;
4564 offset = size - fragsz;
4568 nc->offset = offset;
4570 return nc->va + offset;
4572 EXPORT_SYMBOL(page_frag_alloc);
4575 * Frees a page fragment allocated out of either a compound or order 0 page.
4577 void page_frag_free(void *addr)
4579 struct page *page = virt_to_head_page(addr);
4581 if (unlikely(put_page_testzero(page)))
4582 free_the_page(page, compound_order(page));
4584 EXPORT_SYMBOL(page_frag_free);
4586 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4590 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4591 unsigned long used = addr + PAGE_ALIGN(size);
4593 split_page(virt_to_page((void *)addr), order);
4594 while (used < alloc_end) {
4599 return (void *)addr;
4603 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4604 * @size: the number of bytes to allocate
4605 * @gfp_mask: GFP flags for the allocation
4607 * This function is similar to alloc_pages(), except that it allocates the
4608 * minimum number of pages to satisfy the request. alloc_pages() can only
4609 * allocate memory in power-of-two pages.
4611 * This function is also limited by MAX_ORDER.
4613 * Memory allocated by this function must be released by free_pages_exact().
4615 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4617 unsigned int order = get_order(size);
4620 addr = __get_free_pages(gfp_mask, order);
4621 return make_alloc_exact(addr, order, size);
4623 EXPORT_SYMBOL(alloc_pages_exact);
4626 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4628 * @nid: the preferred node ID where memory should be allocated
4629 * @size: the number of bytes to allocate
4630 * @gfp_mask: GFP flags for the allocation
4632 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4635 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4637 unsigned int order = get_order(size);
4638 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4641 return make_alloc_exact((unsigned long)page_address(p), order, size);
4645 * free_pages_exact - release memory allocated via alloc_pages_exact()
4646 * @virt: the value returned by alloc_pages_exact.
4647 * @size: size of allocation, same value as passed to alloc_pages_exact().
4649 * Release the memory allocated by a previous call to alloc_pages_exact.
4651 void free_pages_exact(void *virt, size_t size)
4653 unsigned long addr = (unsigned long)virt;
4654 unsigned long end = addr + PAGE_ALIGN(size);
4656 while (addr < end) {
4661 EXPORT_SYMBOL(free_pages_exact);
4664 * nr_free_zone_pages - count number of pages beyond high watermark
4665 * @offset: The zone index of the highest zone
4667 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4668 * high watermark within all zones at or below a given zone index. For each
4669 * zone, the number of pages is calculated as:
4671 * nr_free_zone_pages = managed_pages - high_pages
4673 static unsigned long nr_free_zone_pages(int offset)
4678 /* Just pick one node, since fallback list is circular */
4679 unsigned long sum = 0;
4681 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4683 for_each_zone_zonelist(zone, z, zonelist, offset) {
4684 unsigned long size = zone->managed_pages;
4685 unsigned long high = high_wmark_pages(zone);
4694 * nr_free_buffer_pages - count number of pages beyond high watermark
4696 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4697 * watermark within ZONE_DMA and ZONE_NORMAL.
4699 unsigned long nr_free_buffer_pages(void)
4701 return nr_free_zone_pages(gfp_zone(GFP_USER));
4703 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4706 * nr_free_pagecache_pages - count number of pages beyond high watermark
4708 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4709 * high watermark within all zones.
4711 unsigned long nr_free_pagecache_pages(void)
4713 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4716 static inline void show_node(struct zone *zone)
4718 if (IS_ENABLED(CONFIG_NUMA))
4719 printk("Node %d ", zone_to_nid(zone));
4722 long si_mem_available(void)
4725 unsigned long pagecache;
4726 unsigned long wmark_low = 0;
4727 unsigned long pages[NR_LRU_LISTS];
4731 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4732 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4735 wmark_low += zone->watermark[WMARK_LOW];
4738 * Estimate the amount of memory available for userspace allocations,
4739 * without causing swapping.
4741 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4744 * Not all the page cache can be freed, otherwise the system will
4745 * start swapping. Assume at least half of the page cache, or the
4746 * low watermark worth of cache, needs to stay.
4748 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4749 pagecache -= min(pagecache / 2, wmark_low);
4750 available += pagecache;
4753 * Part of the reclaimable slab consists of items that are in use,
4754 * and cannot be freed. Cap this estimate at the low watermark.
4756 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4757 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4761 * Part of the kernel memory, which can be released under memory
4764 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4771 EXPORT_SYMBOL_GPL(si_mem_available);
4773 void si_meminfo(struct sysinfo *val)
4775 val->totalram = totalram_pages;
4776 val->sharedram = global_node_page_state(NR_SHMEM);
4777 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4778 val->bufferram = nr_blockdev_pages();
4779 val->totalhigh = totalhigh_pages;
4780 val->freehigh = nr_free_highpages();
4781 val->mem_unit = PAGE_SIZE;
4784 EXPORT_SYMBOL(si_meminfo);
4787 void si_meminfo_node(struct sysinfo *val, int nid)
4789 int zone_type; /* needs to be signed */
4790 unsigned long managed_pages = 0;
4791 unsigned long managed_highpages = 0;
4792 unsigned long free_highpages = 0;
4793 pg_data_t *pgdat = NODE_DATA(nid);
4795 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4796 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4797 val->totalram = managed_pages;
4798 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4799 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4800 #ifdef CONFIG_HIGHMEM
4801 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4802 struct zone *zone = &pgdat->node_zones[zone_type];
4804 if (is_highmem(zone)) {
4805 managed_highpages += zone->managed_pages;
4806 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4809 val->totalhigh = managed_highpages;
4810 val->freehigh = free_highpages;
4812 val->totalhigh = managed_highpages;
4813 val->freehigh = free_highpages;
4815 val->mem_unit = PAGE_SIZE;
4820 * Determine whether the node should be displayed or not, depending on whether
4821 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4823 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4825 if (!(flags & SHOW_MEM_FILTER_NODES))
4829 * no node mask - aka implicit memory numa policy. Do not bother with
4830 * the synchronization - read_mems_allowed_begin - because we do not
4831 * have to be precise here.
4834 nodemask = &cpuset_current_mems_allowed;
4836 return !node_isset(nid, *nodemask);
4839 #define K(x) ((x) << (PAGE_SHIFT-10))
4841 static void show_migration_types(unsigned char type)
4843 static const char types[MIGRATE_TYPES] = {
4844 [MIGRATE_UNMOVABLE] = 'U',
4845 [MIGRATE_MOVABLE] = 'M',
4846 [MIGRATE_RECLAIMABLE] = 'E',
4847 [MIGRATE_HIGHATOMIC] = 'H',
4849 [MIGRATE_CMA] = 'C',
4851 #ifdef CONFIG_MEMORY_ISOLATION
4852 [MIGRATE_ISOLATE] = 'I',
4855 char tmp[MIGRATE_TYPES + 1];
4859 for (i = 0; i < MIGRATE_TYPES; i++) {
4860 if (type & (1 << i))
4865 printk(KERN_CONT "(%s) ", tmp);
4869 * Show free area list (used inside shift_scroll-lock stuff)
4870 * We also calculate the percentage fragmentation. We do this by counting the
4871 * memory on each free list with the exception of the first item on the list.
4874 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4877 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4879 unsigned long free_pcp = 0;
4884 for_each_populated_zone(zone) {
4885 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4888 for_each_online_cpu(cpu)
4889 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4892 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4893 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4894 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4895 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4896 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4897 " free:%lu free_pcp:%lu free_cma:%lu\n",
4898 global_node_page_state(NR_ACTIVE_ANON),
4899 global_node_page_state(NR_INACTIVE_ANON),
4900 global_node_page_state(NR_ISOLATED_ANON),
4901 global_node_page_state(NR_ACTIVE_FILE),
4902 global_node_page_state(NR_INACTIVE_FILE),
4903 global_node_page_state(NR_ISOLATED_FILE),
4904 global_node_page_state(NR_UNEVICTABLE),
4905 global_node_page_state(NR_FILE_DIRTY),
4906 global_node_page_state(NR_WRITEBACK),
4907 global_node_page_state(NR_UNSTABLE_NFS),
4908 global_node_page_state(NR_SLAB_RECLAIMABLE),
4909 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4910 global_node_page_state(NR_FILE_MAPPED),
4911 global_node_page_state(NR_SHMEM),
4912 global_zone_page_state(NR_PAGETABLE),
4913 global_zone_page_state(NR_BOUNCE),
4914 global_zone_page_state(NR_FREE_PAGES),
4916 global_zone_page_state(NR_FREE_CMA_PAGES));
4918 for_each_online_pgdat(pgdat) {
4919 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4923 " active_anon:%lukB"
4924 " inactive_anon:%lukB"
4925 " active_file:%lukB"
4926 " inactive_file:%lukB"
4927 " unevictable:%lukB"
4928 " isolated(anon):%lukB"
4929 " isolated(file):%lukB"
4934 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4936 " shmem_pmdmapped: %lukB"
4939 " writeback_tmp:%lukB"
4941 " all_unreclaimable? %s"
4944 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4945 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4946 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4947 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4948 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4949 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4950 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4951 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4952 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4953 K(node_page_state(pgdat, NR_WRITEBACK)),
4954 K(node_page_state(pgdat, NR_SHMEM)),
4955 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4956 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4957 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4959 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4961 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4962 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4963 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4967 for_each_populated_zone(zone) {
4970 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4974 for_each_online_cpu(cpu)
4975 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4984 " active_anon:%lukB"
4985 " inactive_anon:%lukB"
4986 " active_file:%lukB"
4987 " inactive_file:%lukB"
4988 " unevictable:%lukB"
4989 " writepending:%lukB"
4993 " kernel_stack:%lukB"
5001 K(zone_page_state(zone, NR_FREE_PAGES)),
5002 K(min_wmark_pages(zone)),
5003 K(low_wmark_pages(zone)),
5004 K(high_wmark_pages(zone)),
5005 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5006 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5007 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5008 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5009 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5010 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5011 K(zone->present_pages),
5012 K(zone->managed_pages),
5013 K(zone_page_state(zone, NR_MLOCK)),
5014 zone_page_state(zone, NR_KERNEL_STACK_KB),
5015 K(zone_page_state(zone, NR_PAGETABLE)),
5016 K(zone_page_state(zone, NR_BOUNCE)),
5018 K(this_cpu_read(zone->pageset->pcp.count)),
5019 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5020 printk("lowmem_reserve[]:");
5021 for (i = 0; i < MAX_NR_ZONES; i++)
5022 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5023 printk(KERN_CONT "\n");
5026 for_each_populated_zone(zone) {
5028 unsigned long nr[MAX_ORDER], flags, total = 0;
5029 unsigned char types[MAX_ORDER];
5031 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5034 printk(KERN_CONT "%s: ", zone->name);
5036 spin_lock_irqsave(&zone->lock, flags);
5037 for (order = 0; order < MAX_ORDER; order++) {
5038 struct free_area *area = &zone->free_area[order];
5041 nr[order] = area->nr_free;
5042 total += nr[order] << order;
5045 for (type = 0; type < MIGRATE_TYPES; type++) {
5046 if (!list_empty(&area->free_list[type]))
5047 types[order] |= 1 << type;
5050 spin_unlock_irqrestore(&zone->lock, flags);
5051 for (order = 0; order < MAX_ORDER; order++) {
5052 printk(KERN_CONT "%lu*%lukB ",
5053 nr[order], K(1UL) << order);
5055 show_migration_types(types[order]);
5057 printk(KERN_CONT "= %lukB\n", K(total));
5060 hugetlb_show_meminfo();
5062 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5064 show_swap_cache_info();
5067 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5069 zoneref->zone = zone;
5070 zoneref->zone_idx = zone_idx(zone);
5074 * Builds allocation fallback zone lists.
5076 * Add all populated zones of a node to the zonelist.
5078 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5081 enum zone_type zone_type = MAX_NR_ZONES;
5086 zone = pgdat->node_zones + zone_type;
5087 if (managed_zone(zone)) {
5088 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5089 check_highest_zone(zone_type);
5091 } while (zone_type);
5098 static int __parse_numa_zonelist_order(char *s)
5101 * We used to support different zonlists modes but they turned
5102 * out to be just not useful. Let's keep the warning in place
5103 * if somebody still use the cmd line parameter so that we do
5104 * not fail it silently
5106 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5107 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5113 static __init int setup_numa_zonelist_order(char *s)
5118 return __parse_numa_zonelist_order(s);
5120 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5122 char numa_zonelist_order[] = "Node";
5125 * sysctl handler for numa_zonelist_order
5127 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5128 void __user *buffer, size_t *length,
5135 return proc_dostring(table, write, buffer, length, ppos);
5136 str = memdup_user_nul(buffer, 16);
5138 return PTR_ERR(str);
5140 ret = __parse_numa_zonelist_order(str);
5146 #define MAX_NODE_LOAD (nr_online_nodes)
5147 static int node_load[MAX_NUMNODES];
5150 * find_next_best_node - find the next node that should appear in a given node's fallback list
5151 * @node: node whose fallback list we're appending
5152 * @used_node_mask: nodemask_t of already used nodes
5154 * We use a number of factors to determine which is the next node that should
5155 * appear on a given node's fallback list. The node should not have appeared
5156 * already in @node's fallback list, and it should be the next closest node
5157 * according to the distance array (which contains arbitrary distance values
5158 * from each node to each node in the system), and should also prefer nodes
5159 * with no CPUs, since presumably they'll have very little allocation pressure
5160 * on them otherwise.
5161 * It returns -1 if no node is found.
5163 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5166 int min_val = INT_MAX;
5167 int best_node = NUMA_NO_NODE;
5168 const struct cpumask *tmp = cpumask_of_node(0);
5170 /* Use the local node if we haven't already */
5171 if (!node_isset(node, *used_node_mask)) {
5172 node_set(node, *used_node_mask);
5176 for_each_node_state(n, N_MEMORY) {
5178 /* Don't want a node to appear more than once */
5179 if (node_isset(n, *used_node_mask))
5182 /* Use the distance array to find the distance */
5183 val = node_distance(node, n);
5185 /* Penalize nodes under us ("prefer the next node") */
5188 /* Give preference to headless and unused nodes */
5189 tmp = cpumask_of_node(n);
5190 if (!cpumask_empty(tmp))
5191 val += PENALTY_FOR_NODE_WITH_CPUS;
5193 /* Slight preference for less loaded node */
5194 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5195 val += node_load[n];
5197 if (val < min_val) {
5204 node_set(best_node, *used_node_mask);
5211 * Build zonelists ordered by node and zones within node.
5212 * This results in maximum locality--normal zone overflows into local
5213 * DMA zone, if any--but risks exhausting DMA zone.
5215 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5218 struct zoneref *zonerefs;
5221 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5223 for (i = 0; i < nr_nodes; i++) {
5226 pg_data_t *node = NODE_DATA(node_order[i]);
5228 nr_zones = build_zonerefs_node(node, zonerefs);
5229 zonerefs += nr_zones;
5231 zonerefs->zone = NULL;
5232 zonerefs->zone_idx = 0;
5236 * Build gfp_thisnode zonelists
5238 static void build_thisnode_zonelists(pg_data_t *pgdat)
5240 struct zoneref *zonerefs;
5243 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5244 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5245 zonerefs += nr_zones;
5246 zonerefs->zone = NULL;
5247 zonerefs->zone_idx = 0;
5251 * Build zonelists ordered by zone and nodes within zones.
5252 * This results in conserving DMA zone[s] until all Normal memory is
5253 * exhausted, but results in overflowing to remote node while memory
5254 * may still exist in local DMA zone.
5257 static void build_zonelists(pg_data_t *pgdat)
5259 static int node_order[MAX_NUMNODES];
5260 int node, load, nr_nodes = 0;
5261 nodemask_t used_mask;
5262 int local_node, prev_node;
5264 /* NUMA-aware ordering of nodes */
5265 local_node = pgdat->node_id;
5266 load = nr_online_nodes;
5267 prev_node = local_node;
5268 nodes_clear(used_mask);
5270 memset(node_order, 0, sizeof(node_order));
5271 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5273 * We don't want to pressure a particular node.
5274 * So adding penalty to the first node in same
5275 * distance group to make it round-robin.
5277 if (node_distance(local_node, node) !=
5278 node_distance(local_node, prev_node))
5279 node_load[node] = load;
5281 node_order[nr_nodes++] = node;
5286 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5287 build_thisnode_zonelists(pgdat);
5290 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5292 * Return node id of node used for "local" allocations.
5293 * I.e., first node id of first zone in arg node's generic zonelist.
5294 * Used for initializing percpu 'numa_mem', which is used primarily
5295 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5297 int local_memory_node(int node)
5301 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5302 gfp_zone(GFP_KERNEL),
5304 return zone_to_nid(z->zone);
5308 static void setup_min_unmapped_ratio(void);
5309 static void setup_min_slab_ratio(void);
5310 #else /* CONFIG_NUMA */
5312 static void build_zonelists(pg_data_t *pgdat)
5314 int node, local_node;
5315 struct zoneref *zonerefs;
5318 local_node = pgdat->node_id;
5320 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5321 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5322 zonerefs += nr_zones;
5325 * Now we build the zonelist so that it contains the zones
5326 * of all the other nodes.
5327 * We don't want to pressure a particular node, so when
5328 * building the zones for node N, we make sure that the
5329 * zones coming right after the local ones are those from
5330 * node N+1 (modulo N)
5332 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5333 if (!node_online(node))
5335 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5336 zonerefs += nr_zones;
5338 for (node = 0; node < local_node; node++) {
5339 if (!node_online(node))
5341 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5342 zonerefs += nr_zones;
5345 zonerefs->zone = NULL;
5346 zonerefs->zone_idx = 0;
5349 #endif /* CONFIG_NUMA */
5352 * Boot pageset table. One per cpu which is going to be used for all
5353 * zones and all nodes. The parameters will be set in such a way
5354 * that an item put on a list will immediately be handed over to
5355 * the buddy list. This is safe since pageset manipulation is done
5356 * with interrupts disabled.
5358 * The boot_pagesets must be kept even after bootup is complete for
5359 * unused processors and/or zones. They do play a role for bootstrapping
5360 * hotplugged processors.
5362 * zoneinfo_show() and maybe other functions do
5363 * not check if the processor is online before following the pageset pointer.
5364 * Other parts of the kernel may not check if the zone is available.
5366 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5367 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5368 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5370 static void __build_all_zonelists(void *data)
5373 int __maybe_unused cpu;
5374 pg_data_t *self = data;
5375 static DEFINE_SPINLOCK(lock);
5380 memset(node_load, 0, sizeof(node_load));
5384 * This node is hotadded and no memory is yet present. So just
5385 * building zonelists is fine - no need to touch other nodes.
5387 if (self && !node_online(self->node_id)) {
5388 build_zonelists(self);
5390 for_each_online_node(nid) {
5391 pg_data_t *pgdat = NODE_DATA(nid);
5393 build_zonelists(pgdat);
5396 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5398 * We now know the "local memory node" for each node--
5399 * i.e., the node of the first zone in the generic zonelist.
5400 * Set up numa_mem percpu variable for on-line cpus. During
5401 * boot, only the boot cpu should be on-line; we'll init the
5402 * secondary cpus' numa_mem as they come on-line. During
5403 * node/memory hotplug, we'll fixup all on-line cpus.
5405 for_each_online_cpu(cpu)
5406 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5413 static noinline void __init
5414 build_all_zonelists_init(void)
5418 __build_all_zonelists(NULL);
5421 * Initialize the boot_pagesets that are going to be used
5422 * for bootstrapping processors. The real pagesets for
5423 * each zone will be allocated later when the per cpu
5424 * allocator is available.
5426 * boot_pagesets are used also for bootstrapping offline
5427 * cpus if the system is already booted because the pagesets
5428 * are needed to initialize allocators on a specific cpu too.
5429 * F.e. the percpu allocator needs the page allocator which
5430 * needs the percpu allocator in order to allocate its pagesets
5431 * (a chicken-egg dilemma).
5433 for_each_possible_cpu(cpu)
5434 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5436 mminit_verify_zonelist();
5437 cpuset_init_current_mems_allowed();
5441 * unless system_state == SYSTEM_BOOTING.
5443 * __ref due to call of __init annotated helper build_all_zonelists_init
5444 * [protected by SYSTEM_BOOTING].
5446 void __ref build_all_zonelists(pg_data_t *pgdat)
5448 if (system_state == SYSTEM_BOOTING) {
5449 build_all_zonelists_init();
5451 __build_all_zonelists(pgdat);
5452 /* cpuset refresh routine should be here */
5454 vm_total_pages = nr_free_pagecache_pages();
5456 * Disable grouping by mobility if the number of pages in the
5457 * system is too low to allow the mechanism to work. It would be
5458 * more accurate, but expensive to check per-zone. This check is
5459 * made on memory-hotadd so a system can start with mobility
5460 * disabled and enable it later
5462 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5463 page_group_by_mobility_disabled = 1;
5465 page_group_by_mobility_disabled = 0;
5467 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5469 page_group_by_mobility_disabled ? "off" : "on",
5472 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5477 * Initially all pages are reserved - free ones are freed
5478 * up by free_all_bootmem() once the early boot process is
5479 * done. Non-atomic initialization, single-pass.
5481 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5482 unsigned long start_pfn, enum memmap_context context,
5483 struct vmem_altmap *altmap)
5485 unsigned long end_pfn = start_pfn + size;
5486 pg_data_t *pgdat = NODE_DATA(nid);
5488 unsigned long nr_initialised = 0;
5490 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5491 struct memblock_region *r = NULL, *tmp;
5494 if (highest_memmap_pfn < end_pfn - 1)
5495 highest_memmap_pfn = end_pfn - 1;
5498 * Honor reservation requested by the driver for this ZONE_DEVICE
5501 if (altmap && start_pfn == altmap->base_pfn)
5502 start_pfn += altmap->reserve;
5504 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5506 * There can be holes in boot-time mem_map[]s handed to this
5507 * function. They do not exist on hotplugged memory.
5509 if (context != MEMMAP_EARLY)
5512 if (!early_pfn_valid(pfn))
5514 if (!early_pfn_in_nid(pfn, nid))
5516 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5519 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5521 * Check given memblock attribute by firmware which can affect
5522 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5523 * mirrored, it's an overlapped memmap init. skip it.
5525 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5526 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5527 for_each_memblock(memory, tmp)
5528 if (pfn < memblock_region_memory_end_pfn(tmp))
5532 if (pfn >= memblock_region_memory_base_pfn(r) &&
5533 memblock_is_mirror(r)) {
5534 /* already initialized as NORMAL */
5535 pfn = memblock_region_memory_end_pfn(r);
5542 page = pfn_to_page(pfn);
5543 __init_single_page(page, pfn, zone, nid);
5544 if (context == MEMMAP_HOTPLUG)
5545 SetPageReserved(page);
5548 * Mark the block movable so that blocks are reserved for
5549 * movable at startup. This will force kernel allocations
5550 * to reserve their blocks rather than leaking throughout
5551 * the address space during boot when many long-lived
5552 * kernel allocations are made.
5554 * bitmap is created for zone's valid pfn range. but memmap
5555 * can be created for invalid pages (for alignment)
5556 * check here not to call set_pageblock_migratetype() against
5559 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5560 * because this is done early in sparse_add_one_section
5562 if (!(pfn & (pageblock_nr_pages - 1))) {
5563 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5569 static void __meminit zone_init_free_lists(struct zone *zone)
5571 unsigned int order, t;
5572 for_each_migratetype_order(order, t) {
5573 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5574 zone->free_area[order].nr_free = 0;
5578 #ifndef __HAVE_ARCH_MEMMAP_INIT
5579 #define memmap_init(size, nid, zone, start_pfn) \
5580 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5583 static int zone_batchsize(struct zone *zone)
5589 * The per-cpu-pages pools are set to around 1000th of the
5592 batch = zone->managed_pages / 1024;
5593 /* But no more than a meg. */
5594 if (batch * PAGE_SIZE > 1024 * 1024)
5595 batch = (1024 * 1024) / PAGE_SIZE;
5596 batch /= 4; /* We effectively *= 4 below */
5601 * Clamp the batch to a 2^n - 1 value. Having a power
5602 * of 2 value was found to be more likely to have
5603 * suboptimal cache aliasing properties in some cases.
5605 * For example if 2 tasks are alternately allocating
5606 * batches of pages, one task can end up with a lot
5607 * of pages of one half of the possible page colors
5608 * and the other with pages of the other colors.
5610 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5615 /* The deferral and batching of frees should be suppressed under NOMMU
5618 * The problem is that NOMMU needs to be able to allocate large chunks
5619 * of contiguous memory as there's no hardware page translation to
5620 * assemble apparent contiguous memory from discontiguous pages.
5622 * Queueing large contiguous runs of pages for batching, however,
5623 * causes the pages to actually be freed in smaller chunks. As there
5624 * can be a significant delay between the individual batches being
5625 * recycled, this leads to the once large chunks of space being
5626 * fragmented and becoming unavailable for high-order allocations.
5633 * pcp->high and pcp->batch values are related and dependent on one another:
5634 * ->batch must never be higher then ->high.
5635 * The following function updates them in a safe manner without read side
5638 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5639 * those fields changing asynchronously (acording the the above rule).
5641 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5642 * outside of boot time (or some other assurance that no concurrent updaters
5645 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5646 unsigned long batch)
5648 /* start with a fail safe value for batch */
5652 /* Update high, then batch, in order */
5659 /* a companion to pageset_set_high() */
5660 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5662 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5665 static void pageset_init(struct per_cpu_pageset *p)
5667 struct per_cpu_pages *pcp;
5670 memset(p, 0, sizeof(*p));
5674 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5675 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5678 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5681 pageset_set_batch(p, batch);
5685 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5686 * to the value high for the pageset p.
5688 static void pageset_set_high(struct per_cpu_pageset *p,
5691 unsigned long batch = max(1UL, high / 4);
5692 if ((high / 4) > (PAGE_SHIFT * 8))
5693 batch = PAGE_SHIFT * 8;
5695 pageset_update(&p->pcp, high, batch);
5698 static void pageset_set_high_and_batch(struct zone *zone,
5699 struct per_cpu_pageset *pcp)
5701 if (percpu_pagelist_fraction)
5702 pageset_set_high(pcp,
5703 (zone->managed_pages /
5704 percpu_pagelist_fraction));
5706 pageset_set_batch(pcp, zone_batchsize(zone));
5709 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5711 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5714 pageset_set_high_and_batch(zone, pcp);
5717 void __meminit setup_zone_pageset(struct zone *zone)
5720 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5721 for_each_possible_cpu(cpu)
5722 zone_pageset_init(zone, cpu);
5726 * Allocate per cpu pagesets and initialize them.
5727 * Before this call only boot pagesets were available.
5729 void __init setup_per_cpu_pageset(void)
5731 struct pglist_data *pgdat;
5734 for_each_populated_zone(zone)
5735 setup_zone_pageset(zone);
5737 for_each_online_pgdat(pgdat)
5738 pgdat->per_cpu_nodestats =
5739 alloc_percpu(struct per_cpu_nodestat);
5742 static __meminit void zone_pcp_init(struct zone *zone)
5745 * per cpu subsystem is not up at this point. The following code
5746 * relies on the ability of the linker to provide the
5747 * offset of a (static) per cpu variable into the per cpu area.
5749 zone->pageset = &boot_pageset;
5751 if (populated_zone(zone))
5752 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5753 zone->name, zone->present_pages,
5754 zone_batchsize(zone));
5757 void __meminit init_currently_empty_zone(struct zone *zone,
5758 unsigned long zone_start_pfn,
5761 struct pglist_data *pgdat = zone->zone_pgdat;
5762 int zone_idx = zone_idx(zone) + 1;
5764 if (zone_idx > pgdat->nr_zones)
5765 pgdat->nr_zones = zone_idx;
5767 zone->zone_start_pfn = zone_start_pfn;
5769 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5770 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5772 (unsigned long)zone_idx(zone),
5773 zone_start_pfn, (zone_start_pfn + size));
5775 zone_init_free_lists(zone);
5776 zone->initialized = 1;
5779 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5780 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5783 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5785 int __meminit __early_pfn_to_nid(unsigned long pfn,
5786 struct mminit_pfnnid_cache *state)
5788 unsigned long start_pfn, end_pfn;
5791 if (state->last_start <= pfn && pfn < state->last_end)
5792 return state->last_nid;
5794 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5796 state->last_start = start_pfn;
5797 state->last_end = end_pfn;
5798 state->last_nid = nid;
5803 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5806 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5807 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5808 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5810 * If an architecture guarantees that all ranges registered contain no holes
5811 * and may be freed, this this function may be used instead of calling
5812 * memblock_free_early_nid() manually.
5814 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5816 unsigned long start_pfn, end_pfn;
5819 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5820 start_pfn = min(start_pfn, max_low_pfn);
5821 end_pfn = min(end_pfn, max_low_pfn);
5823 if (start_pfn < end_pfn)
5824 memblock_free_early_nid(PFN_PHYS(start_pfn),
5825 (end_pfn - start_pfn) << PAGE_SHIFT,
5831 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5832 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5834 * If an architecture guarantees that all ranges registered contain no holes and may
5835 * be freed, this function may be used instead of calling memory_present() manually.
5837 void __init sparse_memory_present_with_active_regions(int nid)
5839 unsigned long start_pfn, end_pfn;
5842 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5843 memory_present(this_nid, start_pfn, end_pfn);
5847 * get_pfn_range_for_nid - Return the start and end page frames for a node
5848 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5849 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5850 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5852 * It returns the start and end page frame of a node based on information
5853 * provided by memblock_set_node(). If called for a node
5854 * with no available memory, a warning is printed and the start and end
5857 void __meminit get_pfn_range_for_nid(unsigned int nid,
5858 unsigned long *start_pfn, unsigned long *end_pfn)
5860 unsigned long this_start_pfn, this_end_pfn;
5866 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5867 *start_pfn = min(*start_pfn, this_start_pfn);
5868 *end_pfn = max(*end_pfn, this_end_pfn);
5871 if (*start_pfn == -1UL)
5876 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5877 * assumption is made that zones within a node are ordered in monotonic
5878 * increasing memory addresses so that the "highest" populated zone is used
5880 static void __init find_usable_zone_for_movable(void)
5883 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5884 if (zone_index == ZONE_MOVABLE)
5887 if (arch_zone_highest_possible_pfn[zone_index] >
5888 arch_zone_lowest_possible_pfn[zone_index])
5892 VM_BUG_ON(zone_index == -1);
5893 movable_zone = zone_index;
5897 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5898 * because it is sized independent of architecture. Unlike the other zones,
5899 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5900 * in each node depending on the size of each node and how evenly kernelcore
5901 * is distributed. This helper function adjusts the zone ranges
5902 * provided by the architecture for a given node by using the end of the
5903 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5904 * zones within a node are in order of monotonic increases memory addresses
5906 static void __meminit adjust_zone_range_for_zone_movable(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)
5913 /* Only adjust if ZONE_MOVABLE is on this node */
5914 if (zone_movable_pfn[nid]) {
5915 /* Size ZONE_MOVABLE */
5916 if (zone_type == ZONE_MOVABLE) {
5917 *zone_start_pfn = zone_movable_pfn[nid];
5918 *zone_end_pfn = min(node_end_pfn,
5919 arch_zone_highest_possible_pfn[movable_zone]);
5921 /* Adjust for ZONE_MOVABLE starting within this range */
5922 } else if (!mirrored_kernelcore &&
5923 *zone_start_pfn < zone_movable_pfn[nid] &&
5924 *zone_end_pfn > zone_movable_pfn[nid]) {
5925 *zone_end_pfn = zone_movable_pfn[nid];
5927 /* Check if this whole range is within ZONE_MOVABLE */
5928 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5929 *zone_start_pfn = *zone_end_pfn;
5934 * Return the number of pages a zone spans in a node, including holes
5935 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5937 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5938 unsigned long zone_type,
5939 unsigned long node_start_pfn,
5940 unsigned long node_end_pfn,
5941 unsigned long *zone_start_pfn,
5942 unsigned long *zone_end_pfn,
5943 unsigned long *ignored)
5945 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5946 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5947 /* When hotadd a new node from cpu_up(), the node should be empty */
5948 if (!node_start_pfn && !node_end_pfn)
5951 /* Get the start and end of the zone */
5952 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5953 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5954 adjust_zone_range_for_zone_movable(nid, zone_type,
5955 node_start_pfn, node_end_pfn,
5956 zone_start_pfn, zone_end_pfn);
5958 /* Check that this node has pages within the zone's required range */
5959 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5962 /* Move the zone boundaries inside the node if necessary */
5963 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5964 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5966 /* Return the spanned pages */
5967 return *zone_end_pfn - *zone_start_pfn;
5971 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5972 * then all holes in the requested range will be accounted for.
5974 unsigned long __meminit __absent_pages_in_range(int nid,
5975 unsigned long range_start_pfn,
5976 unsigned long range_end_pfn)
5978 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5979 unsigned long start_pfn, end_pfn;
5982 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5983 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5984 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5985 nr_absent -= end_pfn - start_pfn;
5991 * absent_pages_in_range - Return number of page frames in holes within a range
5992 * @start_pfn: The start PFN to start searching for holes
5993 * @end_pfn: The end PFN to stop searching for holes
5995 * It returns the number of pages frames in memory holes within a range.
5997 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5998 unsigned long end_pfn)
6000 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6003 /* Return the number of page frames in holes in a zone on a node */
6004 static unsigned long __meminit zone_absent_pages_in_node(int nid,
6005 unsigned long zone_type,
6006 unsigned long node_start_pfn,
6007 unsigned long node_end_pfn,
6008 unsigned long *ignored)
6010 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6011 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6012 unsigned long zone_start_pfn, zone_end_pfn;
6013 unsigned long nr_absent;
6015 /* When hotadd a new node from cpu_up(), the node should be empty */
6016 if (!node_start_pfn && !node_end_pfn)
6019 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6020 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6022 adjust_zone_range_for_zone_movable(nid, zone_type,
6023 node_start_pfn, node_end_pfn,
6024 &zone_start_pfn, &zone_end_pfn);
6025 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6028 * ZONE_MOVABLE handling.
6029 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6032 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6033 unsigned long start_pfn, end_pfn;
6034 struct memblock_region *r;
6036 for_each_memblock(memory, r) {
6037 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6038 zone_start_pfn, zone_end_pfn);
6039 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6040 zone_start_pfn, zone_end_pfn);
6042 if (zone_type == ZONE_MOVABLE &&
6043 memblock_is_mirror(r))
6044 nr_absent += end_pfn - start_pfn;
6046 if (zone_type == ZONE_NORMAL &&
6047 !memblock_is_mirror(r))
6048 nr_absent += end_pfn - start_pfn;
6055 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6056 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6057 unsigned long zone_type,
6058 unsigned long node_start_pfn,
6059 unsigned long node_end_pfn,
6060 unsigned long *zone_start_pfn,
6061 unsigned long *zone_end_pfn,
6062 unsigned long *zones_size)
6066 *zone_start_pfn = node_start_pfn;
6067 for (zone = 0; zone < zone_type; zone++)
6068 *zone_start_pfn += zones_size[zone];
6070 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6072 return zones_size[zone_type];
6075 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6076 unsigned long zone_type,
6077 unsigned long node_start_pfn,
6078 unsigned long node_end_pfn,
6079 unsigned long *zholes_size)
6084 return zholes_size[zone_type];
6087 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6089 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6090 unsigned long node_start_pfn,
6091 unsigned long node_end_pfn,
6092 unsigned long *zones_size,
6093 unsigned long *zholes_size)
6095 unsigned long realtotalpages = 0, totalpages = 0;
6098 for (i = 0; i < MAX_NR_ZONES; i++) {
6099 struct zone *zone = pgdat->node_zones + i;
6100 unsigned long zone_start_pfn, zone_end_pfn;
6101 unsigned long size, real_size;
6103 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6109 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6110 node_start_pfn, node_end_pfn,
6113 zone->zone_start_pfn = zone_start_pfn;
6115 zone->zone_start_pfn = 0;
6116 zone->spanned_pages = size;
6117 zone->present_pages = real_size;
6120 realtotalpages += real_size;
6123 pgdat->node_spanned_pages = totalpages;
6124 pgdat->node_present_pages = realtotalpages;
6125 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6129 #ifndef CONFIG_SPARSEMEM
6131 * Calculate the size of the zone->blockflags rounded to an unsigned long
6132 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6133 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6134 * round what is now in bits to nearest long in bits, then return it in
6137 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6139 unsigned long usemapsize;
6141 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6142 usemapsize = roundup(zonesize, pageblock_nr_pages);
6143 usemapsize = usemapsize >> pageblock_order;
6144 usemapsize *= NR_PAGEBLOCK_BITS;
6145 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6147 return usemapsize / 8;
6150 static void __ref setup_usemap(struct pglist_data *pgdat,
6152 unsigned long zone_start_pfn,
6153 unsigned long zonesize)
6155 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6156 zone->pageblock_flags = NULL;
6158 zone->pageblock_flags =
6159 memblock_virt_alloc_node_nopanic(usemapsize,
6163 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6164 unsigned long zone_start_pfn, unsigned long zonesize) {}
6165 #endif /* CONFIG_SPARSEMEM */
6167 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6169 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6170 void __init set_pageblock_order(void)
6174 /* Check that pageblock_nr_pages has not already been setup */
6175 if (pageblock_order)
6178 if (HPAGE_SHIFT > PAGE_SHIFT)
6179 order = HUGETLB_PAGE_ORDER;
6181 order = MAX_ORDER - 1;
6184 * Assume the largest contiguous order of interest is a huge page.
6185 * This value may be variable depending on boot parameters on IA64 and
6188 pageblock_order = order;
6190 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6193 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6194 * is unused as pageblock_order is set at compile-time. See
6195 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6198 void __init set_pageblock_order(void)
6202 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6204 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6205 unsigned long present_pages)
6207 unsigned long pages = spanned_pages;
6210 * Provide a more accurate estimation if there are holes within
6211 * the zone and SPARSEMEM is in use. If there are holes within the
6212 * zone, each populated memory region may cost us one or two extra
6213 * memmap pages due to alignment because memmap pages for each
6214 * populated regions may not be naturally aligned on page boundary.
6215 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6217 if (spanned_pages > present_pages + (present_pages >> 4) &&
6218 IS_ENABLED(CONFIG_SPARSEMEM))
6219 pages = present_pages;
6221 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6224 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6225 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6227 spin_lock_init(&pgdat->split_queue_lock);
6228 INIT_LIST_HEAD(&pgdat->split_queue);
6229 pgdat->split_queue_len = 0;
6232 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6235 #ifdef CONFIG_COMPACTION
6236 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6238 init_waitqueue_head(&pgdat->kcompactd_wait);
6241 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6244 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6246 pgdat_resize_init(pgdat);
6248 pgdat_init_split_queue(pgdat);
6249 pgdat_init_kcompactd(pgdat);
6251 init_waitqueue_head(&pgdat->kswapd_wait);
6252 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6254 pgdat_page_ext_init(pgdat);
6255 spin_lock_init(&pgdat->lru_lock);
6256 lruvec_init(node_lruvec(pgdat));
6259 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6260 unsigned long remaining_pages)
6262 zone->managed_pages = remaining_pages;
6263 zone_set_nid(zone, nid);
6264 zone->name = zone_names[idx];
6265 zone->zone_pgdat = NODE_DATA(nid);
6266 spin_lock_init(&zone->lock);
6267 zone_seqlock_init(zone);
6268 zone_pcp_init(zone);
6272 * Set up the zone data structures
6273 * - init pgdat internals
6274 * - init all zones belonging to this node
6276 * NOTE: this function is only called during memory hotplug
6278 #ifdef CONFIG_MEMORY_HOTPLUG
6279 void __ref free_area_init_core_hotplug(int nid)
6282 pg_data_t *pgdat = NODE_DATA(nid);
6284 pgdat_init_internals(pgdat);
6285 for (z = 0; z < MAX_NR_ZONES; z++)
6286 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6291 * Set up the zone data structures:
6292 * - mark all pages reserved
6293 * - mark all memory queues empty
6294 * - clear the memory bitmaps
6296 * NOTE: pgdat should get zeroed by caller.
6297 * NOTE: this function is only called during early init.
6299 static void __init free_area_init_core(struct pglist_data *pgdat)
6302 int nid = pgdat->node_id;
6304 pgdat_init_internals(pgdat);
6305 pgdat->per_cpu_nodestats = &boot_nodestats;
6307 for (j = 0; j < MAX_NR_ZONES; j++) {
6308 struct zone *zone = pgdat->node_zones + j;
6309 unsigned long size, freesize, memmap_pages;
6310 unsigned long zone_start_pfn = zone->zone_start_pfn;
6312 size = zone->spanned_pages;
6313 freesize = zone->present_pages;
6316 * Adjust freesize so that it accounts for how much memory
6317 * is used by this zone for memmap. This affects the watermark
6318 * and per-cpu initialisations
6320 memmap_pages = calc_memmap_size(size, freesize);
6321 if (!is_highmem_idx(j)) {
6322 if (freesize >= memmap_pages) {
6323 freesize -= memmap_pages;
6326 " %s zone: %lu pages used for memmap\n",
6327 zone_names[j], memmap_pages);
6329 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6330 zone_names[j], memmap_pages, freesize);
6333 /* Account for reserved pages */
6334 if (j == 0 && freesize > dma_reserve) {
6335 freesize -= dma_reserve;
6336 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6337 zone_names[0], dma_reserve);
6340 if (!is_highmem_idx(j))
6341 nr_kernel_pages += freesize;
6342 /* Charge for highmem memmap if there are enough kernel pages */
6343 else if (nr_kernel_pages > memmap_pages * 2)
6344 nr_kernel_pages -= memmap_pages;
6345 nr_all_pages += freesize;
6348 * Set an approximate value for lowmem here, it will be adjusted
6349 * when the bootmem allocator frees pages into the buddy system.
6350 * And all highmem pages will be managed by the buddy system.
6352 zone_init_internals(zone, j, nid, freesize);
6357 set_pageblock_order();
6358 setup_usemap(pgdat, zone, zone_start_pfn, size);
6359 init_currently_empty_zone(zone, zone_start_pfn, size);
6360 memmap_init(size, nid, j, zone_start_pfn);
6364 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6365 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6367 unsigned long __maybe_unused start = 0;
6368 unsigned long __maybe_unused offset = 0;
6370 /* Skip empty nodes */
6371 if (!pgdat->node_spanned_pages)
6374 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6375 offset = pgdat->node_start_pfn - start;
6376 /* ia64 gets its own node_mem_map, before this, without bootmem */
6377 if (!pgdat->node_mem_map) {
6378 unsigned long size, end;
6382 * The zone's endpoints aren't required to be MAX_ORDER
6383 * aligned but the node_mem_map endpoints must be in order
6384 * for the buddy allocator to function correctly.
6386 end = pgdat_end_pfn(pgdat);
6387 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6388 size = (end - start) * sizeof(struct page);
6389 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6390 pgdat->node_mem_map = map + offset;
6392 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6393 __func__, pgdat->node_id, (unsigned long)pgdat,
6394 (unsigned long)pgdat->node_mem_map);
6395 #ifndef CONFIG_NEED_MULTIPLE_NODES
6397 * With no DISCONTIG, the global mem_map is just set as node 0's
6399 if (pgdat == NODE_DATA(0)) {
6400 mem_map = NODE_DATA(0)->node_mem_map;
6401 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6402 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6404 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6409 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6410 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6412 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6413 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6416 * We start only with one section of pages, more pages are added as
6417 * needed until the rest of deferred pages are initialized.
6419 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6420 pgdat->node_spanned_pages);
6421 pgdat->first_deferred_pfn = ULONG_MAX;
6424 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6427 void __init free_area_init_node(int nid, unsigned long *zones_size,
6428 unsigned long node_start_pfn,
6429 unsigned long *zholes_size)
6431 pg_data_t *pgdat = NODE_DATA(nid);
6432 unsigned long start_pfn = 0;
6433 unsigned long end_pfn = 0;
6435 /* pg_data_t should be reset to zero when it's allocated */
6436 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6438 pgdat->node_id = nid;
6439 pgdat->node_start_pfn = node_start_pfn;
6440 pgdat->per_cpu_nodestats = NULL;
6441 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6442 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6443 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6444 (u64)start_pfn << PAGE_SHIFT,
6445 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6447 start_pfn = node_start_pfn;
6449 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6450 zones_size, zholes_size);
6452 alloc_node_mem_map(pgdat);
6453 pgdat_set_deferred_range(pgdat);
6455 free_area_init_core(pgdat);
6458 #if defined(CONFIG_HAVE_MEMBLOCK) && !defined(CONFIG_FLAT_NODE_MEM_MAP)
6460 * Only struct pages that are backed by physical memory are zeroed and
6461 * initialized by going through __init_single_page(). But, there are some
6462 * struct pages which are reserved in memblock allocator and their fields
6463 * may be accessed (for example page_to_pfn() on some configuration accesses
6464 * flags). We must explicitly zero those struct pages.
6466 void __init zero_resv_unavail(void)
6468 phys_addr_t start, end;
6473 * Loop through ranges that are reserved, but do not have reported
6474 * physical memory backing.
6477 for_each_resv_unavail_range(i, &start, &end) {
6478 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6479 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6480 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6481 + pageblock_nr_pages - 1;
6484 mm_zero_struct_page(pfn_to_page(pfn));
6490 * Struct pages that do not have backing memory. This could be because
6491 * firmware is using some of this memory, or for some other reasons.
6492 * Once memblock is changed so such behaviour is not allowed: i.e.
6493 * list of "reserved" memory must be a subset of list of "memory", then
6494 * this code can be removed.
6497 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6499 #endif /* CONFIG_HAVE_MEMBLOCK && !CONFIG_FLAT_NODE_MEM_MAP */
6501 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6503 #if MAX_NUMNODES > 1
6505 * Figure out the number of possible node ids.
6507 void __init setup_nr_node_ids(void)
6509 unsigned int highest;
6511 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6512 nr_node_ids = highest + 1;
6517 * node_map_pfn_alignment - determine the maximum internode alignment
6519 * This function should be called after node map is populated and sorted.
6520 * It calculates the maximum power of two alignment which can distinguish
6523 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6524 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6525 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6526 * shifted, 1GiB is enough and this function will indicate so.
6528 * This is used to test whether pfn -> nid mapping of the chosen memory
6529 * model has fine enough granularity to avoid incorrect mapping for the
6530 * populated node map.
6532 * Returns the determined alignment in pfn's. 0 if there is no alignment
6533 * requirement (single node).
6535 unsigned long __init node_map_pfn_alignment(void)
6537 unsigned long accl_mask = 0, last_end = 0;
6538 unsigned long start, end, mask;
6542 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6543 if (!start || last_nid < 0 || last_nid == nid) {
6550 * Start with a mask granular enough to pin-point to the
6551 * start pfn and tick off bits one-by-one until it becomes
6552 * too coarse to separate the current node from the last.
6554 mask = ~((1 << __ffs(start)) - 1);
6555 while (mask && last_end <= (start & (mask << 1)))
6558 /* accumulate all internode masks */
6562 /* convert mask to number of pages */
6563 return ~accl_mask + 1;
6566 /* Find the lowest pfn for a node */
6567 static unsigned long __init find_min_pfn_for_node(int nid)
6569 unsigned long min_pfn = ULONG_MAX;
6570 unsigned long start_pfn;
6573 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6574 min_pfn = min(min_pfn, start_pfn);
6576 if (min_pfn == ULONG_MAX) {
6577 pr_warn("Could not find start_pfn for node %d\n", nid);
6585 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6587 * It returns the minimum PFN based on information provided via
6588 * memblock_set_node().
6590 unsigned long __init find_min_pfn_with_active_regions(void)
6592 return find_min_pfn_for_node(MAX_NUMNODES);
6596 * early_calculate_totalpages()
6597 * Sum pages in active regions for movable zone.
6598 * Populate N_MEMORY for calculating usable_nodes.
6600 static unsigned long __init early_calculate_totalpages(void)
6602 unsigned long totalpages = 0;
6603 unsigned long start_pfn, end_pfn;
6606 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6607 unsigned long pages = end_pfn - start_pfn;
6609 totalpages += pages;
6611 node_set_state(nid, N_MEMORY);
6617 * Find the PFN the Movable zone begins in each node. Kernel memory
6618 * is spread evenly between nodes as long as the nodes have enough
6619 * memory. When they don't, some nodes will have more kernelcore than
6622 static void __init find_zone_movable_pfns_for_nodes(void)
6625 unsigned long usable_startpfn;
6626 unsigned long kernelcore_node, kernelcore_remaining;
6627 /* save the state before borrow the nodemask */
6628 nodemask_t saved_node_state = node_states[N_MEMORY];
6629 unsigned long totalpages = early_calculate_totalpages();
6630 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6631 struct memblock_region *r;
6633 /* Need to find movable_zone earlier when movable_node is specified. */
6634 find_usable_zone_for_movable();
6637 * If movable_node is specified, ignore kernelcore and movablecore
6640 if (movable_node_is_enabled()) {
6641 for_each_memblock(memory, r) {
6642 if (!memblock_is_hotpluggable(r))
6647 usable_startpfn = PFN_DOWN(r->base);
6648 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6649 min(usable_startpfn, zone_movable_pfn[nid]) :
6657 * If kernelcore=mirror is specified, ignore movablecore option
6659 if (mirrored_kernelcore) {
6660 bool mem_below_4gb_not_mirrored = false;
6662 for_each_memblock(memory, r) {
6663 if (memblock_is_mirror(r))
6668 usable_startpfn = memblock_region_memory_base_pfn(r);
6670 if (usable_startpfn < 0x100000) {
6671 mem_below_4gb_not_mirrored = true;
6675 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6676 min(usable_startpfn, zone_movable_pfn[nid]) :
6680 if (mem_below_4gb_not_mirrored)
6681 pr_warn("This configuration results in unmirrored kernel memory.");
6687 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6688 * amount of necessary memory.
6690 if (required_kernelcore_percent)
6691 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6693 if (required_movablecore_percent)
6694 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6698 * If movablecore= was specified, calculate what size of
6699 * kernelcore that corresponds so that memory usable for
6700 * any allocation type is evenly spread. If both kernelcore
6701 * and movablecore are specified, then the value of kernelcore
6702 * will be used for required_kernelcore if it's greater than
6703 * what movablecore would have allowed.
6705 if (required_movablecore) {
6706 unsigned long corepages;
6709 * Round-up so that ZONE_MOVABLE is at least as large as what
6710 * was requested by the user
6712 required_movablecore =
6713 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6714 required_movablecore = min(totalpages, required_movablecore);
6715 corepages = totalpages - required_movablecore;
6717 required_kernelcore = max(required_kernelcore, corepages);
6721 * If kernelcore was not specified or kernelcore size is larger
6722 * than totalpages, there is no ZONE_MOVABLE.
6724 if (!required_kernelcore || required_kernelcore >= totalpages)
6727 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6728 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6731 /* Spread kernelcore memory as evenly as possible throughout nodes */
6732 kernelcore_node = required_kernelcore / usable_nodes;
6733 for_each_node_state(nid, N_MEMORY) {
6734 unsigned long start_pfn, end_pfn;
6737 * Recalculate kernelcore_node if the division per node
6738 * now exceeds what is necessary to satisfy the requested
6739 * amount of memory for the kernel
6741 if (required_kernelcore < kernelcore_node)
6742 kernelcore_node = required_kernelcore / usable_nodes;
6745 * As the map is walked, we track how much memory is usable
6746 * by the kernel using kernelcore_remaining. When it is
6747 * 0, the rest of the node is usable by ZONE_MOVABLE
6749 kernelcore_remaining = kernelcore_node;
6751 /* Go through each range of PFNs within this node */
6752 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6753 unsigned long size_pages;
6755 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6756 if (start_pfn >= end_pfn)
6759 /* Account for what is only usable for kernelcore */
6760 if (start_pfn < usable_startpfn) {
6761 unsigned long kernel_pages;
6762 kernel_pages = min(end_pfn, usable_startpfn)
6765 kernelcore_remaining -= min(kernel_pages,
6766 kernelcore_remaining);
6767 required_kernelcore -= min(kernel_pages,
6768 required_kernelcore);
6770 /* Continue if range is now fully accounted */
6771 if (end_pfn <= usable_startpfn) {
6774 * Push zone_movable_pfn to the end so
6775 * that if we have to rebalance
6776 * kernelcore across nodes, we will
6777 * not double account here
6779 zone_movable_pfn[nid] = end_pfn;
6782 start_pfn = usable_startpfn;
6786 * The usable PFN range for ZONE_MOVABLE is from
6787 * start_pfn->end_pfn. Calculate size_pages as the
6788 * number of pages used as kernelcore
6790 size_pages = end_pfn - start_pfn;
6791 if (size_pages > kernelcore_remaining)
6792 size_pages = kernelcore_remaining;
6793 zone_movable_pfn[nid] = start_pfn + size_pages;
6796 * Some kernelcore has been met, update counts and
6797 * break if the kernelcore for this node has been
6800 required_kernelcore -= min(required_kernelcore,
6802 kernelcore_remaining -= size_pages;
6803 if (!kernelcore_remaining)
6809 * If there is still required_kernelcore, we do another pass with one
6810 * less node in the count. This will push zone_movable_pfn[nid] further
6811 * along on the nodes that still have memory until kernelcore is
6815 if (usable_nodes && required_kernelcore > usable_nodes)
6819 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6820 for (nid = 0; nid < MAX_NUMNODES; nid++)
6821 zone_movable_pfn[nid] =
6822 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6825 /* restore the node_state */
6826 node_states[N_MEMORY] = saved_node_state;
6829 /* Any regular or high memory on that node ? */
6830 static void check_for_memory(pg_data_t *pgdat, int nid)
6832 enum zone_type zone_type;
6834 if (N_MEMORY == N_NORMAL_MEMORY)
6837 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6838 struct zone *zone = &pgdat->node_zones[zone_type];
6839 if (populated_zone(zone)) {
6840 node_set_state(nid, N_HIGH_MEMORY);
6841 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6842 zone_type <= ZONE_NORMAL)
6843 node_set_state(nid, N_NORMAL_MEMORY);
6850 * free_area_init_nodes - Initialise all pg_data_t and zone data
6851 * @max_zone_pfn: an array of max PFNs for each zone
6853 * This will call free_area_init_node() for each active node in the system.
6854 * Using the page ranges provided by memblock_set_node(), the size of each
6855 * zone in each node and their holes is calculated. If the maximum PFN
6856 * between two adjacent zones match, it is assumed that the zone is empty.
6857 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6858 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6859 * starts where the previous one ended. For example, ZONE_DMA32 starts
6860 * at arch_max_dma_pfn.
6862 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6864 unsigned long start_pfn, end_pfn;
6867 /* Record where the zone boundaries are */
6868 memset(arch_zone_lowest_possible_pfn, 0,
6869 sizeof(arch_zone_lowest_possible_pfn));
6870 memset(arch_zone_highest_possible_pfn, 0,
6871 sizeof(arch_zone_highest_possible_pfn));
6873 start_pfn = find_min_pfn_with_active_regions();
6875 for (i = 0; i < MAX_NR_ZONES; i++) {
6876 if (i == ZONE_MOVABLE)
6879 end_pfn = max(max_zone_pfn[i], start_pfn);
6880 arch_zone_lowest_possible_pfn[i] = start_pfn;
6881 arch_zone_highest_possible_pfn[i] = end_pfn;
6883 start_pfn = end_pfn;
6886 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6887 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6888 find_zone_movable_pfns_for_nodes();
6890 /* Print out the zone ranges */
6891 pr_info("Zone ranges:\n");
6892 for (i = 0; i < MAX_NR_ZONES; i++) {
6893 if (i == ZONE_MOVABLE)
6895 pr_info(" %-8s ", zone_names[i]);
6896 if (arch_zone_lowest_possible_pfn[i] ==
6897 arch_zone_highest_possible_pfn[i])
6900 pr_cont("[mem %#018Lx-%#018Lx]\n",
6901 (u64)arch_zone_lowest_possible_pfn[i]
6903 ((u64)arch_zone_highest_possible_pfn[i]
6904 << PAGE_SHIFT) - 1);
6907 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6908 pr_info("Movable zone start for each node\n");
6909 for (i = 0; i < MAX_NUMNODES; i++) {
6910 if (zone_movable_pfn[i])
6911 pr_info(" Node %d: %#018Lx\n", i,
6912 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6915 /* Print out the early node map */
6916 pr_info("Early memory node ranges\n");
6917 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6918 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6919 (u64)start_pfn << PAGE_SHIFT,
6920 ((u64)end_pfn << PAGE_SHIFT) - 1);
6922 /* Initialise every node */
6923 mminit_verify_pageflags_layout();
6924 setup_nr_node_ids();
6925 zero_resv_unavail();
6926 for_each_online_node(nid) {
6927 pg_data_t *pgdat = NODE_DATA(nid);
6928 free_area_init_node(nid, NULL,
6929 find_min_pfn_for_node(nid), NULL);
6931 /* Any memory on that node */
6932 if (pgdat->node_present_pages)
6933 node_set_state(nid, N_MEMORY);
6934 check_for_memory(pgdat, nid);
6938 static int __init cmdline_parse_core(char *p, unsigned long *core,
6939 unsigned long *percent)
6941 unsigned long long coremem;
6947 /* Value may be a percentage of total memory, otherwise bytes */
6948 coremem = simple_strtoull(p, &endptr, 0);
6949 if (*endptr == '%') {
6950 /* Paranoid check for percent values greater than 100 */
6951 WARN_ON(coremem > 100);
6955 coremem = memparse(p, &p);
6956 /* Paranoid check that UL is enough for the coremem value */
6957 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6959 *core = coremem >> PAGE_SHIFT;
6966 * kernelcore=size sets the amount of memory for use for allocations that
6967 * cannot be reclaimed or migrated.
6969 static int __init cmdline_parse_kernelcore(char *p)
6971 /* parse kernelcore=mirror */
6972 if (parse_option_str(p, "mirror")) {
6973 mirrored_kernelcore = true;
6977 return cmdline_parse_core(p, &required_kernelcore,
6978 &required_kernelcore_percent);
6982 * movablecore=size sets the amount of memory for use for allocations that
6983 * can be reclaimed or migrated.
6985 static int __init cmdline_parse_movablecore(char *p)
6987 return cmdline_parse_core(p, &required_movablecore,
6988 &required_movablecore_percent);
6991 early_param("kernelcore", cmdline_parse_kernelcore);
6992 early_param("movablecore", cmdline_parse_movablecore);
6994 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6996 void adjust_managed_page_count(struct page *page, long count)
6998 spin_lock(&managed_page_count_lock);
6999 page_zone(page)->managed_pages += count;
7000 totalram_pages += count;
7001 #ifdef CONFIG_HIGHMEM
7002 if (PageHighMem(page))
7003 totalhigh_pages += count;
7005 spin_unlock(&managed_page_count_lock);
7007 EXPORT_SYMBOL(adjust_managed_page_count);
7009 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
7012 unsigned long pages = 0;
7014 start = (void *)PAGE_ALIGN((unsigned long)start);
7015 end = (void *)((unsigned long)end & PAGE_MASK);
7016 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7017 struct page *page = virt_to_page(pos);
7018 void *direct_map_addr;
7021 * 'direct_map_addr' might be different from 'pos'
7022 * because some architectures' virt_to_page()
7023 * work with aliases. Getting the direct map
7024 * address ensures that we get a _writeable_
7025 * alias for the memset().
7027 direct_map_addr = page_address(page);
7028 if ((unsigned int)poison <= 0xFF)
7029 memset(direct_map_addr, poison, PAGE_SIZE);
7031 free_reserved_page(page);
7035 pr_info("Freeing %s memory: %ldK\n",
7036 s, pages << (PAGE_SHIFT - 10));
7040 EXPORT_SYMBOL(free_reserved_area);
7042 #ifdef CONFIG_HIGHMEM
7043 void free_highmem_page(struct page *page)
7045 __free_reserved_page(page);
7047 page_zone(page)->managed_pages++;
7053 void __init mem_init_print_info(const char *str)
7055 unsigned long physpages, codesize, datasize, rosize, bss_size;
7056 unsigned long init_code_size, init_data_size;
7058 physpages = get_num_physpages();
7059 codesize = _etext - _stext;
7060 datasize = _edata - _sdata;
7061 rosize = __end_rodata - __start_rodata;
7062 bss_size = __bss_stop - __bss_start;
7063 init_data_size = __init_end - __init_begin;
7064 init_code_size = _einittext - _sinittext;
7067 * Detect special cases and adjust section sizes accordingly:
7068 * 1) .init.* may be embedded into .data sections
7069 * 2) .init.text.* may be out of [__init_begin, __init_end],
7070 * please refer to arch/tile/kernel/vmlinux.lds.S.
7071 * 3) .rodata.* may be embedded into .text or .data sections.
7073 #define adj_init_size(start, end, size, pos, adj) \
7075 if (start <= pos && pos < end && size > adj) \
7079 adj_init_size(__init_begin, __init_end, init_data_size,
7080 _sinittext, init_code_size);
7081 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7082 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7083 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7084 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7086 #undef adj_init_size
7088 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7089 #ifdef CONFIG_HIGHMEM
7093 nr_free_pages() << (PAGE_SHIFT - 10),
7094 physpages << (PAGE_SHIFT - 10),
7095 codesize >> 10, datasize >> 10, rosize >> 10,
7096 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7097 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7098 totalcma_pages << (PAGE_SHIFT - 10),
7099 #ifdef CONFIG_HIGHMEM
7100 totalhigh_pages << (PAGE_SHIFT - 10),
7102 str ? ", " : "", str ? str : "");
7106 * set_dma_reserve - set the specified number of pages reserved in the first zone
7107 * @new_dma_reserve: The number of pages to mark reserved
7109 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7110 * In the DMA zone, a significant percentage may be consumed by kernel image
7111 * and other unfreeable allocations which can skew the watermarks badly. This
7112 * function may optionally be used to account for unfreeable pages in the
7113 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7114 * smaller per-cpu batchsize.
7116 void __init set_dma_reserve(unsigned long new_dma_reserve)
7118 dma_reserve = new_dma_reserve;
7121 void __init free_area_init(unsigned long *zones_size)
7123 zero_resv_unavail();
7124 free_area_init_node(0, zones_size,
7125 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7128 static int page_alloc_cpu_dead(unsigned int cpu)
7131 lru_add_drain_cpu(cpu);
7135 * Spill the event counters of the dead processor
7136 * into the current processors event counters.
7137 * This artificially elevates the count of the current
7140 vm_events_fold_cpu(cpu);
7143 * Zero the differential counters of the dead processor
7144 * so that the vm statistics are consistent.
7146 * This is only okay since the processor is dead and cannot
7147 * race with what we are doing.
7149 cpu_vm_stats_fold(cpu);
7153 void __init page_alloc_init(void)
7157 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7158 "mm/page_alloc:dead", NULL,
7159 page_alloc_cpu_dead);
7164 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7165 * or min_free_kbytes changes.
7167 static void calculate_totalreserve_pages(void)
7169 struct pglist_data *pgdat;
7170 unsigned long reserve_pages = 0;
7171 enum zone_type i, j;
7173 for_each_online_pgdat(pgdat) {
7175 pgdat->totalreserve_pages = 0;
7177 for (i = 0; i < MAX_NR_ZONES; i++) {
7178 struct zone *zone = pgdat->node_zones + i;
7181 /* Find valid and maximum lowmem_reserve in the zone */
7182 for (j = i; j < MAX_NR_ZONES; j++) {
7183 if (zone->lowmem_reserve[j] > max)
7184 max = zone->lowmem_reserve[j];
7187 /* we treat the high watermark as reserved pages. */
7188 max += high_wmark_pages(zone);
7190 if (max > zone->managed_pages)
7191 max = zone->managed_pages;
7193 pgdat->totalreserve_pages += max;
7195 reserve_pages += max;
7198 totalreserve_pages = reserve_pages;
7202 * setup_per_zone_lowmem_reserve - called whenever
7203 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7204 * has a correct pages reserved value, so an adequate number of
7205 * pages are left in the zone after a successful __alloc_pages().
7207 static void setup_per_zone_lowmem_reserve(void)
7209 struct pglist_data *pgdat;
7210 enum zone_type j, idx;
7212 for_each_online_pgdat(pgdat) {
7213 for (j = 0; j < MAX_NR_ZONES; j++) {
7214 struct zone *zone = pgdat->node_zones + j;
7215 unsigned long managed_pages = zone->managed_pages;
7217 zone->lowmem_reserve[j] = 0;
7221 struct zone *lower_zone;
7224 lower_zone = pgdat->node_zones + idx;
7226 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7227 sysctl_lowmem_reserve_ratio[idx] = 0;
7228 lower_zone->lowmem_reserve[j] = 0;
7230 lower_zone->lowmem_reserve[j] =
7231 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7233 managed_pages += lower_zone->managed_pages;
7238 /* update totalreserve_pages */
7239 calculate_totalreserve_pages();
7242 static void __setup_per_zone_wmarks(void)
7244 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7245 unsigned long lowmem_pages = 0;
7247 unsigned long flags;
7249 /* Calculate total number of !ZONE_HIGHMEM pages */
7250 for_each_zone(zone) {
7251 if (!is_highmem(zone))
7252 lowmem_pages += zone->managed_pages;
7255 for_each_zone(zone) {
7258 spin_lock_irqsave(&zone->lock, flags);
7259 tmp = (u64)pages_min * zone->managed_pages;
7260 do_div(tmp, lowmem_pages);
7261 if (is_highmem(zone)) {
7263 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7264 * need highmem pages, so cap pages_min to a small
7267 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7268 * deltas control asynch page reclaim, and so should
7269 * not be capped for highmem.
7271 unsigned long min_pages;
7273 min_pages = zone->managed_pages / 1024;
7274 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7275 zone->watermark[WMARK_MIN] = min_pages;
7278 * If it's a lowmem zone, reserve a number of pages
7279 * proportionate to the zone's size.
7281 zone->watermark[WMARK_MIN] = tmp;
7285 * Set the kswapd watermarks distance according to the
7286 * scale factor in proportion to available memory, but
7287 * ensure a minimum size on small systems.
7289 tmp = max_t(u64, tmp >> 2,
7290 mult_frac(zone->managed_pages,
7291 watermark_scale_factor, 10000));
7293 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7294 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7296 spin_unlock_irqrestore(&zone->lock, flags);
7299 /* update totalreserve_pages */
7300 calculate_totalreserve_pages();
7304 * setup_per_zone_wmarks - called when min_free_kbytes changes
7305 * or when memory is hot-{added|removed}
7307 * Ensures that the watermark[min,low,high] values for each zone are set
7308 * correctly with respect to min_free_kbytes.
7310 void setup_per_zone_wmarks(void)
7312 static DEFINE_SPINLOCK(lock);
7315 __setup_per_zone_wmarks();
7320 * Initialise min_free_kbytes.
7322 * For small machines we want it small (128k min). For large machines
7323 * we want it large (64MB max). But it is not linear, because network
7324 * bandwidth does not increase linearly with machine size. We use
7326 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7327 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7343 int __meminit init_per_zone_wmark_min(void)
7345 unsigned long lowmem_kbytes;
7346 int new_min_free_kbytes;
7348 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7349 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7351 if (new_min_free_kbytes > user_min_free_kbytes) {
7352 min_free_kbytes = new_min_free_kbytes;
7353 if (min_free_kbytes < 128)
7354 min_free_kbytes = 128;
7355 if (min_free_kbytes > 65536)
7356 min_free_kbytes = 65536;
7358 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7359 new_min_free_kbytes, user_min_free_kbytes);
7361 setup_per_zone_wmarks();
7362 refresh_zone_stat_thresholds();
7363 setup_per_zone_lowmem_reserve();
7366 setup_min_unmapped_ratio();
7367 setup_min_slab_ratio();
7372 core_initcall(init_per_zone_wmark_min)
7375 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7376 * that we can call two helper functions whenever min_free_kbytes
7379 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7380 void __user *buffer, size_t *length, loff_t *ppos)
7384 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7389 user_min_free_kbytes = min_free_kbytes;
7390 setup_per_zone_wmarks();
7395 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7396 void __user *buffer, size_t *length, loff_t *ppos)
7400 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7405 setup_per_zone_wmarks();
7411 static void setup_min_unmapped_ratio(void)
7416 for_each_online_pgdat(pgdat)
7417 pgdat->min_unmapped_pages = 0;
7420 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7421 sysctl_min_unmapped_ratio) / 100;
7425 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7426 void __user *buffer, size_t *length, loff_t *ppos)
7430 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7434 setup_min_unmapped_ratio();
7439 static void setup_min_slab_ratio(void)
7444 for_each_online_pgdat(pgdat)
7445 pgdat->min_slab_pages = 0;
7448 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7449 sysctl_min_slab_ratio) / 100;
7452 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7453 void __user *buffer, size_t *length, loff_t *ppos)
7457 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7461 setup_min_slab_ratio();
7468 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7469 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7470 * whenever sysctl_lowmem_reserve_ratio changes.
7472 * The reserve ratio obviously has absolutely no relation with the
7473 * minimum watermarks. The lowmem reserve ratio can only make sense
7474 * if in function of the boot time zone sizes.
7476 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7477 void __user *buffer, size_t *length, loff_t *ppos)
7479 proc_dointvec_minmax(table, write, buffer, length, ppos);
7480 setup_per_zone_lowmem_reserve();
7485 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7486 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7487 * pagelist can have before it gets flushed back to buddy allocator.
7489 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7490 void __user *buffer, size_t *length, loff_t *ppos)
7493 int old_percpu_pagelist_fraction;
7496 mutex_lock(&pcp_batch_high_lock);
7497 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7499 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7500 if (!write || ret < 0)
7503 /* Sanity checking to avoid pcp imbalance */
7504 if (percpu_pagelist_fraction &&
7505 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7506 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7512 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7515 for_each_populated_zone(zone) {
7518 for_each_possible_cpu(cpu)
7519 pageset_set_high_and_batch(zone,
7520 per_cpu_ptr(zone->pageset, cpu));
7523 mutex_unlock(&pcp_batch_high_lock);
7528 int hashdist = HASHDIST_DEFAULT;
7530 static int __init set_hashdist(char *str)
7534 hashdist = simple_strtoul(str, &str, 0);
7537 __setup("hashdist=", set_hashdist);
7540 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7542 * Returns the number of pages that arch has reserved but
7543 * is not known to alloc_large_system_hash().
7545 static unsigned long __init arch_reserved_kernel_pages(void)
7552 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7553 * machines. As memory size is increased the scale is also increased but at
7554 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7555 * quadruples the scale is increased by one, which means the size of hash table
7556 * only doubles, instead of quadrupling as well.
7557 * Because 32-bit systems cannot have large physical memory, where this scaling
7558 * makes sense, it is disabled on such platforms.
7560 #if __BITS_PER_LONG > 32
7561 #define ADAPT_SCALE_BASE (64ul << 30)
7562 #define ADAPT_SCALE_SHIFT 2
7563 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7567 * allocate a large system hash table from bootmem
7568 * - it is assumed that the hash table must contain an exact power-of-2
7569 * quantity of entries
7570 * - limit is the number of hash buckets, not the total allocation size
7572 void *__init alloc_large_system_hash(const char *tablename,
7573 unsigned long bucketsize,
7574 unsigned long numentries,
7577 unsigned int *_hash_shift,
7578 unsigned int *_hash_mask,
7579 unsigned long low_limit,
7580 unsigned long high_limit)
7582 unsigned long long max = high_limit;
7583 unsigned long log2qty, size;
7587 /* allow the kernel cmdline to have a say */
7589 /* round applicable memory size up to nearest megabyte */
7590 numentries = nr_kernel_pages;
7591 numentries -= arch_reserved_kernel_pages();
7593 /* It isn't necessary when PAGE_SIZE >= 1MB */
7594 if (PAGE_SHIFT < 20)
7595 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7597 #if __BITS_PER_LONG > 32
7599 unsigned long adapt;
7601 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7602 adapt <<= ADAPT_SCALE_SHIFT)
7607 /* limit to 1 bucket per 2^scale bytes of low memory */
7608 if (scale > PAGE_SHIFT)
7609 numentries >>= (scale - PAGE_SHIFT);
7611 numentries <<= (PAGE_SHIFT - scale);
7613 /* Make sure we've got at least a 0-order allocation.. */
7614 if (unlikely(flags & HASH_SMALL)) {
7615 /* Makes no sense without HASH_EARLY */
7616 WARN_ON(!(flags & HASH_EARLY));
7617 if (!(numentries >> *_hash_shift)) {
7618 numentries = 1UL << *_hash_shift;
7619 BUG_ON(!numentries);
7621 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7622 numentries = PAGE_SIZE / bucketsize;
7624 numentries = roundup_pow_of_two(numentries);
7626 /* limit allocation size to 1/16 total memory by default */
7628 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7629 do_div(max, bucketsize);
7631 max = min(max, 0x80000000ULL);
7633 if (numentries < low_limit)
7634 numentries = low_limit;
7635 if (numentries > max)
7638 log2qty = ilog2(numentries);
7640 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7642 size = bucketsize << log2qty;
7643 if (flags & HASH_EARLY) {
7644 if (flags & HASH_ZERO)
7645 table = memblock_virt_alloc_nopanic(size, 0);
7647 table = memblock_virt_alloc_raw(size, 0);
7648 } else if (hashdist) {
7649 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7652 * If bucketsize is not a power-of-two, we may free
7653 * some pages at the end of hash table which
7654 * alloc_pages_exact() automatically does
7656 if (get_order(size) < MAX_ORDER) {
7657 table = alloc_pages_exact(size, gfp_flags);
7658 kmemleak_alloc(table, size, 1, gfp_flags);
7661 } while (!table && size > PAGE_SIZE && --log2qty);
7664 panic("Failed to allocate %s hash table\n", tablename);
7666 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7667 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7670 *_hash_shift = log2qty;
7672 *_hash_mask = (1 << log2qty) - 1;
7678 * This function checks whether pageblock includes unmovable pages or not.
7679 * If @count is not zero, it is okay to include less @count unmovable pages
7681 * PageLRU check without isolation or lru_lock could race so that
7682 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7683 * check without lock_page also may miss some movable non-lru pages at
7684 * race condition. So you can't expect this function should be exact.
7686 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7688 bool skip_hwpoisoned_pages)
7690 unsigned long pfn, iter, found;
7693 * TODO we could make this much more efficient by not checking every
7694 * page in the range if we know all of them are in MOVABLE_ZONE and
7695 * that the movable zone guarantees that pages are migratable but
7696 * the later is not the case right now unfortunatelly. E.g. movablecore
7697 * can still lead to having bootmem allocations in zone_movable.
7701 * CMA allocations (alloc_contig_range) really need to mark isolate
7702 * CMA pageblocks even when they are not movable in fact so consider
7703 * them movable here.
7705 if (is_migrate_cma(migratetype) &&
7706 is_migrate_cma(get_pageblock_migratetype(page)))
7709 pfn = page_to_pfn(page);
7710 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7711 unsigned long check = pfn + iter;
7713 if (!pfn_valid_within(check))
7716 page = pfn_to_page(check);
7718 if (PageReserved(page))
7722 * If the zone is movable and we have ruled out all reserved
7723 * pages then it should be reasonably safe to assume the rest
7726 if (zone_idx(zone) == ZONE_MOVABLE)
7730 * Hugepages are not in LRU lists, but they're movable.
7731 * We need not scan over tail pages bacause we don't
7732 * handle each tail page individually in migration.
7734 if (PageHuge(page)) {
7735 struct page *head = compound_head(page);
7736 unsigned int skip_pages;
7738 if (!hugepage_migration_supported(page_hstate(head)))
7741 skip_pages = (1 << compound_order(head)) - (page - head);
7742 iter += skip_pages - 1;
7747 * We can't use page_count without pin a page
7748 * because another CPU can free compound page.
7749 * This check already skips compound tails of THP
7750 * because their page->_refcount is zero at all time.
7752 if (!page_ref_count(page)) {
7753 if (PageBuddy(page))
7754 iter += (1 << page_order(page)) - 1;
7759 * The HWPoisoned page may be not in buddy system, and
7760 * page_count() is not 0.
7762 if (skip_hwpoisoned_pages && PageHWPoison(page))
7765 if (__PageMovable(page))
7771 * If there are RECLAIMABLE pages, we need to check
7772 * it. But now, memory offline itself doesn't call
7773 * shrink_node_slabs() and it still to be fixed.
7776 * If the page is not RAM, page_count()should be 0.
7777 * we don't need more check. This is an _used_ not-movable page.
7779 * The problematic thing here is PG_reserved pages. PG_reserved
7780 * is set to both of a memory hole page and a _used_ kernel
7788 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7792 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7794 static unsigned long pfn_max_align_down(unsigned long pfn)
7796 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7797 pageblock_nr_pages) - 1);
7800 static unsigned long pfn_max_align_up(unsigned long pfn)
7802 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7803 pageblock_nr_pages));
7806 /* [start, end) must belong to a single zone. */
7807 static int __alloc_contig_migrate_range(struct compact_control *cc,
7808 unsigned long start, unsigned long end)
7810 /* This function is based on compact_zone() from compaction.c. */
7811 unsigned long nr_reclaimed;
7812 unsigned long pfn = start;
7813 unsigned int tries = 0;
7818 while (pfn < end || !list_empty(&cc->migratepages)) {
7819 if (fatal_signal_pending(current)) {
7824 if (list_empty(&cc->migratepages)) {
7825 cc->nr_migratepages = 0;
7826 pfn = isolate_migratepages_range(cc, pfn, end);
7832 } else if (++tries == 5) {
7833 ret = ret < 0 ? ret : -EBUSY;
7837 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7839 cc->nr_migratepages -= nr_reclaimed;
7841 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7842 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7845 putback_movable_pages(&cc->migratepages);
7852 * alloc_contig_range() -- tries to allocate given range of pages
7853 * @start: start PFN to allocate
7854 * @end: one-past-the-last PFN to allocate
7855 * @migratetype: migratetype of the underlaying pageblocks (either
7856 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7857 * in range must have the same migratetype and it must
7858 * be either of the two.
7859 * @gfp_mask: GFP mask to use during compaction
7861 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7862 * aligned. The PFN range must belong to a single zone.
7864 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7865 * pageblocks in the range. Once isolated, the pageblocks should not
7866 * be modified by others.
7868 * Returns zero on success or negative error code. On success all
7869 * pages which PFN is in [start, end) are allocated for the caller and
7870 * need to be freed with free_contig_range().
7872 int alloc_contig_range(unsigned long start, unsigned long end,
7873 unsigned migratetype, gfp_t gfp_mask)
7875 unsigned long outer_start, outer_end;
7879 struct compact_control cc = {
7880 .nr_migratepages = 0,
7882 .zone = page_zone(pfn_to_page(start)),
7883 .mode = MIGRATE_SYNC,
7884 .ignore_skip_hint = true,
7885 .no_set_skip_hint = true,
7886 .gfp_mask = current_gfp_context(gfp_mask),
7888 INIT_LIST_HEAD(&cc.migratepages);
7891 * What we do here is we mark all pageblocks in range as
7892 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7893 * have different sizes, and due to the way page allocator
7894 * work, we align the range to biggest of the two pages so
7895 * that page allocator won't try to merge buddies from
7896 * different pageblocks and change MIGRATE_ISOLATE to some
7897 * other migration type.
7899 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7900 * migrate the pages from an unaligned range (ie. pages that
7901 * we are interested in). This will put all the pages in
7902 * range back to page allocator as MIGRATE_ISOLATE.
7904 * When this is done, we take the pages in range from page
7905 * allocator removing them from the buddy system. This way
7906 * page allocator will never consider using them.
7908 * This lets us mark the pageblocks back as
7909 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7910 * aligned range but not in the unaligned, original range are
7911 * put back to page allocator so that buddy can use them.
7914 ret = start_isolate_page_range(pfn_max_align_down(start),
7915 pfn_max_align_up(end), migratetype,
7921 * In case of -EBUSY, we'd like to know which page causes problem.
7922 * So, just fall through. test_pages_isolated() has a tracepoint
7923 * which will report the busy page.
7925 * It is possible that busy pages could become available before
7926 * the call to test_pages_isolated, and the range will actually be
7927 * allocated. So, if we fall through be sure to clear ret so that
7928 * -EBUSY is not accidentally used or returned to caller.
7930 ret = __alloc_contig_migrate_range(&cc, start, end);
7931 if (ret && ret != -EBUSY)
7936 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7937 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7938 * more, all pages in [start, end) are free in page allocator.
7939 * What we are going to do is to allocate all pages from
7940 * [start, end) (that is remove them from page allocator).
7942 * The only problem is that pages at the beginning and at the
7943 * end of interesting range may be not aligned with pages that
7944 * page allocator holds, ie. they can be part of higher order
7945 * pages. Because of this, we reserve the bigger range and
7946 * once this is done free the pages we are not interested in.
7948 * We don't have to hold zone->lock here because the pages are
7949 * isolated thus they won't get removed from buddy.
7952 lru_add_drain_all();
7953 drain_all_pages(cc.zone);
7956 outer_start = start;
7957 while (!PageBuddy(pfn_to_page(outer_start))) {
7958 if (++order >= MAX_ORDER) {
7959 outer_start = start;
7962 outer_start &= ~0UL << order;
7965 if (outer_start != start) {
7966 order = page_order(pfn_to_page(outer_start));
7969 * outer_start page could be small order buddy page and
7970 * it doesn't include start page. Adjust outer_start
7971 * in this case to report failed page properly
7972 * on tracepoint in test_pages_isolated()
7974 if (outer_start + (1UL << order) <= start)
7975 outer_start = start;
7978 /* Make sure the range is really isolated. */
7979 if (test_pages_isolated(outer_start, end, false)) {
7980 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7981 __func__, outer_start, end);
7986 /* Grab isolated pages from freelists. */
7987 outer_end = isolate_freepages_range(&cc, outer_start, end);
7993 /* Free head and tail (if any) */
7994 if (start != outer_start)
7995 free_contig_range(outer_start, start - outer_start);
7996 if (end != outer_end)
7997 free_contig_range(end, outer_end - end);
8000 undo_isolate_page_range(pfn_max_align_down(start),
8001 pfn_max_align_up(end), migratetype);
8005 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8007 unsigned int count = 0;
8009 for (; nr_pages--; pfn++) {
8010 struct page *page = pfn_to_page(pfn);
8012 count += page_count(page) != 1;
8015 WARN(count != 0, "%d pages are still in use!\n", count);
8020 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8021 * page high values need to be recalulated.
8023 void __meminit zone_pcp_update(struct zone *zone)
8026 mutex_lock(&pcp_batch_high_lock);
8027 for_each_possible_cpu(cpu)
8028 pageset_set_high_and_batch(zone,
8029 per_cpu_ptr(zone->pageset, cpu));
8030 mutex_unlock(&pcp_batch_high_lock);
8033 void zone_pcp_reset(struct zone *zone)
8035 unsigned long flags;
8037 struct per_cpu_pageset *pset;
8039 /* avoid races with drain_pages() */
8040 local_irq_save(flags);
8041 if (zone->pageset != &boot_pageset) {
8042 for_each_online_cpu(cpu) {
8043 pset = per_cpu_ptr(zone->pageset, cpu);
8044 drain_zonestat(zone, pset);
8046 free_percpu(zone->pageset);
8047 zone->pageset = &boot_pageset;
8049 local_irq_restore(flags);
8052 #ifdef CONFIG_MEMORY_HOTREMOVE
8054 * All pages in the range must be in a single zone and isolated
8055 * before calling this.
8058 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8062 unsigned int order, i;
8064 unsigned long flags;
8065 /* find the first valid pfn */
8066 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8071 offline_mem_sections(pfn, end_pfn);
8072 zone = page_zone(pfn_to_page(pfn));
8073 spin_lock_irqsave(&zone->lock, flags);
8075 while (pfn < end_pfn) {
8076 if (!pfn_valid(pfn)) {
8080 page = pfn_to_page(pfn);
8082 * The HWPoisoned page may be not in buddy system, and
8083 * page_count() is not 0.
8085 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8087 SetPageReserved(page);
8091 BUG_ON(page_count(page));
8092 BUG_ON(!PageBuddy(page));
8093 order = page_order(page);
8094 #ifdef CONFIG_DEBUG_VM
8095 pr_info("remove from free list %lx %d %lx\n",
8096 pfn, 1 << order, end_pfn);
8098 list_del(&page->lru);
8099 rmv_page_order(page);
8100 zone->free_area[order].nr_free--;
8101 for (i = 0; i < (1 << order); i++)
8102 SetPageReserved((page+i));
8103 pfn += (1 << order);
8105 spin_unlock_irqrestore(&zone->lock, flags);
8109 bool is_free_buddy_page(struct page *page)
8111 struct zone *zone = page_zone(page);
8112 unsigned long pfn = page_to_pfn(page);
8113 unsigned long flags;
8116 spin_lock_irqsave(&zone->lock, flags);
8117 for (order = 0; order < MAX_ORDER; order++) {
8118 struct page *page_head = page - (pfn & ((1 << order) - 1));
8120 if (PageBuddy(page_head) && page_order(page_head) >= order)
8123 spin_unlock_irqrestore(&zone->lock, flags);
8125 return order < MAX_ORDER;
8128 #ifdef CONFIG_MEMORY_FAILURE
8130 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8131 * test is performed under the zone lock to prevent a race against page
8134 bool set_hwpoison_free_buddy_page(struct page *page)
8136 struct zone *zone = page_zone(page);
8137 unsigned long pfn = page_to_pfn(page);
8138 unsigned long flags;
8140 bool hwpoisoned = false;
8142 spin_lock_irqsave(&zone->lock, flags);
8143 for (order = 0; order < MAX_ORDER; order++) {
8144 struct page *page_head = page - (pfn & ((1 << order) - 1));
8146 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8147 if (!TestSetPageHWPoison(page))
8152 spin_unlock_irqrestore(&zone->lock, flags);