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/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 atomic_long_t _totalram_pages __read_mostly;
126 EXPORT_SYMBOL(_totalram_pages);
127 unsigned long totalreserve_pages __read_mostly;
128 unsigned long totalcma_pages __read_mostly;
130 int percpu_pagelist_fraction;
131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page *page)
146 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
148 page->index = migratetype;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with system_transition_mutex held
157 * (gfp_allowed_mask also should only be modified with system_transition_mutex
158 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
159 * with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&system_transition_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&system_transition_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
214 #ifdef CONFIG_HIGHMEM
220 EXPORT_SYMBOL(totalram_pages);
222 static char * const zone_names[MAX_NR_ZONES] = {
223 #ifdef CONFIG_ZONE_DMA
226 #ifdef CONFIG_ZONE_DMA32
230 #ifdef CONFIG_HIGHMEM
234 #ifdef CONFIG_ZONE_DEVICE
239 char * const migratetype_names[MIGRATE_TYPES] = {
247 #ifdef CONFIG_MEMORY_ISOLATION
252 compound_page_dtor * const compound_page_dtors[] = {
255 #ifdef CONFIG_HUGETLB_PAGE
258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
263 int min_free_kbytes = 1024;
264 int user_min_free_kbytes = -1;
265 int watermark_scale_factor = 10;
267 static unsigned long nr_kernel_pages __meminitdata;
268 static unsigned long nr_all_pages __meminitdata;
269 static unsigned long dma_reserve __meminitdata;
271 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
272 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
273 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274 static unsigned long required_kernelcore __initdata;
275 static unsigned long required_kernelcore_percent __initdata;
276 static unsigned long required_movablecore __initdata;
277 static unsigned long required_movablecore_percent __initdata;
278 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
279 static bool mirrored_kernelcore __meminitdata;
281 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
283 EXPORT_SYMBOL(movable_zone);
284 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
287 int nr_node_ids __read_mostly = MAX_NUMNODES;
288 int nr_online_nodes __read_mostly = 1;
289 EXPORT_SYMBOL(nr_node_ids);
290 EXPORT_SYMBOL(nr_online_nodes);
293 int page_group_by_mobility_disabled __read_mostly;
295 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
296 /* Returns true if the struct page for the pfn is uninitialised */
297 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
299 int nid = early_pfn_to_nid(pfn);
301 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
308 * Returns true when the remaining initialisation should be deferred until
309 * later in the boot cycle when it can be parallelised.
311 static bool __meminit
312 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
314 static unsigned long prev_end_pfn, nr_initialised;
317 * prev_end_pfn static that contains the end of previous zone
318 * No need to protect because called very early in boot before smp_init.
320 if (prev_end_pfn != end_pfn) {
321 prev_end_pfn = end_pfn;
325 /* Always populate low zones for address-constrained allocations */
326 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
329 if ((nr_initialised > NODE_DATA(nid)->static_init_pgcnt) &&
330 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
331 NODE_DATA(nid)->first_deferred_pfn = pfn;
337 static inline bool early_page_uninitialised(unsigned long pfn)
342 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
348 /* Return a pointer to the bitmap storing bits affecting a block of pages */
349 static inline unsigned long *get_pageblock_bitmap(struct page *page,
352 #ifdef CONFIG_SPARSEMEM
353 return __pfn_to_section(pfn)->pageblock_flags;
355 return page_zone(page)->pageblock_flags;
356 #endif /* CONFIG_SPARSEMEM */
359 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
361 #ifdef CONFIG_SPARSEMEM
362 pfn &= (PAGES_PER_SECTION-1);
363 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
365 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
366 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
367 #endif /* CONFIG_SPARSEMEM */
371 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
372 * @page: The page within the block of interest
373 * @pfn: The target page frame number
374 * @end_bitidx: The last bit of interest to retrieve
375 * @mask: mask of bits that the caller is interested in
377 * Return: pageblock_bits flags
379 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
381 unsigned long end_bitidx,
384 unsigned long *bitmap;
385 unsigned long bitidx, word_bitidx;
388 bitmap = get_pageblock_bitmap(page, pfn);
389 bitidx = pfn_to_bitidx(page, pfn);
390 word_bitidx = bitidx / BITS_PER_LONG;
391 bitidx &= (BITS_PER_LONG-1);
393 word = bitmap[word_bitidx];
394 bitidx += end_bitidx;
395 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
398 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
399 unsigned long end_bitidx,
402 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
405 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
407 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
411 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
412 * @page: The page within the block of interest
413 * @flags: The flags to set
414 * @pfn: The target page frame number
415 * @end_bitidx: The last bit of interest
416 * @mask: mask of bits that the caller is interested in
418 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
420 unsigned long end_bitidx,
423 unsigned long *bitmap;
424 unsigned long bitidx, word_bitidx;
425 unsigned long old_word, word;
427 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
429 bitmap = get_pageblock_bitmap(page, pfn);
430 bitidx = pfn_to_bitidx(page, pfn);
431 word_bitidx = bitidx / BITS_PER_LONG;
432 bitidx &= (BITS_PER_LONG-1);
434 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
436 bitidx += end_bitidx;
437 mask <<= (BITS_PER_LONG - bitidx - 1);
438 flags <<= (BITS_PER_LONG - bitidx - 1);
440 word = READ_ONCE(bitmap[word_bitidx]);
442 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
443 if (word == old_word)
449 void set_pageblock_migratetype(struct page *page, int migratetype)
451 if (unlikely(page_group_by_mobility_disabled &&
452 migratetype < MIGRATE_PCPTYPES))
453 migratetype = MIGRATE_UNMOVABLE;
455 set_pageblock_flags_group(page, (unsigned long)migratetype,
456 PB_migrate, PB_migrate_end);
459 #ifdef CONFIG_DEBUG_VM
460 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
464 unsigned long pfn = page_to_pfn(page);
465 unsigned long sp, start_pfn;
468 seq = zone_span_seqbegin(zone);
469 start_pfn = zone->zone_start_pfn;
470 sp = zone->spanned_pages;
471 if (!zone_spans_pfn(zone, pfn))
473 } while (zone_span_seqretry(zone, seq));
476 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
477 pfn, zone_to_nid(zone), zone->name,
478 start_pfn, start_pfn + sp);
483 static int page_is_consistent(struct zone *zone, struct page *page)
485 if (!pfn_valid_within(page_to_pfn(page)))
487 if (zone != page_zone(page))
493 * Temporary debugging check for pages not lying within a given zone.
495 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
497 if (page_outside_zone_boundaries(zone, page))
499 if (!page_is_consistent(zone, page))
505 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
511 static void bad_page(struct page *page, const char *reason,
512 unsigned long bad_flags)
514 static unsigned long resume;
515 static unsigned long nr_shown;
516 static unsigned long nr_unshown;
519 * Allow a burst of 60 reports, then keep quiet for that minute;
520 * or allow a steady drip of one report per second.
522 if (nr_shown == 60) {
523 if (time_before(jiffies, resume)) {
529 "BUG: Bad page state: %lu messages suppressed\n",
536 resume = jiffies + 60 * HZ;
538 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
539 current->comm, page_to_pfn(page));
540 __dump_page(page, reason);
541 bad_flags &= page->flags;
543 pr_alert("bad because of flags: %#lx(%pGp)\n",
544 bad_flags, &bad_flags);
545 dump_page_owner(page);
550 /* Leave bad fields for debug, except PageBuddy could make trouble */
551 page_mapcount_reset(page); /* remove PageBuddy */
552 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
556 * Higher-order pages are called "compound pages". They are structured thusly:
558 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
560 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
561 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
563 * The first tail page's ->compound_dtor holds the offset in array of compound
564 * page destructors. See compound_page_dtors.
566 * The first tail page's ->compound_order holds the order of allocation.
567 * This usage means that zero-order pages may not be compound.
570 void free_compound_page(struct page *page)
572 __free_pages_ok(page, compound_order(page));
575 void prep_compound_page(struct page *page, unsigned int order)
578 int nr_pages = 1 << order;
580 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
581 set_compound_order(page, order);
583 for (i = 1; i < nr_pages; i++) {
584 struct page *p = page + i;
585 set_page_count(p, 0);
586 p->mapping = TAIL_MAPPING;
587 set_compound_head(p, page);
589 atomic_set(compound_mapcount_ptr(page), -1);
592 #ifdef CONFIG_DEBUG_PAGEALLOC
593 unsigned int _debug_guardpage_minorder;
594 bool _debug_pagealloc_enabled __read_mostly
595 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
596 EXPORT_SYMBOL(_debug_pagealloc_enabled);
597 bool _debug_guardpage_enabled __read_mostly;
599 static int __init early_debug_pagealloc(char *buf)
603 return kstrtobool(buf, &_debug_pagealloc_enabled);
605 early_param("debug_pagealloc", early_debug_pagealloc);
607 static bool need_debug_guardpage(void)
609 /* If we don't use debug_pagealloc, we don't need guard page */
610 if (!debug_pagealloc_enabled())
613 if (!debug_guardpage_minorder())
619 static void init_debug_guardpage(void)
621 if (!debug_pagealloc_enabled())
624 if (!debug_guardpage_minorder())
627 _debug_guardpage_enabled = true;
630 struct page_ext_operations debug_guardpage_ops = {
631 .need = need_debug_guardpage,
632 .init = init_debug_guardpage,
635 static int __init debug_guardpage_minorder_setup(char *buf)
639 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
640 pr_err("Bad debug_guardpage_minorder value\n");
643 _debug_guardpage_minorder = res;
644 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
647 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
649 static inline bool set_page_guard(struct zone *zone, struct page *page,
650 unsigned int order, int migratetype)
652 struct page_ext *page_ext;
654 if (!debug_guardpage_enabled())
657 if (order >= debug_guardpage_minorder())
660 page_ext = lookup_page_ext(page);
661 if (unlikely(!page_ext))
664 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
666 INIT_LIST_HEAD(&page->lru);
667 set_page_private(page, order);
668 /* Guard pages are not available for any usage */
669 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
674 static inline void clear_page_guard(struct zone *zone, struct page *page,
675 unsigned int order, int migratetype)
677 struct page_ext *page_ext;
679 if (!debug_guardpage_enabled())
682 page_ext = lookup_page_ext(page);
683 if (unlikely(!page_ext))
686 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
688 set_page_private(page, 0);
689 if (!is_migrate_isolate(migratetype))
690 __mod_zone_freepage_state(zone, (1 << order), migratetype);
693 struct page_ext_operations debug_guardpage_ops;
694 static inline bool set_page_guard(struct zone *zone, struct page *page,
695 unsigned int order, int migratetype) { return false; }
696 static inline void clear_page_guard(struct zone *zone, struct page *page,
697 unsigned int order, int migratetype) {}
700 static inline void set_page_order(struct page *page, unsigned int order)
702 set_page_private(page, order);
703 __SetPageBuddy(page);
706 static inline void rmv_page_order(struct page *page)
708 __ClearPageBuddy(page);
709 set_page_private(page, 0);
713 * This function checks whether a page is free && is the buddy
714 * we can coalesce a page and its buddy if
715 * (a) the buddy is not in a hole (check before calling!) &&
716 * (b) the buddy is in the buddy system &&
717 * (c) a page and its buddy have the same order &&
718 * (d) a page and its buddy are in the same zone.
720 * For recording whether a page is in the buddy system, we set PageBuddy.
721 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
723 * For recording page's order, we use page_private(page).
725 static inline int page_is_buddy(struct page *page, struct page *buddy,
728 if (page_is_guard(buddy) && page_order(buddy) == order) {
729 if (page_zone_id(page) != page_zone_id(buddy))
732 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
737 if (PageBuddy(buddy) && page_order(buddy) == order) {
739 * zone check is done late to avoid uselessly
740 * calculating zone/node ids for pages that could
743 if (page_zone_id(page) != page_zone_id(buddy))
746 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
754 * Freeing function for a buddy system allocator.
756 * The concept of a buddy system is to maintain direct-mapped table
757 * (containing bit values) for memory blocks of various "orders".
758 * The bottom level table contains the map for the smallest allocatable
759 * units of memory (here, pages), and each level above it describes
760 * pairs of units from the levels below, hence, "buddies".
761 * At a high level, all that happens here is marking the table entry
762 * at the bottom level available, and propagating the changes upward
763 * as necessary, plus some accounting needed to play nicely with other
764 * parts of the VM system.
765 * At each level, we keep a list of pages, which are heads of continuous
766 * free pages of length of (1 << order) and marked with PageBuddy.
767 * Page's order is recorded in page_private(page) field.
768 * So when we are allocating or freeing one, we can derive the state of the
769 * other. That is, if we allocate a small block, and both were
770 * free, the remainder of the region must be split into blocks.
771 * If a block is freed, and its buddy is also free, then this
772 * triggers coalescing into a block of larger size.
777 static inline void __free_one_page(struct page *page,
779 struct zone *zone, unsigned int order,
782 unsigned long combined_pfn;
783 unsigned long uninitialized_var(buddy_pfn);
785 unsigned int max_order;
787 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
789 VM_BUG_ON(!zone_is_initialized(zone));
790 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
792 VM_BUG_ON(migratetype == -1);
793 if (likely(!is_migrate_isolate(migratetype)))
794 __mod_zone_freepage_state(zone, 1 << order, migratetype);
796 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
797 VM_BUG_ON_PAGE(bad_range(zone, page), page);
800 while (order < max_order - 1) {
801 buddy_pfn = __find_buddy_pfn(pfn, order);
802 buddy = page + (buddy_pfn - pfn);
804 if (!pfn_valid_within(buddy_pfn))
806 if (!page_is_buddy(page, buddy, order))
809 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
810 * merge with it and move up one order.
812 if (page_is_guard(buddy)) {
813 clear_page_guard(zone, buddy, order, migratetype);
815 list_del(&buddy->lru);
816 zone->free_area[order].nr_free--;
817 rmv_page_order(buddy);
819 combined_pfn = buddy_pfn & pfn;
820 page = page + (combined_pfn - pfn);
824 if (max_order < MAX_ORDER) {
825 /* If we are here, it means order is >= pageblock_order.
826 * We want to prevent merge between freepages on isolate
827 * pageblock and normal pageblock. Without this, pageblock
828 * isolation could cause incorrect freepage or CMA accounting.
830 * We don't want to hit this code for the more frequent
833 if (unlikely(has_isolate_pageblock(zone))) {
836 buddy_pfn = __find_buddy_pfn(pfn, order);
837 buddy = page + (buddy_pfn - pfn);
838 buddy_mt = get_pageblock_migratetype(buddy);
840 if (migratetype != buddy_mt
841 && (is_migrate_isolate(migratetype) ||
842 is_migrate_isolate(buddy_mt)))
846 goto continue_merging;
850 set_page_order(page, order);
853 * If this is not the largest possible page, check if the buddy
854 * of the next-highest order is free. If it is, it's possible
855 * that pages are being freed that will coalesce soon. In case,
856 * that is happening, add the free page to the tail of the list
857 * so it's less likely to be used soon and more likely to be merged
858 * as a higher order page
860 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
861 struct page *higher_page, *higher_buddy;
862 combined_pfn = buddy_pfn & pfn;
863 higher_page = page + (combined_pfn - pfn);
864 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
865 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
866 if (pfn_valid_within(buddy_pfn) &&
867 page_is_buddy(higher_page, higher_buddy, order + 1)) {
868 list_add_tail(&page->lru,
869 &zone->free_area[order].free_list[migratetype]);
874 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
876 zone->free_area[order].nr_free++;
880 * A bad page could be due to a number of fields. Instead of multiple branches,
881 * try and check multiple fields with one check. The caller must do a detailed
882 * check if necessary.
884 static inline bool page_expected_state(struct page *page,
885 unsigned long check_flags)
887 if (unlikely(atomic_read(&page->_mapcount) != -1))
890 if (unlikely((unsigned long)page->mapping |
891 page_ref_count(page) |
893 (unsigned long)page->mem_cgroup |
895 (page->flags & check_flags)))
901 static void free_pages_check_bad(struct page *page)
903 const char *bad_reason;
904 unsigned long bad_flags;
909 if (unlikely(atomic_read(&page->_mapcount) != -1))
910 bad_reason = "nonzero mapcount";
911 if (unlikely(page->mapping != NULL))
912 bad_reason = "non-NULL mapping";
913 if (unlikely(page_ref_count(page) != 0))
914 bad_reason = "nonzero _refcount";
915 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
916 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
917 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
920 if (unlikely(page->mem_cgroup))
921 bad_reason = "page still charged to cgroup";
923 bad_page(page, bad_reason, bad_flags);
926 static inline int free_pages_check(struct page *page)
928 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
931 /* Something has gone sideways, find it */
932 free_pages_check_bad(page);
936 static int free_tail_pages_check(struct page *head_page, struct page *page)
941 * We rely page->lru.next never has bit 0 set, unless the page
942 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
944 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
946 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
950 switch (page - head_page) {
952 /* the first tail page: ->mapping may be compound_mapcount() */
953 if (unlikely(compound_mapcount(page))) {
954 bad_page(page, "nonzero compound_mapcount", 0);
960 * the second tail page: ->mapping is
961 * deferred_list.next -- ignore value.
965 if (page->mapping != TAIL_MAPPING) {
966 bad_page(page, "corrupted mapping in tail page", 0);
971 if (unlikely(!PageTail(page))) {
972 bad_page(page, "PageTail not set", 0);
975 if (unlikely(compound_head(page) != head_page)) {
976 bad_page(page, "compound_head not consistent", 0);
981 page->mapping = NULL;
982 clear_compound_head(page);
986 static __always_inline bool free_pages_prepare(struct page *page,
987 unsigned int order, bool check_free)
991 VM_BUG_ON_PAGE(PageTail(page), page);
993 trace_mm_page_free(page, order);
996 * Check tail pages before head page information is cleared to
997 * avoid checking PageCompound for order-0 pages.
999 if (unlikely(order)) {
1000 bool compound = PageCompound(page);
1003 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1006 ClearPageDoubleMap(page);
1007 for (i = 1; i < (1 << order); i++) {
1009 bad += free_tail_pages_check(page, page + i);
1010 if (unlikely(free_pages_check(page + i))) {
1014 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1017 if (PageMappingFlags(page))
1018 page->mapping = NULL;
1019 if (memcg_kmem_enabled() && PageKmemcg(page))
1020 memcg_kmem_uncharge(page, order);
1022 bad += free_pages_check(page);
1026 page_cpupid_reset_last(page);
1027 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1028 reset_page_owner(page, order);
1030 if (!PageHighMem(page)) {
1031 debug_check_no_locks_freed(page_address(page),
1032 PAGE_SIZE << order);
1033 debug_check_no_obj_freed(page_address(page),
1034 PAGE_SIZE << order);
1036 arch_free_page(page, order);
1037 kernel_poison_pages(page, 1 << order, 0);
1038 kernel_map_pages(page, 1 << order, 0);
1039 kasan_free_pages(page, order);
1044 #ifdef CONFIG_DEBUG_VM
1045 static inline bool free_pcp_prepare(struct page *page)
1047 return free_pages_prepare(page, 0, true);
1050 static inline bool bulkfree_pcp_prepare(struct page *page)
1055 static bool free_pcp_prepare(struct page *page)
1057 return free_pages_prepare(page, 0, false);
1060 static bool bulkfree_pcp_prepare(struct page *page)
1062 return free_pages_check(page);
1064 #endif /* CONFIG_DEBUG_VM */
1066 static inline void prefetch_buddy(struct page *page)
1068 unsigned long pfn = page_to_pfn(page);
1069 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1070 struct page *buddy = page + (buddy_pfn - pfn);
1076 * Frees a number of pages from the PCP lists
1077 * Assumes all pages on list are in same zone, and of same order.
1078 * count is the number of pages to free.
1080 * If the zone was previously in an "all pages pinned" state then look to
1081 * see if this freeing clears that state.
1083 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1084 * pinned" detection logic.
1086 static void free_pcppages_bulk(struct zone *zone, int count,
1087 struct per_cpu_pages *pcp)
1089 int migratetype = 0;
1091 int prefetch_nr = 0;
1092 bool isolated_pageblocks;
1093 struct page *page, *tmp;
1097 struct list_head *list;
1100 * Remove pages from lists in a round-robin fashion. A
1101 * batch_free count is maintained that is incremented when an
1102 * empty list is encountered. This is so more pages are freed
1103 * off fuller lists instead of spinning excessively around empty
1108 if (++migratetype == MIGRATE_PCPTYPES)
1110 list = &pcp->lists[migratetype];
1111 } while (list_empty(list));
1113 /* This is the only non-empty list. Free them all. */
1114 if (batch_free == MIGRATE_PCPTYPES)
1118 page = list_last_entry(list, struct page, lru);
1119 /* must delete to avoid corrupting pcp list */
1120 list_del(&page->lru);
1123 if (bulkfree_pcp_prepare(page))
1126 list_add_tail(&page->lru, &head);
1129 * We are going to put the page back to the global
1130 * pool, prefetch its buddy to speed up later access
1131 * under zone->lock. It is believed the overhead of
1132 * an additional test and calculating buddy_pfn here
1133 * can be offset by reduced memory latency later. To
1134 * avoid excessive prefetching due to large count, only
1135 * prefetch buddy for the first pcp->batch nr of pages.
1137 if (prefetch_nr++ < pcp->batch)
1138 prefetch_buddy(page);
1139 } while (--count && --batch_free && !list_empty(list));
1142 spin_lock(&zone->lock);
1143 isolated_pageblocks = has_isolate_pageblock(zone);
1146 * Use safe version since after __free_one_page(),
1147 * page->lru.next will not point to original list.
1149 list_for_each_entry_safe(page, tmp, &head, lru) {
1150 int mt = get_pcppage_migratetype(page);
1151 /* MIGRATE_ISOLATE page should not go to pcplists */
1152 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1153 /* Pageblock could have been isolated meanwhile */
1154 if (unlikely(isolated_pageblocks))
1155 mt = get_pageblock_migratetype(page);
1157 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1158 trace_mm_page_pcpu_drain(page, 0, mt);
1160 spin_unlock(&zone->lock);
1163 static void free_one_page(struct zone *zone,
1164 struct page *page, unsigned long pfn,
1168 spin_lock(&zone->lock);
1169 if (unlikely(has_isolate_pageblock(zone) ||
1170 is_migrate_isolate(migratetype))) {
1171 migratetype = get_pfnblock_migratetype(page, pfn);
1173 __free_one_page(page, pfn, zone, order, migratetype);
1174 spin_unlock(&zone->lock);
1177 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1178 unsigned long zone, int nid)
1180 mm_zero_struct_page(page);
1181 set_page_links(page, zone, nid, pfn);
1182 init_page_count(page);
1183 page_mapcount_reset(page);
1184 page_cpupid_reset_last(page);
1185 page_kasan_tag_reset(page);
1187 INIT_LIST_HEAD(&page->lru);
1188 #ifdef WANT_PAGE_VIRTUAL
1189 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1190 if (!is_highmem_idx(zone))
1191 set_page_address(page, __va(pfn << PAGE_SHIFT));
1195 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1196 static void __meminit init_reserved_page(unsigned long pfn)
1201 if (!early_page_uninitialised(pfn))
1204 nid = early_pfn_to_nid(pfn);
1205 pgdat = NODE_DATA(nid);
1207 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1208 struct zone *zone = &pgdat->node_zones[zid];
1210 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1213 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1216 static inline void init_reserved_page(unsigned long pfn)
1219 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1222 * Initialised pages do not have PageReserved set. This function is
1223 * called for each range allocated by the bootmem allocator and
1224 * marks the pages PageReserved. The remaining valid pages are later
1225 * sent to the buddy page allocator.
1227 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1229 unsigned long start_pfn = PFN_DOWN(start);
1230 unsigned long end_pfn = PFN_UP(end);
1232 for (; start_pfn < end_pfn; start_pfn++) {
1233 if (pfn_valid(start_pfn)) {
1234 struct page *page = pfn_to_page(start_pfn);
1236 init_reserved_page(start_pfn);
1238 /* Avoid false-positive PageTail() */
1239 INIT_LIST_HEAD(&page->lru);
1242 * no need for atomic set_bit because the struct
1243 * page is not visible yet so nobody should
1246 __SetPageReserved(page);
1251 static void __free_pages_ok(struct page *page, unsigned int order)
1253 unsigned long flags;
1255 unsigned long pfn = page_to_pfn(page);
1257 if (!free_pages_prepare(page, order, true))
1260 migratetype = get_pfnblock_migratetype(page, pfn);
1261 local_irq_save(flags);
1262 __count_vm_events(PGFREE, 1 << order);
1263 free_one_page(page_zone(page), page, pfn, order, migratetype);
1264 local_irq_restore(flags);
1267 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1269 unsigned int nr_pages = 1 << order;
1270 struct page *p = page;
1274 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1276 __ClearPageReserved(p);
1277 set_page_count(p, 0);
1279 __ClearPageReserved(p);
1280 set_page_count(p, 0);
1282 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1283 set_page_refcounted(page);
1284 __free_pages(page, order);
1287 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1288 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1290 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1292 int __meminit early_pfn_to_nid(unsigned long pfn)
1294 static DEFINE_SPINLOCK(early_pfn_lock);
1297 spin_lock(&early_pfn_lock);
1298 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1300 nid = first_online_node;
1301 spin_unlock(&early_pfn_lock);
1307 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1308 static inline bool __meminit __maybe_unused
1309 meminit_pfn_in_nid(unsigned long pfn, int node,
1310 struct mminit_pfnnid_cache *state)
1314 nid = __early_pfn_to_nid(pfn, state);
1315 if (nid >= 0 && nid != node)
1320 /* Only safe to use early in boot when initialisation is single-threaded */
1321 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1323 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1328 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1332 static inline bool __meminit __maybe_unused
1333 meminit_pfn_in_nid(unsigned long pfn, int node,
1334 struct mminit_pfnnid_cache *state)
1341 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1344 if (early_page_uninitialised(pfn))
1346 return __free_pages_boot_core(page, order);
1350 * Check that the whole (or subset of) a pageblock given by the interval of
1351 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1352 * with the migration of free compaction scanner. The scanners then need to
1353 * use only pfn_valid_within() check for arches that allow holes within
1356 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1358 * It's possible on some configurations to have a setup like node0 node1 node0
1359 * i.e. it's possible that all pages within a zones range of pages do not
1360 * belong to a single zone. We assume that a border between node0 and node1
1361 * can occur within a single pageblock, but not a node0 node1 node0
1362 * interleaving within a single pageblock. It is therefore sufficient to check
1363 * the first and last page of a pageblock and avoid checking each individual
1364 * page in a pageblock.
1366 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1367 unsigned long end_pfn, struct zone *zone)
1369 struct page *start_page;
1370 struct page *end_page;
1372 /* end_pfn is one past the range we are checking */
1375 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1378 start_page = pfn_to_online_page(start_pfn);
1382 if (page_zone(start_page) != zone)
1385 end_page = pfn_to_page(end_pfn);
1387 /* This gives a shorter code than deriving page_zone(end_page) */
1388 if (page_zone_id(start_page) != page_zone_id(end_page))
1394 void set_zone_contiguous(struct zone *zone)
1396 unsigned long block_start_pfn = zone->zone_start_pfn;
1397 unsigned long block_end_pfn;
1399 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1400 for (; block_start_pfn < zone_end_pfn(zone);
1401 block_start_pfn = block_end_pfn,
1402 block_end_pfn += pageblock_nr_pages) {
1404 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1406 if (!__pageblock_pfn_to_page(block_start_pfn,
1407 block_end_pfn, zone))
1411 /* We confirm that there is no hole */
1412 zone->contiguous = true;
1415 void clear_zone_contiguous(struct zone *zone)
1417 zone->contiguous = false;
1420 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1421 static void __init deferred_free_range(unsigned long pfn,
1422 unsigned long nr_pages)
1430 page = pfn_to_page(pfn);
1432 /* Free a large naturally-aligned chunk if possible */
1433 if (nr_pages == pageblock_nr_pages &&
1434 (pfn & (pageblock_nr_pages - 1)) == 0) {
1435 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1436 __free_pages_boot_core(page, pageblock_order);
1440 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1441 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1442 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1443 __free_pages_boot_core(page, 0);
1447 /* Completion tracking for deferred_init_memmap() threads */
1448 static atomic_t pgdat_init_n_undone __initdata;
1449 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1451 static inline void __init pgdat_init_report_one_done(void)
1453 if (atomic_dec_and_test(&pgdat_init_n_undone))
1454 complete(&pgdat_init_all_done_comp);
1458 * Returns true if page needs to be initialized or freed to buddy allocator.
1460 * First we check if pfn is valid on architectures where it is possible to have
1461 * holes within pageblock_nr_pages. On systems where it is not possible, this
1462 * function is optimized out.
1464 * Then, we check if a current large page is valid by only checking the validity
1467 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1468 * within a node: a pfn is between start and end of a node, but does not belong
1469 * to this memory node.
1471 static inline bool __init
1472 deferred_pfn_valid(int nid, unsigned long pfn,
1473 struct mminit_pfnnid_cache *nid_init_state)
1475 if (!pfn_valid_within(pfn))
1477 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1479 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1485 * Free pages to buddy allocator. Try to free aligned pages in
1486 * pageblock_nr_pages sizes.
1488 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1489 unsigned long end_pfn)
1491 struct mminit_pfnnid_cache nid_init_state = { };
1492 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1493 unsigned long nr_free = 0;
1495 for (; pfn < end_pfn; pfn++) {
1496 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1497 deferred_free_range(pfn - nr_free, nr_free);
1499 } else if (!(pfn & nr_pgmask)) {
1500 deferred_free_range(pfn - nr_free, nr_free);
1502 touch_nmi_watchdog();
1507 /* Free the last block of pages to allocator */
1508 deferred_free_range(pfn - nr_free, nr_free);
1512 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1513 * by performing it only once every pageblock_nr_pages.
1514 * Return number of pages initialized.
1516 static unsigned long __init deferred_init_pages(int nid, int zid,
1518 unsigned long end_pfn)
1520 struct mminit_pfnnid_cache nid_init_state = { };
1521 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1522 unsigned long nr_pages = 0;
1523 struct page *page = NULL;
1525 for (; pfn < end_pfn; pfn++) {
1526 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1529 } else if (!page || !(pfn & nr_pgmask)) {
1530 page = pfn_to_page(pfn);
1531 touch_nmi_watchdog();
1535 __init_single_page(page, pfn, zid, nid);
1541 /* Initialise remaining memory on a node */
1542 static int __init deferred_init_memmap(void *data)
1544 pg_data_t *pgdat = data;
1545 int nid = pgdat->node_id;
1546 unsigned long start = jiffies;
1547 unsigned long nr_pages = 0;
1548 unsigned long spfn, epfn, first_init_pfn, flags;
1549 phys_addr_t spa, epa;
1552 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1555 /* Bind memory initialisation thread to a local node if possible */
1556 if (!cpumask_empty(cpumask))
1557 set_cpus_allowed_ptr(current, cpumask);
1559 pgdat_resize_lock(pgdat, &flags);
1560 first_init_pfn = pgdat->first_deferred_pfn;
1561 if (first_init_pfn == ULONG_MAX) {
1562 pgdat_resize_unlock(pgdat, &flags);
1563 pgdat_init_report_one_done();
1567 /* Sanity check boundaries */
1568 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1569 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1570 pgdat->first_deferred_pfn = ULONG_MAX;
1572 /* Only the highest zone is deferred so find it */
1573 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1574 zone = pgdat->node_zones + zid;
1575 if (first_init_pfn < zone_end_pfn(zone))
1578 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1581 * Initialize and free pages. We do it in two loops: first we initialize
1582 * struct page, than free to buddy allocator, because while we are
1583 * freeing pages we can access pages that are ahead (computing buddy
1584 * page in __free_one_page()).
1586 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1587 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1588 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1589 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1591 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1592 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1593 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1594 deferred_free_pages(nid, zid, spfn, epfn);
1596 pgdat_resize_unlock(pgdat, &flags);
1598 /* Sanity check that the next zone really is unpopulated */
1599 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1601 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1602 jiffies_to_msecs(jiffies - start));
1604 pgdat_init_report_one_done();
1609 * During boot we initialize deferred pages on-demand, as needed, but once
1610 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1611 * and we can permanently disable that path.
1613 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1616 * If this zone has deferred pages, try to grow it by initializing enough
1617 * deferred pages to satisfy the allocation specified by order, rounded up to
1618 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1619 * of SECTION_SIZE bytes by initializing struct pages in increments of
1620 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1622 * Return true when zone was grown, otherwise return false. We return true even
1623 * when we grow less than requested, to let the caller decide if there are
1624 * enough pages to satisfy the allocation.
1626 * Note: We use noinline because this function is needed only during boot, and
1627 * it is called from a __ref function _deferred_grow_zone. This way we are
1628 * making sure that it is not inlined into permanent text section.
1630 static noinline bool __init
1631 deferred_grow_zone(struct zone *zone, unsigned int order)
1633 int zid = zone_idx(zone);
1634 int nid = zone_to_nid(zone);
1635 pg_data_t *pgdat = NODE_DATA(nid);
1636 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1637 unsigned long nr_pages = 0;
1638 unsigned long first_init_pfn, spfn, epfn, t, flags;
1639 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1640 phys_addr_t spa, epa;
1643 /* Only the last zone may have deferred pages */
1644 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1647 pgdat_resize_lock(pgdat, &flags);
1650 * If deferred pages have been initialized while we were waiting for
1651 * the lock, return true, as the zone was grown. The caller will retry
1652 * this zone. We won't return to this function since the caller also
1653 * has this static branch.
1655 if (!static_branch_unlikely(&deferred_pages)) {
1656 pgdat_resize_unlock(pgdat, &flags);
1661 * If someone grew this zone while we were waiting for spinlock, return
1662 * true, as there might be enough pages already.
1664 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1665 pgdat_resize_unlock(pgdat, &flags);
1669 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1671 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1672 pgdat_resize_unlock(pgdat, &flags);
1676 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1677 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1678 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1680 while (spfn < epfn && nr_pages < nr_pages_needed) {
1681 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1682 first_deferred_pfn = min(t, epfn);
1683 nr_pages += deferred_init_pages(nid, zid, spfn,
1684 first_deferred_pfn);
1685 spfn = first_deferred_pfn;
1688 if (nr_pages >= nr_pages_needed)
1692 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1693 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1694 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1695 deferred_free_pages(nid, zid, spfn, epfn);
1697 if (first_deferred_pfn == epfn)
1700 pgdat->first_deferred_pfn = first_deferred_pfn;
1701 pgdat_resize_unlock(pgdat, &flags);
1703 return nr_pages > 0;
1707 * deferred_grow_zone() is __init, but it is called from
1708 * get_page_from_freelist() during early boot until deferred_pages permanently
1709 * disables this call. This is why we have refdata wrapper to avoid warning,
1710 * and to ensure that the function body gets unloaded.
1713 _deferred_grow_zone(struct zone *zone, unsigned int order)
1715 return deferred_grow_zone(zone, order);
1718 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1720 void __init page_alloc_init_late(void)
1724 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1727 /* There will be num_node_state(N_MEMORY) threads */
1728 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1729 for_each_node_state(nid, N_MEMORY) {
1730 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1733 /* Block until all are initialised */
1734 wait_for_completion(&pgdat_init_all_done_comp);
1737 * We initialized the rest of the deferred pages. Permanently disable
1738 * on-demand struct page initialization.
1740 static_branch_disable(&deferred_pages);
1742 /* Reinit limits that are based on free pages after the kernel is up */
1743 files_maxfiles_init();
1745 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1746 /* Discard memblock private memory */
1750 for_each_populated_zone(zone)
1751 set_zone_contiguous(zone);
1755 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1756 void __init init_cma_reserved_pageblock(struct page *page)
1758 unsigned i = pageblock_nr_pages;
1759 struct page *p = page;
1762 __ClearPageReserved(p);
1763 set_page_count(p, 0);
1766 set_pageblock_migratetype(page, MIGRATE_CMA);
1768 if (pageblock_order >= MAX_ORDER) {
1769 i = pageblock_nr_pages;
1772 set_page_refcounted(p);
1773 __free_pages(p, MAX_ORDER - 1);
1774 p += MAX_ORDER_NR_PAGES;
1775 } while (i -= MAX_ORDER_NR_PAGES);
1777 set_page_refcounted(page);
1778 __free_pages(page, pageblock_order);
1781 adjust_managed_page_count(page, pageblock_nr_pages);
1786 * The order of subdivision here is critical for the IO subsystem.
1787 * Please do not alter this order without good reasons and regression
1788 * testing. Specifically, as large blocks of memory are subdivided,
1789 * the order in which smaller blocks are delivered depends on the order
1790 * they're subdivided in this function. This is the primary factor
1791 * influencing the order in which pages are delivered to the IO
1792 * subsystem according to empirical testing, and this is also justified
1793 * by considering the behavior of a buddy system containing a single
1794 * large block of memory acted on by a series of small allocations.
1795 * This behavior is a critical factor in sglist merging's success.
1799 static inline void expand(struct zone *zone, struct page *page,
1800 int low, int high, struct free_area *area,
1803 unsigned long size = 1 << high;
1805 while (high > low) {
1809 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1812 * Mark as guard pages (or page), that will allow to
1813 * merge back to allocator when buddy will be freed.
1814 * Corresponding page table entries will not be touched,
1815 * pages will stay not present in virtual address space
1817 if (set_page_guard(zone, &page[size], high, migratetype))
1820 list_add(&page[size].lru, &area->free_list[migratetype]);
1822 set_page_order(&page[size], high);
1826 static void check_new_page_bad(struct page *page)
1828 const char *bad_reason = NULL;
1829 unsigned long bad_flags = 0;
1831 if (unlikely(atomic_read(&page->_mapcount) != -1))
1832 bad_reason = "nonzero mapcount";
1833 if (unlikely(page->mapping != NULL))
1834 bad_reason = "non-NULL mapping";
1835 if (unlikely(page_ref_count(page) != 0))
1836 bad_reason = "nonzero _count";
1837 if (unlikely(page->flags & __PG_HWPOISON)) {
1838 bad_reason = "HWPoisoned (hardware-corrupted)";
1839 bad_flags = __PG_HWPOISON;
1840 /* Don't complain about hwpoisoned pages */
1841 page_mapcount_reset(page); /* remove PageBuddy */
1844 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1845 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1846 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1849 if (unlikely(page->mem_cgroup))
1850 bad_reason = "page still charged to cgroup";
1852 bad_page(page, bad_reason, bad_flags);
1856 * This page is about to be returned from the page allocator
1858 static inline int check_new_page(struct page *page)
1860 if (likely(page_expected_state(page,
1861 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1864 check_new_page_bad(page);
1868 static inline bool free_pages_prezeroed(void)
1870 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1871 page_poisoning_enabled();
1874 #ifdef CONFIG_DEBUG_VM
1875 static bool check_pcp_refill(struct page *page)
1880 static bool check_new_pcp(struct page *page)
1882 return check_new_page(page);
1885 static bool check_pcp_refill(struct page *page)
1887 return check_new_page(page);
1889 static bool check_new_pcp(struct page *page)
1893 #endif /* CONFIG_DEBUG_VM */
1895 static bool check_new_pages(struct page *page, unsigned int order)
1898 for (i = 0; i < (1 << order); i++) {
1899 struct page *p = page + i;
1901 if (unlikely(check_new_page(p)))
1908 inline void post_alloc_hook(struct page *page, unsigned int order,
1911 set_page_private(page, 0);
1912 set_page_refcounted(page);
1914 arch_alloc_page(page, order);
1915 kernel_map_pages(page, 1 << order, 1);
1916 kernel_poison_pages(page, 1 << order, 1);
1917 kasan_alloc_pages(page, order);
1918 set_page_owner(page, order, gfp_flags);
1921 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1922 unsigned int alloc_flags)
1926 post_alloc_hook(page, order, gfp_flags);
1928 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1929 for (i = 0; i < (1 << order); i++)
1930 clear_highpage(page + i);
1932 if (order && (gfp_flags & __GFP_COMP))
1933 prep_compound_page(page, order);
1936 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1937 * allocate the page. The expectation is that the caller is taking
1938 * steps that will free more memory. The caller should avoid the page
1939 * being used for !PFMEMALLOC purposes.
1941 if (alloc_flags & ALLOC_NO_WATERMARKS)
1942 set_page_pfmemalloc(page);
1944 clear_page_pfmemalloc(page);
1948 * Go through the free lists for the given migratetype and remove
1949 * the smallest available page from the freelists
1951 static __always_inline
1952 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1955 unsigned int current_order;
1956 struct free_area *area;
1959 /* Find a page of the appropriate size in the preferred list */
1960 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1961 area = &(zone->free_area[current_order]);
1962 page = list_first_entry_or_null(&area->free_list[migratetype],
1966 list_del(&page->lru);
1967 rmv_page_order(page);
1969 expand(zone, page, order, current_order, area, migratetype);
1970 set_pcppage_migratetype(page, migratetype);
1979 * This array describes the order lists are fallen back to when
1980 * the free lists for the desirable migrate type are depleted
1982 static int fallbacks[MIGRATE_TYPES][4] = {
1983 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1984 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1985 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1987 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1989 #ifdef CONFIG_MEMORY_ISOLATION
1990 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1995 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1998 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2001 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2002 unsigned int order) { return NULL; }
2006 * Move the free pages in a range to the free lists of the requested type.
2007 * Note that start_page and end_pages are not aligned on a pageblock
2008 * boundary. If alignment is required, use move_freepages_block()
2010 static int move_freepages(struct zone *zone,
2011 struct page *start_page, struct page *end_page,
2012 int migratetype, int *num_movable)
2016 int pages_moved = 0;
2018 #ifndef CONFIG_HOLES_IN_ZONE
2020 * page_zone is not safe to call in this context when
2021 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2022 * anyway as we check zone boundaries in move_freepages_block().
2023 * Remove at a later date when no bug reports exist related to
2024 * grouping pages by mobility
2026 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2027 pfn_valid(page_to_pfn(end_page)) &&
2028 page_zone(start_page) != page_zone(end_page));
2030 for (page = start_page; page <= end_page;) {
2031 if (!pfn_valid_within(page_to_pfn(page))) {
2036 /* Make sure we are not inadvertently changing nodes */
2037 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2039 if (!PageBuddy(page)) {
2041 * We assume that pages that could be isolated for
2042 * migration are movable. But we don't actually try
2043 * isolating, as that would be expensive.
2046 (PageLRU(page) || __PageMovable(page)))
2053 order = page_order(page);
2054 list_move(&page->lru,
2055 &zone->free_area[order].free_list[migratetype]);
2057 pages_moved += 1 << order;
2063 int move_freepages_block(struct zone *zone, struct page *page,
2064 int migratetype, int *num_movable)
2066 unsigned long start_pfn, end_pfn;
2067 struct page *start_page, *end_page;
2072 start_pfn = page_to_pfn(page);
2073 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2074 start_page = pfn_to_page(start_pfn);
2075 end_page = start_page + pageblock_nr_pages - 1;
2076 end_pfn = start_pfn + pageblock_nr_pages - 1;
2078 /* Do not cross zone boundaries */
2079 if (!zone_spans_pfn(zone, start_pfn))
2081 if (!zone_spans_pfn(zone, end_pfn))
2084 return move_freepages(zone, start_page, end_page, migratetype,
2088 static void change_pageblock_range(struct page *pageblock_page,
2089 int start_order, int migratetype)
2091 int nr_pageblocks = 1 << (start_order - pageblock_order);
2093 while (nr_pageblocks--) {
2094 set_pageblock_migratetype(pageblock_page, migratetype);
2095 pageblock_page += pageblock_nr_pages;
2100 * When we are falling back to another migratetype during allocation, try to
2101 * steal extra free pages from the same pageblocks to satisfy further
2102 * allocations, instead of polluting multiple pageblocks.
2104 * If we are stealing a relatively large buddy page, it is likely there will
2105 * be more free pages in the pageblock, so try to steal them all. For
2106 * reclaimable and unmovable allocations, we steal regardless of page size,
2107 * as fragmentation caused by those allocations polluting movable pageblocks
2108 * is worse than movable allocations stealing from unmovable and reclaimable
2111 static bool can_steal_fallback(unsigned int order, int start_mt)
2114 * Leaving this order check is intended, although there is
2115 * relaxed order check in next check. The reason is that
2116 * we can actually steal whole pageblock if this condition met,
2117 * but, below check doesn't guarantee it and that is just heuristic
2118 * so could be changed anytime.
2120 if (order >= pageblock_order)
2123 if (order >= pageblock_order / 2 ||
2124 start_mt == MIGRATE_RECLAIMABLE ||
2125 start_mt == MIGRATE_UNMOVABLE ||
2126 page_group_by_mobility_disabled)
2133 * This function implements actual steal behaviour. If order is large enough,
2134 * we can steal whole pageblock. If not, we first move freepages in this
2135 * pageblock to our migratetype and determine how many already-allocated pages
2136 * are there in the pageblock with a compatible migratetype. If at least half
2137 * of pages are free or compatible, we can change migratetype of the pageblock
2138 * itself, so pages freed in the future will be put on the correct free list.
2140 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2141 int start_type, bool whole_block)
2143 unsigned int current_order = page_order(page);
2144 struct free_area *area;
2145 int free_pages, movable_pages, alike_pages;
2148 old_block_type = get_pageblock_migratetype(page);
2151 * This can happen due to races and we want to prevent broken
2152 * highatomic accounting.
2154 if (is_migrate_highatomic(old_block_type))
2157 /* Take ownership for orders >= pageblock_order */
2158 if (current_order >= pageblock_order) {
2159 change_pageblock_range(page, current_order, start_type);
2163 /* We are not allowed to try stealing from the whole block */
2167 free_pages = move_freepages_block(zone, page, start_type,
2170 * Determine how many pages are compatible with our allocation.
2171 * For movable allocation, it's the number of movable pages which
2172 * we just obtained. For other types it's a bit more tricky.
2174 if (start_type == MIGRATE_MOVABLE) {
2175 alike_pages = movable_pages;
2178 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2179 * to MOVABLE pageblock, consider all non-movable pages as
2180 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2181 * vice versa, be conservative since we can't distinguish the
2182 * exact migratetype of non-movable pages.
2184 if (old_block_type == MIGRATE_MOVABLE)
2185 alike_pages = pageblock_nr_pages
2186 - (free_pages + movable_pages);
2191 /* moving whole block can fail due to zone boundary conditions */
2196 * If a sufficient number of pages in the block are either free or of
2197 * comparable migratability as our allocation, claim the whole block.
2199 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2200 page_group_by_mobility_disabled)
2201 set_pageblock_migratetype(page, start_type);
2206 area = &zone->free_area[current_order];
2207 list_move(&page->lru, &area->free_list[start_type]);
2211 * Check whether there is a suitable fallback freepage with requested order.
2212 * If only_stealable is true, this function returns fallback_mt only if
2213 * we can steal other freepages all together. This would help to reduce
2214 * fragmentation due to mixed migratetype pages in one pageblock.
2216 int find_suitable_fallback(struct free_area *area, unsigned int order,
2217 int migratetype, bool only_stealable, bool *can_steal)
2222 if (area->nr_free == 0)
2227 fallback_mt = fallbacks[migratetype][i];
2228 if (fallback_mt == MIGRATE_TYPES)
2231 if (list_empty(&area->free_list[fallback_mt]))
2234 if (can_steal_fallback(order, migratetype))
2237 if (!only_stealable)
2248 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2249 * there are no empty page blocks that contain a page with a suitable order
2251 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2252 unsigned int alloc_order)
2255 unsigned long max_managed, flags;
2258 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2259 * Check is race-prone but harmless.
2261 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2262 if (zone->nr_reserved_highatomic >= max_managed)
2265 spin_lock_irqsave(&zone->lock, flags);
2267 /* Recheck the nr_reserved_highatomic limit under the lock */
2268 if (zone->nr_reserved_highatomic >= max_managed)
2272 mt = get_pageblock_migratetype(page);
2273 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2274 && !is_migrate_cma(mt)) {
2275 zone->nr_reserved_highatomic += pageblock_nr_pages;
2276 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2277 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2281 spin_unlock_irqrestore(&zone->lock, flags);
2285 * Used when an allocation is about to fail under memory pressure. This
2286 * potentially hurts the reliability of high-order allocations when under
2287 * intense memory pressure but failed atomic allocations should be easier
2288 * to recover from than an OOM.
2290 * If @force is true, try to unreserve a pageblock even though highatomic
2291 * pageblock is exhausted.
2293 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2296 struct zonelist *zonelist = ac->zonelist;
2297 unsigned long flags;
2304 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2307 * Preserve at least one pageblock unless memory pressure
2310 if (!force && zone->nr_reserved_highatomic <=
2314 spin_lock_irqsave(&zone->lock, flags);
2315 for (order = 0; order < MAX_ORDER; order++) {
2316 struct free_area *area = &(zone->free_area[order]);
2318 page = list_first_entry_or_null(
2319 &area->free_list[MIGRATE_HIGHATOMIC],
2325 * In page freeing path, migratetype change is racy so
2326 * we can counter several free pages in a pageblock
2327 * in this loop althoug we changed the pageblock type
2328 * from highatomic to ac->migratetype. So we should
2329 * adjust the count once.
2331 if (is_migrate_highatomic_page(page)) {
2333 * It should never happen but changes to
2334 * locking could inadvertently allow a per-cpu
2335 * drain to add pages to MIGRATE_HIGHATOMIC
2336 * while unreserving so be safe and watch for
2339 zone->nr_reserved_highatomic -= min(
2341 zone->nr_reserved_highatomic);
2345 * Convert to ac->migratetype and avoid the normal
2346 * pageblock stealing heuristics. Minimally, the caller
2347 * is doing the work and needs the pages. More
2348 * importantly, if the block was always converted to
2349 * MIGRATE_UNMOVABLE or another type then the number
2350 * of pageblocks that cannot be completely freed
2353 set_pageblock_migratetype(page, ac->migratetype);
2354 ret = move_freepages_block(zone, page, ac->migratetype,
2357 spin_unlock_irqrestore(&zone->lock, flags);
2361 spin_unlock_irqrestore(&zone->lock, flags);
2368 * Try finding a free buddy page on the fallback list and put it on the free
2369 * list of requested migratetype, possibly along with other pages from the same
2370 * block, depending on fragmentation avoidance heuristics. Returns true if
2371 * fallback was found so that __rmqueue_smallest() can grab it.
2373 * The use of signed ints for order and current_order is a deliberate
2374 * deviation from the rest of this file, to make the for loop
2375 * condition simpler.
2377 static __always_inline bool
2378 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2380 struct free_area *area;
2387 * Find the largest available free page in the other list. This roughly
2388 * approximates finding the pageblock with the most free pages, which
2389 * would be too costly to do exactly.
2391 for (current_order = MAX_ORDER - 1; current_order >= order;
2393 area = &(zone->free_area[current_order]);
2394 fallback_mt = find_suitable_fallback(area, current_order,
2395 start_migratetype, false, &can_steal);
2396 if (fallback_mt == -1)
2400 * We cannot steal all free pages from the pageblock and the
2401 * requested migratetype is movable. In that case it's better to
2402 * steal and split the smallest available page instead of the
2403 * largest available page, because even if the next movable
2404 * allocation falls back into a different pageblock than this
2405 * one, it won't cause permanent fragmentation.
2407 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2408 && current_order > order)
2417 for (current_order = order; current_order < MAX_ORDER;
2419 area = &(zone->free_area[current_order]);
2420 fallback_mt = find_suitable_fallback(area, current_order,
2421 start_migratetype, false, &can_steal);
2422 if (fallback_mt != -1)
2427 * This should not happen - we already found a suitable fallback
2428 * when looking for the largest page.
2430 VM_BUG_ON(current_order == MAX_ORDER);
2433 page = list_first_entry(&area->free_list[fallback_mt],
2436 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2438 trace_mm_page_alloc_extfrag(page, order, current_order,
2439 start_migratetype, fallback_mt);
2446 * Do the hard work of removing an element from the buddy allocator.
2447 * Call me with the zone->lock already held.
2449 static __always_inline struct page *
2450 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2455 page = __rmqueue_smallest(zone, order, migratetype);
2456 if (unlikely(!page)) {
2457 if (migratetype == MIGRATE_MOVABLE)
2458 page = __rmqueue_cma_fallback(zone, order);
2460 if (!page && __rmqueue_fallback(zone, order, migratetype))
2464 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2469 * Obtain a specified number of elements from the buddy allocator, all under
2470 * a single hold of the lock, for efficiency. Add them to the supplied list.
2471 * Returns the number of new pages which were placed at *list.
2473 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2474 unsigned long count, struct list_head *list,
2479 spin_lock(&zone->lock);
2480 for (i = 0; i < count; ++i) {
2481 struct page *page = __rmqueue(zone, order, migratetype);
2482 if (unlikely(page == NULL))
2485 if (unlikely(check_pcp_refill(page)))
2489 * Split buddy pages returned by expand() are received here in
2490 * physical page order. The page is added to the tail of
2491 * caller's list. From the callers perspective, the linked list
2492 * is ordered by page number under some conditions. This is
2493 * useful for IO devices that can forward direction from the
2494 * head, thus also in the physical page order. This is useful
2495 * for IO devices that can merge IO requests if the physical
2496 * pages are ordered properly.
2498 list_add_tail(&page->lru, list);
2500 if (is_migrate_cma(get_pcppage_migratetype(page)))
2501 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2506 * i pages were removed from the buddy list even if some leak due
2507 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2508 * on i. Do not confuse with 'alloced' which is the number of
2509 * pages added to the pcp list.
2511 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2512 spin_unlock(&zone->lock);
2518 * Called from the vmstat counter updater to drain pagesets of this
2519 * currently executing processor on remote nodes after they have
2522 * Note that this function must be called with the thread pinned to
2523 * a single processor.
2525 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2527 unsigned long flags;
2528 int to_drain, batch;
2530 local_irq_save(flags);
2531 batch = READ_ONCE(pcp->batch);
2532 to_drain = min(pcp->count, batch);
2534 free_pcppages_bulk(zone, to_drain, pcp);
2535 local_irq_restore(flags);
2540 * Drain pcplists of the indicated processor and zone.
2542 * The processor must either be the current processor and the
2543 * thread pinned to the current processor or a processor that
2546 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2548 unsigned long flags;
2549 struct per_cpu_pageset *pset;
2550 struct per_cpu_pages *pcp;
2552 local_irq_save(flags);
2553 pset = per_cpu_ptr(zone->pageset, cpu);
2557 free_pcppages_bulk(zone, pcp->count, pcp);
2558 local_irq_restore(flags);
2562 * Drain pcplists of all zones on the indicated processor.
2564 * The processor must either be the current processor and the
2565 * thread pinned to the current processor or a processor that
2568 static void drain_pages(unsigned int cpu)
2572 for_each_populated_zone(zone) {
2573 drain_pages_zone(cpu, zone);
2578 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2580 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2581 * the single zone's pages.
2583 void drain_local_pages(struct zone *zone)
2585 int cpu = smp_processor_id();
2588 drain_pages_zone(cpu, zone);
2593 static void drain_local_pages_wq(struct work_struct *work)
2596 * drain_all_pages doesn't use proper cpu hotplug protection so
2597 * we can race with cpu offline when the WQ can move this from
2598 * a cpu pinned worker to an unbound one. We can operate on a different
2599 * cpu which is allright but we also have to make sure to not move to
2603 drain_local_pages(NULL);
2608 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2610 * When zone parameter is non-NULL, spill just the single zone's pages.
2612 * Note that this can be extremely slow as the draining happens in a workqueue.
2614 void drain_all_pages(struct zone *zone)
2619 * Allocate in the BSS so we wont require allocation in
2620 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2622 static cpumask_t cpus_with_pcps;
2625 * Make sure nobody triggers this path before mm_percpu_wq is fully
2628 if (WARN_ON_ONCE(!mm_percpu_wq))
2632 * Do not drain if one is already in progress unless it's specific to
2633 * a zone. Such callers are primarily CMA and memory hotplug and need
2634 * the drain to be complete when the call returns.
2636 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2639 mutex_lock(&pcpu_drain_mutex);
2643 * We don't care about racing with CPU hotplug event
2644 * as offline notification will cause the notified
2645 * cpu to drain that CPU pcps and on_each_cpu_mask
2646 * disables preemption as part of its processing
2648 for_each_online_cpu(cpu) {
2649 struct per_cpu_pageset *pcp;
2651 bool has_pcps = false;
2654 pcp = per_cpu_ptr(zone->pageset, cpu);
2658 for_each_populated_zone(z) {
2659 pcp = per_cpu_ptr(z->pageset, cpu);
2660 if (pcp->pcp.count) {
2668 cpumask_set_cpu(cpu, &cpus_with_pcps);
2670 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2673 for_each_cpu(cpu, &cpus_with_pcps) {
2674 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2675 INIT_WORK(work, drain_local_pages_wq);
2676 queue_work_on(cpu, mm_percpu_wq, work);
2678 for_each_cpu(cpu, &cpus_with_pcps)
2679 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2681 mutex_unlock(&pcpu_drain_mutex);
2684 #ifdef CONFIG_HIBERNATION
2687 * Touch the watchdog for every WD_PAGE_COUNT pages.
2689 #define WD_PAGE_COUNT (128*1024)
2691 void mark_free_pages(struct zone *zone)
2693 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2694 unsigned long flags;
2695 unsigned int order, t;
2698 if (zone_is_empty(zone))
2701 spin_lock_irqsave(&zone->lock, flags);
2703 max_zone_pfn = zone_end_pfn(zone);
2704 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2705 if (pfn_valid(pfn)) {
2706 page = pfn_to_page(pfn);
2708 if (!--page_count) {
2709 touch_nmi_watchdog();
2710 page_count = WD_PAGE_COUNT;
2713 if (page_zone(page) != zone)
2716 if (!swsusp_page_is_forbidden(page))
2717 swsusp_unset_page_free(page);
2720 for_each_migratetype_order(order, t) {
2721 list_for_each_entry(page,
2722 &zone->free_area[order].free_list[t], lru) {
2725 pfn = page_to_pfn(page);
2726 for (i = 0; i < (1UL << order); i++) {
2727 if (!--page_count) {
2728 touch_nmi_watchdog();
2729 page_count = WD_PAGE_COUNT;
2731 swsusp_set_page_free(pfn_to_page(pfn + i));
2735 spin_unlock_irqrestore(&zone->lock, flags);
2737 #endif /* CONFIG_PM */
2739 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2743 if (!free_pcp_prepare(page))
2746 migratetype = get_pfnblock_migratetype(page, pfn);
2747 set_pcppage_migratetype(page, migratetype);
2751 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2753 struct zone *zone = page_zone(page);
2754 struct per_cpu_pages *pcp;
2757 migratetype = get_pcppage_migratetype(page);
2758 __count_vm_event(PGFREE);
2761 * We only track unmovable, reclaimable and movable on pcp lists.
2762 * Free ISOLATE pages back to the allocator because they are being
2763 * offlined but treat HIGHATOMIC as movable pages so we can get those
2764 * areas back if necessary. Otherwise, we may have to free
2765 * excessively into the page allocator
2767 if (migratetype >= MIGRATE_PCPTYPES) {
2768 if (unlikely(is_migrate_isolate(migratetype))) {
2769 free_one_page(zone, page, pfn, 0, migratetype);
2772 migratetype = MIGRATE_MOVABLE;
2775 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2776 list_add(&page->lru, &pcp->lists[migratetype]);
2778 if (pcp->count >= pcp->high) {
2779 unsigned long batch = READ_ONCE(pcp->batch);
2780 free_pcppages_bulk(zone, batch, pcp);
2785 * Free a 0-order page
2787 void free_unref_page(struct page *page)
2789 unsigned long flags;
2790 unsigned long pfn = page_to_pfn(page);
2792 if (!free_unref_page_prepare(page, pfn))
2795 local_irq_save(flags);
2796 free_unref_page_commit(page, pfn);
2797 local_irq_restore(flags);
2801 * Free a list of 0-order pages
2803 void free_unref_page_list(struct list_head *list)
2805 struct page *page, *next;
2806 unsigned long flags, pfn;
2807 int batch_count = 0;
2809 /* Prepare pages for freeing */
2810 list_for_each_entry_safe(page, next, list, lru) {
2811 pfn = page_to_pfn(page);
2812 if (!free_unref_page_prepare(page, pfn))
2813 list_del(&page->lru);
2814 set_page_private(page, pfn);
2817 local_irq_save(flags);
2818 list_for_each_entry_safe(page, next, list, lru) {
2819 unsigned long pfn = page_private(page);
2821 set_page_private(page, 0);
2822 trace_mm_page_free_batched(page);
2823 free_unref_page_commit(page, pfn);
2826 * Guard against excessive IRQ disabled times when we get
2827 * a large list of pages to free.
2829 if (++batch_count == SWAP_CLUSTER_MAX) {
2830 local_irq_restore(flags);
2832 local_irq_save(flags);
2835 local_irq_restore(flags);
2839 * split_page takes a non-compound higher-order page, and splits it into
2840 * n (1<<order) sub-pages: page[0..n]
2841 * Each sub-page must be freed individually.
2843 * Note: this is probably too low level an operation for use in drivers.
2844 * Please consult with lkml before using this in your driver.
2846 void split_page(struct page *page, unsigned int order)
2850 VM_BUG_ON_PAGE(PageCompound(page), page);
2851 VM_BUG_ON_PAGE(!page_count(page), page);
2853 for (i = 1; i < (1 << order); i++)
2854 set_page_refcounted(page + i);
2855 split_page_owner(page, order);
2857 EXPORT_SYMBOL_GPL(split_page);
2859 int __isolate_free_page(struct page *page, unsigned int order)
2861 unsigned long watermark;
2865 BUG_ON(!PageBuddy(page));
2867 zone = page_zone(page);
2868 mt = get_pageblock_migratetype(page);
2870 if (!is_migrate_isolate(mt)) {
2872 * Obey watermarks as if the page was being allocated. We can
2873 * emulate a high-order watermark check with a raised order-0
2874 * watermark, because we already know our high-order page
2877 watermark = min_wmark_pages(zone) + (1UL << order);
2878 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2881 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2884 /* Remove page from free list */
2885 list_del(&page->lru);
2886 zone->free_area[order].nr_free--;
2887 rmv_page_order(page);
2890 * Set the pageblock if the isolated page is at least half of a
2893 if (order >= pageblock_order - 1) {
2894 struct page *endpage = page + (1 << order) - 1;
2895 for (; page < endpage; page += pageblock_nr_pages) {
2896 int mt = get_pageblock_migratetype(page);
2897 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2898 && !is_migrate_highatomic(mt))
2899 set_pageblock_migratetype(page,
2905 return 1UL << order;
2909 * Update NUMA hit/miss statistics
2911 * Must be called with interrupts disabled.
2913 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2916 enum numa_stat_item local_stat = NUMA_LOCAL;
2918 /* skip numa counters update if numa stats is disabled */
2919 if (!static_branch_likely(&vm_numa_stat_key))
2922 if (zone_to_nid(z) != numa_node_id())
2923 local_stat = NUMA_OTHER;
2925 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2926 __inc_numa_state(z, NUMA_HIT);
2928 __inc_numa_state(z, NUMA_MISS);
2929 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2931 __inc_numa_state(z, local_stat);
2935 /* Remove page from the per-cpu list, caller must protect the list */
2936 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2937 struct per_cpu_pages *pcp,
2938 struct list_head *list)
2943 if (list_empty(list)) {
2944 pcp->count += rmqueue_bulk(zone, 0,
2947 if (unlikely(list_empty(list)))
2951 page = list_first_entry(list, struct page, lru);
2952 list_del(&page->lru);
2954 } while (check_new_pcp(page));
2959 /* Lock and remove page from the per-cpu list */
2960 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2961 struct zone *zone, unsigned int order,
2962 gfp_t gfp_flags, int migratetype)
2964 struct per_cpu_pages *pcp;
2965 struct list_head *list;
2967 unsigned long flags;
2969 local_irq_save(flags);
2970 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2971 list = &pcp->lists[migratetype];
2972 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2974 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2975 zone_statistics(preferred_zone, zone);
2977 local_irq_restore(flags);
2982 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2985 struct page *rmqueue(struct zone *preferred_zone,
2986 struct zone *zone, unsigned int order,
2987 gfp_t gfp_flags, unsigned int alloc_flags,
2990 unsigned long flags;
2993 if (likely(order == 0)) {
2994 page = rmqueue_pcplist(preferred_zone, zone, order,
2995 gfp_flags, migratetype);
3000 * We most definitely don't want callers attempting to
3001 * allocate greater than order-1 page units with __GFP_NOFAIL.
3003 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3004 spin_lock_irqsave(&zone->lock, flags);
3008 if (alloc_flags & ALLOC_HARDER) {
3009 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3011 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3014 page = __rmqueue(zone, order, migratetype);
3015 } while (page && check_new_pages(page, order));
3016 spin_unlock(&zone->lock);
3019 __mod_zone_freepage_state(zone, -(1 << order),
3020 get_pcppage_migratetype(page));
3022 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3023 zone_statistics(preferred_zone, zone);
3024 local_irq_restore(flags);
3027 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3031 local_irq_restore(flags);
3035 #ifdef CONFIG_FAIL_PAGE_ALLOC
3038 struct fault_attr attr;
3040 bool ignore_gfp_highmem;
3041 bool ignore_gfp_reclaim;
3043 } fail_page_alloc = {
3044 .attr = FAULT_ATTR_INITIALIZER,
3045 .ignore_gfp_reclaim = true,
3046 .ignore_gfp_highmem = true,
3050 static int __init setup_fail_page_alloc(char *str)
3052 return setup_fault_attr(&fail_page_alloc.attr, str);
3054 __setup("fail_page_alloc=", setup_fail_page_alloc);
3056 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3058 if (order < fail_page_alloc.min_order)
3060 if (gfp_mask & __GFP_NOFAIL)
3062 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3064 if (fail_page_alloc.ignore_gfp_reclaim &&
3065 (gfp_mask & __GFP_DIRECT_RECLAIM))
3068 return should_fail(&fail_page_alloc.attr, 1 << order);
3071 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3073 static int __init fail_page_alloc_debugfs(void)
3075 umode_t mode = S_IFREG | 0600;
3078 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3079 &fail_page_alloc.attr);
3081 return PTR_ERR(dir);
3083 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3084 &fail_page_alloc.ignore_gfp_reclaim))
3086 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3087 &fail_page_alloc.ignore_gfp_highmem))
3089 if (!debugfs_create_u32("min-order", mode, dir,
3090 &fail_page_alloc.min_order))
3095 debugfs_remove_recursive(dir);
3100 late_initcall(fail_page_alloc_debugfs);
3102 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3104 #else /* CONFIG_FAIL_PAGE_ALLOC */
3106 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3111 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3114 * Return true if free base pages are above 'mark'. For high-order checks it
3115 * will return true of the order-0 watermark is reached and there is at least
3116 * one free page of a suitable size. Checking now avoids taking the zone lock
3117 * to check in the allocation paths if no pages are free.
3119 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3120 int classzone_idx, unsigned int alloc_flags,
3125 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3127 /* free_pages may go negative - that's OK */
3128 free_pages -= (1 << order) - 1;
3130 if (alloc_flags & ALLOC_HIGH)
3134 * If the caller does not have rights to ALLOC_HARDER then subtract
3135 * the high-atomic reserves. This will over-estimate the size of the
3136 * atomic reserve but it avoids a search.
3138 if (likely(!alloc_harder)) {
3139 free_pages -= z->nr_reserved_highatomic;
3142 * OOM victims can try even harder than normal ALLOC_HARDER
3143 * users on the grounds that it's definitely going to be in
3144 * the exit path shortly and free memory. Any allocation it
3145 * makes during the free path will be small and short-lived.
3147 if (alloc_flags & ALLOC_OOM)
3155 /* If allocation can't use CMA areas don't use free CMA pages */
3156 if (!(alloc_flags & ALLOC_CMA))
3157 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3161 * Check watermarks for an order-0 allocation request. If these
3162 * are not met, then a high-order request also cannot go ahead
3163 * even if a suitable page happened to be free.
3165 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3168 /* If this is an order-0 request then the watermark is fine */
3172 /* For a high-order request, check at least one suitable page is free */
3173 for (o = order; o < MAX_ORDER; o++) {
3174 struct free_area *area = &z->free_area[o];
3180 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3181 if (!list_empty(&area->free_list[mt]))
3186 if ((alloc_flags & ALLOC_CMA) &&
3187 !list_empty(&area->free_list[MIGRATE_CMA])) {
3192 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3198 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3199 int classzone_idx, unsigned int alloc_flags)
3201 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3202 zone_page_state(z, NR_FREE_PAGES));
3205 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3206 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3208 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3212 /* If allocation can't use CMA areas don't use free CMA pages */
3213 if (!(alloc_flags & ALLOC_CMA))
3214 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3218 * Fast check for order-0 only. If this fails then the reserves
3219 * need to be calculated. There is a corner case where the check
3220 * passes but only the high-order atomic reserve are free. If
3221 * the caller is !atomic then it'll uselessly search the free
3222 * list. That corner case is then slower but it is harmless.
3224 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3227 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3231 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3232 unsigned long mark, int classzone_idx)
3234 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3236 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3237 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3239 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3244 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3246 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3249 #else /* CONFIG_NUMA */
3250 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3254 #endif /* CONFIG_NUMA */
3257 * get_page_from_freelist goes through the zonelist trying to allocate
3260 static struct page *
3261 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3262 const struct alloc_context *ac)
3264 struct zoneref *z = ac->preferred_zoneref;
3266 struct pglist_data *last_pgdat_dirty_limit = NULL;
3269 * Scan zonelist, looking for a zone with enough free.
3270 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3272 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3277 if (cpusets_enabled() &&
3278 (alloc_flags & ALLOC_CPUSET) &&
3279 !__cpuset_zone_allowed(zone, gfp_mask))
3282 * When allocating a page cache page for writing, we
3283 * want to get it from a node that is within its dirty
3284 * limit, such that no single node holds more than its
3285 * proportional share of globally allowed dirty pages.
3286 * The dirty limits take into account the node's
3287 * lowmem reserves and high watermark so that kswapd
3288 * should be able to balance it without having to
3289 * write pages from its LRU list.
3291 * XXX: For now, allow allocations to potentially
3292 * exceed the per-node dirty limit in the slowpath
3293 * (spread_dirty_pages unset) before going into reclaim,
3294 * which is important when on a NUMA setup the allowed
3295 * nodes are together not big enough to reach the
3296 * global limit. The proper fix for these situations
3297 * will require awareness of nodes in the
3298 * dirty-throttling and the flusher threads.
3300 if (ac->spread_dirty_pages) {
3301 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3304 if (!node_dirty_ok(zone->zone_pgdat)) {
3305 last_pgdat_dirty_limit = zone->zone_pgdat;
3310 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3311 if (!zone_watermark_fast(zone, order, mark,
3312 ac_classzone_idx(ac), alloc_flags)) {
3315 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3317 * Watermark failed for this zone, but see if we can
3318 * grow this zone if it contains deferred pages.
3320 if (static_branch_unlikely(&deferred_pages)) {
3321 if (_deferred_grow_zone(zone, order))
3325 /* Checked here to keep the fast path fast */
3326 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3327 if (alloc_flags & ALLOC_NO_WATERMARKS)
3330 if (node_reclaim_mode == 0 ||
3331 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3334 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3336 case NODE_RECLAIM_NOSCAN:
3339 case NODE_RECLAIM_FULL:
3340 /* scanned but unreclaimable */
3343 /* did we reclaim enough */
3344 if (zone_watermark_ok(zone, order, mark,
3345 ac_classzone_idx(ac), alloc_flags))
3353 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3354 gfp_mask, alloc_flags, ac->migratetype);
3356 prep_new_page(page, order, gfp_mask, alloc_flags);
3359 * If this is a high-order atomic allocation then check
3360 * if the pageblock should be reserved for the future
3362 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3363 reserve_highatomic_pageblock(page, zone, order);
3367 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3368 /* Try again if zone has deferred pages */
3369 if (static_branch_unlikely(&deferred_pages)) {
3370 if (_deferred_grow_zone(zone, order))
3380 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3382 unsigned int filter = SHOW_MEM_FILTER_NODES;
3383 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3385 if (!__ratelimit(&show_mem_rs))
3389 * This documents exceptions given to allocations in certain
3390 * contexts that are allowed to allocate outside current's set
3393 if (!(gfp_mask & __GFP_NOMEMALLOC))
3394 if (tsk_is_oom_victim(current) ||
3395 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3396 filter &= ~SHOW_MEM_FILTER_NODES;
3397 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3398 filter &= ~SHOW_MEM_FILTER_NODES;
3400 show_mem(filter, nodemask);
3403 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3405 struct va_format vaf;
3407 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3408 DEFAULT_RATELIMIT_BURST);
3410 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3413 va_start(args, fmt);
3416 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3417 current->comm, &vaf, gfp_mask, &gfp_mask,
3418 nodemask_pr_args(nodemask));
3421 cpuset_print_current_mems_allowed();
3424 warn_alloc_show_mem(gfp_mask, nodemask);
3427 static inline struct page *
3428 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3429 unsigned int alloc_flags,
3430 const struct alloc_context *ac)
3434 page = get_page_from_freelist(gfp_mask, order,
3435 alloc_flags|ALLOC_CPUSET, ac);
3437 * fallback to ignore cpuset restriction if our nodes
3441 page = get_page_from_freelist(gfp_mask, order,
3447 static inline struct page *
3448 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3449 const struct alloc_context *ac, unsigned long *did_some_progress)
3451 struct oom_control oc = {
3452 .zonelist = ac->zonelist,
3453 .nodemask = ac->nodemask,
3455 .gfp_mask = gfp_mask,
3460 *did_some_progress = 0;
3463 * Acquire the oom lock. If that fails, somebody else is
3464 * making progress for us.
3466 if (!mutex_trylock(&oom_lock)) {
3467 *did_some_progress = 1;
3468 schedule_timeout_uninterruptible(1);
3473 * Go through the zonelist yet one more time, keep very high watermark
3474 * here, this is only to catch a parallel oom killing, we must fail if
3475 * we're still under heavy pressure. But make sure that this reclaim
3476 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3477 * allocation which will never fail due to oom_lock already held.
3479 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3480 ~__GFP_DIRECT_RECLAIM, order,
3481 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3485 /* Coredumps can quickly deplete all memory reserves */
3486 if (current->flags & PF_DUMPCORE)
3488 /* The OOM killer will not help higher order allocs */
3489 if (order > PAGE_ALLOC_COSTLY_ORDER)
3492 * We have already exhausted all our reclaim opportunities without any
3493 * success so it is time to admit defeat. We will skip the OOM killer
3494 * because it is very likely that the caller has a more reasonable
3495 * fallback than shooting a random task.
3497 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3499 /* The OOM killer does not needlessly kill tasks for lowmem */
3500 if (ac->high_zoneidx < ZONE_NORMAL)
3502 if (pm_suspended_storage())
3505 * XXX: GFP_NOFS allocations should rather fail than rely on
3506 * other request to make a forward progress.
3507 * We are in an unfortunate situation where out_of_memory cannot
3508 * do much for this context but let's try it to at least get
3509 * access to memory reserved if the current task is killed (see
3510 * out_of_memory). Once filesystems are ready to handle allocation
3511 * failures more gracefully we should just bail out here.
3514 /* The OOM killer may not free memory on a specific node */
3515 if (gfp_mask & __GFP_THISNODE)
3518 /* Exhausted what can be done so it's blame time */
3519 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3520 *did_some_progress = 1;
3523 * Help non-failing allocations by giving them access to memory
3526 if (gfp_mask & __GFP_NOFAIL)
3527 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3528 ALLOC_NO_WATERMARKS, ac);
3531 mutex_unlock(&oom_lock);
3536 * Maximum number of compaction retries wit a progress before OOM
3537 * killer is consider as the only way to move forward.
3539 #define MAX_COMPACT_RETRIES 16
3541 #ifdef CONFIG_COMPACTION
3542 /* Try memory compaction for high-order allocations before reclaim */
3543 static struct page *
3544 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3545 unsigned int alloc_flags, const struct alloc_context *ac,
3546 enum compact_priority prio, enum compact_result *compact_result)
3549 unsigned long pflags;
3550 unsigned int noreclaim_flag;
3555 psi_memstall_enter(&pflags);
3556 noreclaim_flag = memalloc_noreclaim_save();
3558 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3561 memalloc_noreclaim_restore(noreclaim_flag);
3562 psi_memstall_leave(&pflags);
3564 if (*compact_result <= COMPACT_INACTIVE)
3568 * At least in one zone compaction wasn't deferred or skipped, so let's
3569 * count a compaction stall
3571 count_vm_event(COMPACTSTALL);
3573 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3576 struct zone *zone = page_zone(page);
3578 zone->compact_blockskip_flush = false;
3579 compaction_defer_reset(zone, order, true);
3580 count_vm_event(COMPACTSUCCESS);
3585 * It's bad if compaction run occurs and fails. The most likely reason
3586 * is that pages exist, but not enough to satisfy watermarks.
3588 count_vm_event(COMPACTFAIL);
3596 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3597 enum compact_result compact_result,
3598 enum compact_priority *compact_priority,
3599 int *compaction_retries)
3601 int max_retries = MAX_COMPACT_RETRIES;
3604 int retries = *compaction_retries;
3605 enum compact_priority priority = *compact_priority;
3610 if (compaction_made_progress(compact_result))
3611 (*compaction_retries)++;
3614 * compaction considers all the zone as desperately out of memory
3615 * so it doesn't really make much sense to retry except when the
3616 * failure could be caused by insufficient priority
3618 if (compaction_failed(compact_result))
3619 goto check_priority;
3622 * make sure the compaction wasn't deferred or didn't bail out early
3623 * due to locks contention before we declare that we should give up.
3624 * But do not retry if the given zonelist is not suitable for
3627 if (compaction_withdrawn(compact_result)) {
3628 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3633 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3634 * costly ones because they are de facto nofail and invoke OOM
3635 * killer to move on while costly can fail and users are ready
3636 * to cope with that. 1/4 retries is rather arbitrary but we
3637 * would need much more detailed feedback from compaction to
3638 * make a better decision.
3640 if (order > PAGE_ALLOC_COSTLY_ORDER)
3642 if (*compaction_retries <= max_retries) {
3648 * Make sure there are attempts at the highest priority if we exhausted
3649 * all retries or failed at the lower priorities.
3652 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3653 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3655 if (*compact_priority > min_priority) {
3656 (*compact_priority)--;
3657 *compaction_retries = 0;
3661 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3665 static inline struct page *
3666 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3667 unsigned int alloc_flags, const struct alloc_context *ac,
3668 enum compact_priority prio, enum compact_result *compact_result)
3670 *compact_result = COMPACT_SKIPPED;
3675 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3676 enum compact_result compact_result,
3677 enum compact_priority *compact_priority,
3678 int *compaction_retries)
3683 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3687 * There are setups with compaction disabled which would prefer to loop
3688 * inside the allocator rather than hit the oom killer prematurely.
3689 * Let's give them a good hope and keep retrying while the order-0
3690 * watermarks are OK.
3692 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3694 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3695 ac_classzone_idx(ac), alloc_flags))
3700 #endif /* CONFIG_COMPACTION */
3702 #ifdef CONFIG_LOCKDEP
3703 static struct lockdep_map __fs_reclaim_map =
3704 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3706 static bool __need_fs_reclaim(gfp_t gfp_mask)
3708 gfp_mask = current_gfp_context(gfp_mask);
3710 /* no reclaim without waiting on it */
3711 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3714 /* this guy won't enter reclaim */
3715 if (current->flags & PF_MEMALLOC)
3718 /* We're only interested __GFP_FS allocations for now */
3719 if (!(gfp_mask & __GFP_FS))
3722 if (gfp_mask & __GFP_NOLOCKDEP)
3728 void __fs_reclaim_acquire(void)
3730 lock_map_acquire(&__fs_reclaim_map);
3733 void __fs_reclaim_release(void)
3735 lock_map_release(&__fs_reclaim_map);
3738 void fs_reclaim_acquire(gfp_t gfp_mask)
3740 if (__need_fs_reclaim(gfp_mask))
3741 __fs_reclaim_acquire();
3743 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3745 void fs_reclaim_release(gfp_t gfp_mask)
3747 if (__need_fs_reclaim(gfp_mask))
3748 __fs_reclaim_release();
3750 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3753 /* Perform direct synchronous page reclaim */
3755 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3756 const struct alloc_context *ac)
3758 struct reclaim_state reclaim_state;
3760 unsigned int noreclaim_flag;
3761 unsigned long pflags;
3765 /* We now go into synchronous reclaim */
3766 cpuset_memory_pressure_bump();
3767 psi_memstall_enter(&pflags);
3768 fs_reclaim_acquire(gfp_mask);
3769 noreclaim_flag = memalloc_noreclaim_save();
3770 reclaim_state.reclaimed_slab = 0;
3771 current->reclaim_state = &reclaim_state;
3773 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3776 current->reclaim_state = NULL;
3777 memalloc_noreclaim_restore(noreclaim_flag);
3778 fs_reclaim_release(gfp_mask);
3779 psi_memstall_leave(&pflags);
3786 /* The really slow allocator path where we enter direct reclaim */
3787 static inline struct page *
3788 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3789 unsigned int alloc_flags, const struct alloc_context *ac,
3790 unsigned long *did_some_progress)
3792 struct page *page = NULL;
3793 bool drained = false;
3795 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3796 if (unlikely(!(*did_some_progress)))
3800 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3803 * If an allocation failed after direct reclaim, it could be because
3804 * pages are pinned on the per-cpu lists or in high alloc reserves.
3805 * Shrink them them and try again
3807 if (!page && !drained) {
3808 unreserve_highatomic_pageblock(ac, false);
3809 drain_all_pages(NULL);
3817 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3818 const struct alloc_context *ac)
3822 pg_data_t *last_pgdat = NULL;
3823 enum zone_type high_zoneidx = ac->high_zoneidx;
3825 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3827 if (last_pgdat != zone->zone_pgdat)
3828 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3829 last_pgdat = zone->zone_pgdat;
3833 static inline unsigned int
3834 gfp_to_alloc_flags(gfp_t gfp_mask)
3836 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3838 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3839 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3842 * The caller may dip into page reserves a bit more if the caller
3843 * cannot run direct reclaim, or if the caller has realtime scheduling
3844 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3845 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3847 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3849 if (gfp_mask & __GFP_ATOMIC) {
3851 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3852 * if it can't schedule.
3854 if (!(gfp_mask & __GFP_NOMEMALLOC))
3855 alloc_flags |= ALLOC_HARDER;
3857 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3858 * comment for __cpuset_node_allowed().
3860 alloc_flags &= ~ALLOC_CPUSET;
3861 } else if (unlikely(rt_task(current)) && !in_interrupt())
3862 alloc_flags |= ALLOC_HARDER;
3865 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3866 alloc_flags |= ALLOC_CMA;
3871 static bool oom_reserves_allowed(struct task_struct *tsk)
3873 if (!tsk_is_oom_victim(tsk))
3877 * !MMU doesn't have oom reaper so give access to memory reserves
3878 * only to the thread with TIF_MEMDIE set
3880 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3887 * Distinguish requests which really need access to full memory
3888 * reserves from oom victims which can live with a portion of it
3890 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3892 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3894 if (gfp_mask & __GFP_MEMALLOC)
3895 return ALLOC_NO_WATERMARKS;
3896 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3897 return ALLOC_NO_WATERMARKS;
3898 if (!in_interrupt()) {
3899 if (current->flags & PF_MEMALLOC)
3900 return ALLOC_NO_WATERMARKS;
3901 else if (oom_reserves_allowed(current))
3908 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3910 return !!__gfp_pfmemalloc_flags(gfp_mask);
3914 * Checks whether it makes sense to retry the reclaim to make a forward progress
3915 * for the given allocation request.
3917 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3918 * without success, or when we couldn't even meet the watermark if we
3919 * reclaimed all remaining pages on the LRU lists.
3921 * Returns true if a retry is viable or false to enter the oom path.
3924 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3925 struct alloc_context *ac, int alloc_flags,
3926 bool did_some_progress, int *no_progress_loops)
3933 * Costly allocations might have made a progress but this doesn't mean
3934 * their order will become available due to high fragmentation so
3935 * always increment the no progress counter for them
3937 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3938 *no_progress_loops = 0;
3940 (*no_progress_loops)++;
3943 * Make sure we converge to OOM if we cannot make any progress
3944 * several times in the row.
3946 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3947 /* Before OOM, exhaust highatomic_reserve */
3948 return unreserve_highatomic_pageblock(ac, true);
3952 * Keep reclaiming pages while there is a chance this will lead
3953 * somewhere. If none of the target zones can satisfy our allocation
3954 * request even if all reclaimable pages are considered then we are
3955 * screwed and have to go OOM.
3957 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3959 unsigned long available;
3960 unsigned long reclaimable;
3961 unsigned long min_wmark = min_wmark_pages(zone);
3964 available = reclaimable = zone_reclaimable_pages(zone);
3965 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3968 * Would the allocation succeed if we reclaimed all
3969 * reclaimable pages?
3971 wmark = __zone_watermark_ok(zone, order, min_wmark,
3972 ac_classzone_idx(ac), alloc_flags, available);
3973 trace_reclaim_retry_zone(z, order, reclaimable,
3974 available, min_wmark, *no_progress_loops, wmark);
3977 * If we didn't make any progress and have a lot of
3978 * dirty + writeback pages then we should wait for
3979 * an IO to complete to slow down the reclaim and
3980 * prevent from pre mature OOM
3982 if (!did_some_progress) {
3983 unsigned long write_pending;
3985 write_pending = zone_page_state_snapshot(zone,
3986 NR_ZONE_WRITE_PENDING);
3988 if (2 * write_pending > reclaimable) {
3989 congestion_wait(BLK_RW_ASYNC, HZ/10);
4001 * Memory allocation/reclaim might be called from a WQ context and the
4002 * current implementation of the WQ concurrency control doesn't
4003 * recognize that a particular WQ is congested if the worker thread is
4004 * looping without ever sleeping. Therefore we have to do a short sleep
4005 * here rather than calling cond_resched().
4007 if (current->flags & PF_WQ_WORKER)
4008 schedule_timeout_uninterruptible(1);
4015 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4018 * It's possible that cpuset's mems_allowed and the nodemask from
4019 * mempolicy don't intersect. This should be normally dealt with by
4020 * policy_nodemask(), but it's possible to race with cpuset update in
4021 * such a way the check therein was true, and then it became false
4022 * before we got our cpuset_mems_cookie here.
4023 * This assumes that for all allocations, ac->nodemask can come only
4024 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4025 * when it does not intersect with the cpuset restrictions) or the
4026 * caller can deal with a violated nodemask.
4028 if (cpusets_enabled() && ac->nodemask &&
4029 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4030 ac->nodemask = NULL;
4035 * When updating a task's mems_allowed or mempolicy nodemask, it is
4036 * possible to race with parallel threads in such a way that our
4037 * allocation can fail while the mask is being updated. If we are about
4038 * to fail, check if the cpuset changed during allocation and if so,
4041 if (read_mems_allowed_retry(cpuset_mems_cookie))
4047 static inline struct page *
4048 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4049 struct alloc_context *ac)
4051 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4052 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4053 struct page *page = NULL;
4054 unsigned int alloc_flags;
4055 unsigned long did_some_progress;
4056 enum compact_priority compact_priority;
4057 enum compact_result compact_result;
4058 int compaction_retries;
4059 int no_progress_loops;
4060 unsigned int cpuset_mems_cookie;
4064 * We also sanity check to catch abuse of atomic reserves being used by
4065 * callers that are not in atomic context.
4067 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4068 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4069 gfp_mask &= ~__GFP_ATOMIC;
4072 compaction_retries = 0;
4073 no_progress_loops = 0;
4074 compact_priority = DEF_COMPACT_PRIORITY;
4075 cpuset_mems_cookie = read_mems_allowed_begin();
4078 * The fast path uses conservative alloc_flags to succeed only until
4079 * kswapd needs to be woken up, and to avoid the cost of setting up
4080 * alloc_flags precisely. So we do that now.
4082 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4085 * We need to recalculate the starting point for the zonelist iterator
4086 * because we might have used different nodemask in the fast path, or
4087 * there was a cpuset modification and we are retrying - otherwise we
4088 * could end up iterating over non-eligible zones endlessly.
4090 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4091 ac->high_zoneidx, ac->nodemask);
4092 if (!ac->preferred_zoneref->zone)
4095 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4096 wake_all_kswapds(order, gfp_mask, ac);
4099 * The adjusted alloc_flags might result in immediate success, so try
4102 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4107 * For costly allocations, try direct compaction first, as it's likely
4108 * that we have enough base pages and don't need to reclaim. For non-
4109 * movable high-order allocations, do that as well, as compaction will
4110 * try prevent permanent fragmentation by migrating from blocks of the
4112 * Don't try this for allocations that are allowed to ignore
4113 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4115 if (can_direct_reclaim &&
4117 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4118 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4119 page = __alloc_pages_direct_compact(gfp_mask, order,
4121 INIT_COMPACT_PRIORITY,
4127 * Checks for costly allocations with __GFP_NORETRY, which
4128 * includes THP page fault allocations
4130 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4132 * If compaction is deferred for high-order allocations,
4133 * it is because sync compaction recently failed. If
4134 * this is the case and the caller requested a THP
4135 * allocation, we do not want to heavily disrupt the
4136 * system, so we fail the allocation instead of entering
4139 if (compact_result == COMPACT_DEFERRED)
4143 * Looks like reclaim/compaction is worth trying, but
4144 * sync compaction could be very expensive, so keep
4145 * using async compaction.
4147 compact_priority = INIT_COMPACT_PRIORITY;
4152 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4153 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4154 wake_all_kswapds(order, gfp_mask, ac);
4156 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4158 alloc_flags = reserve_flags;
4161 * Reset the nodemask and zonelist iterators if memory policies can be
4162 * ignored. These allocations are high priority and system rather than
4165 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4166 ac->nodemask = NULL;
4167 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4168 ac->high_zoneidx, ac->nodemask);
4171 /* Attempt with potentially adjusted zonelist and alloc_flags */
4172 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4176 /* Caller is not willing to reclaim, we can't balance anything */
4177 if (!can_direct_reclaim)
4180 /* Avoid recursion of direct reclaim */
4181 if (current->flags & PF_MEMALLOC)
4184 /* Try direct reclaim and then allocating */
4185 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4186 &did_some_progress);
4190 /* Try direct compaction and then allocating */
4191 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4192 compact_priority, &compact_result);
4196 /* Do not loop if specifically requested */
4197 if (gfp_mask & __GFP_NORETRY)
4201 * Do not retry costly high order allocations unless they are
4202 * __GFP_RETRY_MAYFAIL
4204 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4207 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4208 did_some_progress > 0, &no_progress_loops))
4212 * It doesn't make any sense to retry for the compaction if the order-0
4213 * reclaim is not able to make any progress because the current
4214 * implementation of the compaction depends on the sufficient amount
4215 * of free memory (see __compaction_suitable)
4217 if (did_some_progress > 0 &&
4218 should_compact_retry(ac, order, alloc_flags,
4219 compact_result, &compact_priority,
4220 &compaction_retries))
4224 /* Deal with possible cpuset update races before we start OOM killing */
4225 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4228 /* Reclaim has failed us, start killing things */
4229 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4233 /* Avoid allocations with no watermarks from looping endlessly */
4234 if (tsk_is_oom_victim(current) &&
4235 (alloc_flags == ALLOC_OOM ||
4236 (gfp_mask & __GFP_NOMEMALLOC)))
4239 /* Retry as long as the OOM killer is making progress */
4240 if (did_some_progress) {
4241 no_progress_loops = 0;
4246 /* Deal with possible cpuset update races before we fail */
4247 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4251 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4254 if (gfp_mask & __GFP_NOFAIL) {
4256 * All existing users of the __GFP_NOFAIL are blockable, so warn
4257 * of any new users that actually require GFP_NOWAIT
4259 if (WARN_ON_ONCE(!can_direct_reclaim))
4263 * PF_MEMALLOC request from this context is rather bizarre
4264 * because we cannot reclaim anything and only can loop waiting
4265 * for somebody to do a work for us
4267 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4270 * non failing costly orders are a hard requirement which we
4271 * are not prepared for much so let's warn about these users
4272 * so that we can identify them and convert them to something
4275 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4278 * Help non-failing allocations by giving them access to memory
4279 * reserves but do not use ALLOC_NO_WATERMARKS because this
4280 * could deplete whole memory reserves which would just make
4281 * the situation worse
4283 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4291 warn_alloc(gfp_mask, ac->nodemask,
4292 "page allocation failure: order:%u", order);
4297 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4298 int preferred_nid, nodemask_t *nodemask,
4299 struct alloc_context *ac, gfp_t *alloc_mask,
4300 unsigned int *alloc_flags)
4302 ac->high_zoneidx = gfp_zone(gfp_mask);
4303 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4304 ac->nodemask = nodemask;
4305 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4307 if (cpusets_enabled()) {
4308 *alloc_mask |= __GFP_HARDWALL;
4310 ac->nodemask = &cpuset_current_mems_allowed;
4312 *alloc_flags |= ALLOC_CPUSET;
4315 fs_reclaim_acquire(gfp_mask);
4316 fs_reclaim_release(gfp_mask);
4318 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4320 if (should_fail_alloc_page(gfp_mask, order))
4323 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4324 *alloc_flags |= ALLOC_CMA;
4329 /* Determine whether to spread dirty pages and what the first usable zone */
4330 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4332 /* Dirty zone balancing only done in the fast path */
4333 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4336 * The preferred zone is used for statistics but crucially it is
4337 * also used as the starting point for the zonelist iterator. It
4338 * may get reset for allocations that ignore memory policies.
4340 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4341 ac->high_zoneidx, ac->nodemask);
4345 * This is the 'heart' of the zoned buddy allocator.
4348 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4349 nodemask_t *nodemask)
4352 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4353 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4354 struct alloc_context ac = { };
4357 * There are several places where we assume that the order value is sane
4358 * so bail out early if the request is out of bound.
4360 if (unlikely(order >= MAX_ORDER)) {
4361 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4365 gfp_mask &= gfp_allowed_mask;
4366 alloc_mask = gfp_mask;
4367 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4370 finalise_ac(gfp_mask, &ac);
4372 /* First allocation attempt */
4373 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4378 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4379 * resp. GFP_NOIO which has to be inherited for all allocation requests
4380 * from a particular context which has been marked by
4381 * memalloc_no{fs,io}_{save,restore}.
4383 alloc_mask = current_gfp_context(gfp_mask);
4384 ac.spread_dirty_pages = false;
4387 * Restore the original nodemask if it was potentially replaced with
4388 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4390 if (unlikely(ac.nodemask != nodemask))
4391 ac.nodemask = nodemask;
4393 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4396 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4397 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4398 __free_pages(page, order);
4402 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4406 EXPORT_SYMBOL(__alloc_pages_nodemask);
4409 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4410 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4411 * you need to access high mem.
4413 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4417 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4420 return (unsigned long) page_address(page);
4422 EXPORT_SYMBOL(__get_free_pages);
4424 unsigned long get_zeroed_page(gfp_t gfp_mask)
4426 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4428 EXPORT_SYMBOL(get_zeroed_page);
4430 static inline void free_the_page(struct page *page, unsigned int order)
4432 if (order == 0) /* Via pcp? */
4433 free_unref_page(page);
4435 __free_pages_ok(page, order);
4438 void __free_pages(struct page *page, unsigned int order)
4440 if (put_page_testzero(page))
4441 free_the_page(page, order);
4443 EXPORT_SYMBOL(__free_pages);
4445 void free_pages(unsigned long addr, unsigned int order)
4448 VM_BUG_ON(!virt_addr_valid((void *)addr));
4449 __free_pages(virt_to_page((void *)addr), order);
4453 EXPORT_SYMBOL(free_pages);
4457 * An arbitrary-length arbitrary-offset area of memory which resides
4458 * within a 0 or higher order page. Multiple fragments within that page
4459 * are individually refcounted, in the page's reference counter.
4461 * The page_frag functions below provide a simple allocation framework for
4462 * page fragments. This is used by the network stack and network device
4463 * drivers to provide a backing region of memory for use as either an
4464 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4466 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4469 struct page *page = NULL;
4470 gfp_t gfp = gfp_mask;
4472 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4473 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4475 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4476 PAGE_FRAG_CACHE_MAX_ORDER);
4477 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4479 if (unlikely(!page))
4480 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4482 nc->va = page ? page_address(page) : NULL;
4487 void __page_frag_cache_drain(struct page *page, unsigned int count)
4489 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4491 if (page_ref_sub_and_test(page, count))
4492 free_the_page(page, compound_order(page));
4494 EXPORT_SYMBOL(__page_frag_cache_drain);
4496 void *page_frag_alloc(struct page_frag_cache *nc,
4497 unsigned int fragsz, gfp_t gfp_mask)
4499 unsigned int size = PAGE_SIZE;
4503 if (unlikely(!nc->va)) {
4505 page = __page_frag_cache_refill(nc, gfp_mask);
4509 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4510 /* if size can vary use size else just use PAGE_SIZE */
4513 /* Even if we own the page, we do not use atomic_set().
4514 * This would break get_page_unless_zero() users.
4516 page_ref_add(page, size - 1);
4518 /* reset page count bias and offset to start of new frag */
4519 nc->pfmemalloc = page_is_pfmemalloc(page);
4520 nc->pagecnt_bias = size;
4524 offset = nc->offset - fragsz;
4525 if (unlikely(offset < 0)) {
4526 page = virt_to_page(nc->va);
4528 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4531 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4532 /* if size can vary use size else just use PAGE_SIZE */
4535 /* OK, page count is 0, we can safely set it */
4536 set_page_count(page, size);
4538 /* reset page count bias and offset to start of new frag */
4539 nc->pagecnt_bias = size;
4540 offset = size - fragsz;
4544 nc->offset = offset;
4546 return nc->va + offset;
4548 EXPORT_SYMBOL(page_frag_alloc);
4551 * Frees a page fragment allocated out of either a compound or order 0 page.
4553 void page_frag_free(void *addr)
4555 struct page *page = virt_to_head_page(addr);
4557 if (unlikely(put_page_testzero(page)))
4558 free_the_page(page, compound_order(page));
4560 EXPORT_SYMBOL(page_frag_free);
4562 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4566 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4567 unsigned long used = addr + PAGE_ALIGN(size);
4569 split_page(virt_to_page((void *)addr), order);
4570 while (used < alloc_end) {
4575 return (void *)addr;
4579 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4580 * @size: the number of bytes to allocate
4581 * @gfp_mask: GFP flags for the allocation
4583 * This function is similar to alloc_pages(), except that it allocates the
4584 * minimum number of pages to satisfy the request. alloc_pages() can only
4585 * allocate memory in power-of-two pages.
4587 * This function is also limited by MAX_ORDER.
4589 * Memory allocated by this function must be released by free_pages_exact().
4591 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4593 unsigned int order = get_order(size);
4596 addr = __get_free_pages(gfp_mask, order);
4597 return make_alloc_exact(addr, order, size);
4599 EXPORT_SYMBOL(alloc_pages_exact);
4602 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4604 * @nid: the preferred node ID where memory should be allocated
4605 * @size: the number of bytes to allocate
4606 * @gfp_mask: GFP flags for the allocation
4608 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4611 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4613 unsigned int order = get_order(size);
4614 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4617 return make_alloc_exact((unsigned long)page_address(p), order, size);
4621 * free_pages_exact - release memory allocated via alloc_pages_exact()
4622 * @virt: the value returned by alloc_pages_exact.
4623 * @size: size of allocation, same value as passed to alloc_pages_exact().
4625 * Release the memory allocated by a previous call to alloc_pages_exact.
4627 void free_pages_exact(void *virt, size_t size)
4629 unsigned long addr = (unsigned long)virt;
4630 unsigned long end = addr + PAGE_ALIGN(size);
4632 while (addr < end) {
4637 EXPORT_SYMBOL(free_pages_exact);
4640 * nr_free_zone_pages - count number of pages beyond high watermark
4641 * @offset: The zone index of the highest zone
4643 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4644 * high watermark within all zones at or below a given zone index. For each
4645 * zone, the number of pages is calculated as:
4647 * nr_free_zone_pages = managed_pages - high_pages
4649 static unsigned long nr_free_zone_pages(int offset)
4654 /* Just pick one node, since fallback list is circular */
4655 unsigned long sum = 0;
4657 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4659 for_each_zone_zonelist(zone, z, zonelist, offset) {
4660 unsigned long size = zone_managed_pages(zone);
4661 unsigned long high = high_wmark_pages(zone);
4670 * nr_free_buffer_pages - count number of pages beyond high watermark
4672 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4673 * watermark within ZONE_DMA and ZONE_NORMAL.
4675 unsigned long nr_free_buffer_pages(void)
4677 return nr_free_zone_pages(gfp_zone(GFP_USER));
4679 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4682 * nr_free_pagecache_pages - count number of pages beyond high watermark
4684 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4685 * high watermark within all zones.
4687 unsigned long nr_free_pagecache_pages(void)
4689 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4692 static inline void show_node(struct zone *zone)
4694 if (IS_ENABLED(CONFIG_NUMA))
4695 printk("Node %d ", zone_to_nid(zone));
4698 long si_mem_available(void)
4701 unsigned long pagecache;
4702 unsigned long wmark_low = 0;
4703 unsigned long pages[NR_LRU_LISTS];
4704 unsigned long reclaimable;
4708 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4709 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4712 wmark_low += zone->watermark[WMARK_LOW];
4715 * Estimate the amount of memory available for userspace allocations,
4716 * without causing swapping.
4718 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4721 * Not all the page cache can be freed, otherwise the system will
4722 * start swapping. Assume at least half of the page cache, or the
4723 * low watermark worth of cache, needs to stay.
4725 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4726 pagecache -= min(pagecache / 2, wmark_low);
4727 available += pagecache;
4730 * Part of the reclaimable slab and other kernel memory consists of
4731 * items that are in use, and cannot be freed. Cap this estimate at the
4734 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4735 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4736 available += reclaimable - min(reclaimable / 2, wmark_low);
4742 EXPORT_SYMBOL_GPL(si_mem_available);
4744 void si_meminfo(struct sysinfo *val)
4746 val->totalram = totalram_pages();
4747 val->sharedram = global_node_page_state(NR_SHMEM);
4748 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4749 val->bufferram = nr_blockdev_pages();
4750 val->totalhigh = totalhigh_pages();
4751 val->freehigh = nr_free_highpages();
4752 val->mem_unit = PAGE_SIZE;
4755 EXPORT_SYMBOL(si_meminfo);
4758 void si_meminfo_node(struct sysinfo *val, int nid)
4760 int zone_type; /* needs to be signed */
4761 unsigned long managed_pages = 0;
4762 unsigned long managed_highpages = 0;
4763 unsigned long free_highpages = 0;
4764 pg_data_t *pgdat = NODE_DATA(nid);
4766 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4767 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
4768 val->totalram = managed_pages;
4769 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4770 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4771 #ifdef CONFIG_HIGHMEM
4772 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4773 struct zone *zone = &pgdat->node_zones[zone_type];
4775 if (is_highmem(zone)) {
4776 managed_highpages += zone_managed_pages(zone);
4777 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4780 val->totalhigh = managed_highpages;
4781 val->freehigh = free_highpages;
4783 val->totalhigh = managed_highpages;
4784 val->freehigh = free_highpages;
4786 val->mem_unit = PAGE_SIZE;
4791 * Determine whether the node should be displayed or not, depending on whether
4792 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4794 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4796 if (!(flags & SHOW_MEM_FILTER_NODES))
4800 * no node mask - aka implicit memory numa policy. Do not bother with
4801 * the synchronization - read_mems_allowed_begin - because we do not
4802 * have to be precise here.
4805 nodemask = &cpuset_current_mems_allowed;
4807 return !node_isset(nid, *nodemask);
4810 #define K(x) ((x) << (PAGE_SHIFT-10))
4812 static void show_migration_types(unsigned char type)
4814 static const char types[MIGRATE_TYPES] = {
4815 [MIGRATE_UNMOVABLE] = 'U',
4816 [MIGRATE_MOVABLE] = 'M',
4817 [MIGRATE_RECLAIMABLE] = 'E',
4818 [MIGRATE_HIGHATOMIC] = 'H',
4820 [MIGRATE_CMA] = 'C',
4822 #ifdef CONFIG_MEMORY_ISOLATION
4823 [MIGRATE_ISOLATE] = 'I',
4826 char tmp[MIGRATE_TYPES + 1];
4830 for (i = 0; i < MIGRATE_TYPES; i++) {
4831 if (type & (1 << i))
4836 printk(KERN_CONT "(%s) ", tmp);
4840 * Show free area list (used inside shift_scroll-lock stuff)
4841 * We also calculate the percentage fragmentation. We do this by counting the
4842 * memory on each free list with the exception of the first item on the list.
4845 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4848 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4850 unsigned long free_pcp = 0;
4855 for_each_populated_zone(zone) {
4856 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4859 for_each_online_cpu(cpu)
4860 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4863 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4864 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4865 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4866 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4867 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4868 " free:%lu free_pcp:%lu free_cma:%lu\n",
4869 global_node_page_state(NR_ACTIVE_ANON),
4870 global_node_page_state(NR_INACTIVE_ANON),
4871 global_node_page_state(NR_ISOLATED_ANON),
4872 global_node_page_state(NR_ACTIVE_FILE),
4873 global_node_page_state(NR_INACTIVE_FILE),
4874 global_node_page_state(NR_ISOLATED_FILE),
4875 global_node_page_state(NR_UNEVICTABLE),
4876 global_node_page_state(NR_FILE_DIRTY),
4877 global_node_page_state(NR_WRITEBACK),
4878 global_node_page_state(NR_UNSTABLE_NFS),
4879 global_node_page_state(NR_SLAB_RECLAIMABLE),
4880 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4881 global_node_page_state(NR_FILE_MAPPED),
4882 global_node_page_state(NR_SHMEM),
4883 global_zone_page_state(NR_PAGETABLE),
4884 global_zone_page_state(NR_BOUNCE),
4885 global_zone_page_state(NR_FREE_PAGES),
4887 global_zone_page_state(NR_FREE_CMA_PAGES));
4889 for_each_online_pgdat(pgdat) {
4890 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4894 " active_anon:%lukB"
4895 " inactive_anon:%lukB"
4896 " active_file:%lukB"
4897 " inactive_file:%lukB"
4898 " unevictable:%lukB"
4899 " isolated(anon):%lukB"
4900 " isolated(file):%lukB"
4905 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4907 " shmem_pmdmapped: %lukB"
4910 " writeback_tmp:%lukB"
4912 " all_unreclaimable? %s"
4915 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4916 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4917 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4918 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4919 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4920 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4921 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4922 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4923 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4924 K(node_page_state(pgdat, NR_WRITEBACK)),
4925 K(node_page_state(pgdat, NR_SHMEM)),
4926 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4927 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4928 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4930 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4932 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4933 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4934 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4938 for_each_populated_zone(zone) {
4941 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4945 for_each_online_cpu(cpu)
4946 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4955 " active_anon:%lukB"
4956 " inactive_anon:%lukB"
4957 " active_file:%lukB"
4958 " inactive_file:%lukB"
4959 " unevictable:%lukB"
4960 " writepending:%lukB"
4964 " kernel_stack:%lukB"
4972 K(zone_page_state(zone, NR_FREE_PAGES)),
4973 K(min_wmark_pages(zone)),
4974 K(low_wmark_pages(zone)),
4975 K(high_wmark_pages(zone)),
4976 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4977 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4978 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4979 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4980 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4981 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4982 K(zone->present_pages),
4983 K(zone_managed_pages(zone)),
4984 K(zone_page_state(zone, NR_MLOCK)),
4985 zone_page_state(zone, NR_KERNEL_STACK_KB),
4986 K(zone_page_state(zone, NR_PAGETABLE)),
4987 K(zone_page_state(zone, NR_BOUNCE)),
4989 K(this_cpu_read(zone->pageset->pcp.count)),
4990 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4991 printk("lowmem_reserve[]:");
4992 for (i = 0; i < MAX_NR_ZONES; i++)
4993 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4994 printk(KERN_CONT "\n");
4997 for_each_populated_zone(zone) {
4999 unsigned long nr[MAX_ORDER], flags, total = 0;
5000 unsigned char types[MAX_ORDER];
5002 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5005 printk(KERN_CONT "%s: ", zone->name);
5007 spin_lock_irqsave(&zone->lock, flags);
5008 for (order = 0; order < MAX_ORDER; order++) {
5009 struct free_area *area = &zone->free_area[order];
5012 nr[order] = area->nr_free;
5013 total += nr[order] << order;
5016 for (type = 0; type < MIGRATE_TYPES; type++) {
5017 if (!list_empty(&area->free_list[type]))
5018 types[order] |= 1 << type;
5021 spin_unlock_irqrestore(&zone->lock, flags);
5022 for (order = 0; order < MAX_ORDER; order++) {
5023 printk(KERN_CONT "%lu*%lukB ",
5024 nr[order], K(1UL) << order);
5026 show_migration_types(types[order]);
5028 printk(KERN_CONT "= %lukB\n", K(total));
5031 hugetlb_show_meminfo();
5033 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5035 show_swap_cache_info();
5038 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5040 zoneref->zone = zone;
5041 zoneref->zone_idx = zone_idx(zone);
5045 * Builds allocation fallback zone lists.
5047 * Add all populated zones of a node to the zonelist.
5049 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5052 enum zone_type zone_type = MAX_NR_ZONES;
5057 zone = pgdat->node_zones + zone_type;
5058 if (managed_zone(zone)) {
5059 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5060 check_highest_zone(zone_type);
5062 } while (zone_type);
5069 static int __parse_numa_zonelist_order(char *s)
5072 * We used to support different zonlists modes but they turned
5073 * out to be just not useful. Let's keep the warning in place
5074 * if somebody still use the cmd line parameter so that we do
5075 * not fail it silently
5077 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5078 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5084 static __init int setup_numa_zonelist_order(char *s)
5089 return __parse_numa_zonelist_order(s);
5091 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5093 char numa_zonelist_order[] = "Node";
5096 * sysctl handler for numa_zonelist_order
5098 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5099 void __user *buffer, size_t *length,
5106 return proc_dostring(table, write, buffer, length, ppos);
5107 str = memdup_user_nul(buffer, 16);
5109 return PTR_ERR(str);
5111 ret = __parse_numa_zonelist_order(str);
5117 #define MAX_NODE_LOAD (nr_online_nodes)
5118 static int node_load[MAX_NUMNODES];
5121 * find_next_best_node - find the next node that should appear in a given node's fallback list
5122 * @node: node whose fallback list we're appending
5123 * @used_node_mask: nodemask_t of already used nodes
5125 * We use a number of factors to determine which is the next node that should
5126 * appear on a given node's fallback list. The node should not have appeared
5127 * already in @node's fallback list, and it should be the next closest node
5128 * according to the distance array (which contains arbitrary distance values
5129 * from each node to each node in the system), and should also prefer nodes
5130 * with no CPUs, since presumably they'll have very little allocation pressure
5131 * on them otherwise.
5132 * It returns -1 if no node is found.
5134 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5137 int min_val = INT_MAX;
5138 int best_node = NUMA_NO_NODE;
5139 const struct cpumask *tmp = cpumask_of_node(0);
5141 /* Use the local node if we haven't already */
5142 if (!node_isset(node, *used_node_mask)) {
5143 node_set(node, *used_node_mask);
5147 for_each_node_state(n, N_MEMORY) {
5149 /* Don't want a node to appear more than once */
5150 if (node_isset(n, *used_node_mask))
5153 /* Use the distance array to find the distance */
5154 val = node_distance(node, n);
5156 /* Penalize nodes under us ("prefer the next node") */
5159 /* Give preference to headless and unused nodes */
5160 tmp = cpumask_of_node(n);
5161 if (!cpumask_empty(tmp))
5162 val += PENALTY_FOR_NODE_WITH_CPUS;
5164 /* Slight preference for less loaded node */
5165 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5166 val += node_load[n];
5168 if (val < min_val) {
5175 node_set(best_node, *used_node_mask);
5182 * Build zonelists ordered by node and zones within node.
5183 * This results in maximum locality--normal zone overflows into local
5184 * DMA zone, if any--but risks exhausting DMA zone.
5186 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5189 struct zoneref *zonerefs;
5192 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5194 for (i = 0; i < nr_nodes; i++) {
5197 pg_data_t *node = NODE_DATA(node_order[i]);
5199 nr_zones = build_zonerefs_node(node, zonerefs);
5200 zonerefs += nr_zones;
5202 zonerefs->zone = NULL;
5203 zonerefs->zone_idx = 0;
5207 * Build gfp_thisnode zonelists
5209 static void build_thisnode_zonelists(pg_data_t *pgdat)
5211 struct zoneref *zonerefs;
5214 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5215 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5216 zonerefs += nr_zones;
5217 zonerefs->zone = NULL;
5218 zonerefs->zone_idx = 0;
5222 * Build zonelists ordered by zone and nodes within zones.
5223 * This results in conserving DMA zone[s] until all Normal memory is
5224 * exhausted, but results in overflowing to remote node while memory
5225 * may still exist in local DMA zone.
5228 static void build_zonelists(pg_data_t *pgdat)
5230 static int node_order[MAX_NUMNODES];
5231 int node, load, nr_nodes = 0;
5232 nodemask_t used_mask;
5233 int local_node, prev_node;
5235 /* NUMA-aware ordering of nodes */
5236 local_node = pgdat->node_id;
5237 load = nr_online_nodes;
5238 prev_node = local_node;
5239 nodes_clear(used_mask);
5241 memset(node_order, 0, sizeof(node_order));
5242 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5244 * We don't want to pressure a particular node.
5245 * So adding penalty to the first node in same
5246 * distance group to make it round-robin.
5248 if (node_distance(local_node, node) !=
5249 node_distance(local_node, prev_node))
5250 node_load[node] = load;
5252 node_order[nr_nodes++] = node;
5257 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5258 build_thisnode_zonelists(pgdat);
5261 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5263 * Return node id of node used for "local" allocations.
5264 * I.e., first node id of first zone in arg node's generic zonelist.
5265 * Used for initializing percpu 'numa_mem', which is used primarily
5266 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5268 int local_memory_node(int node)
5272 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5273 gfp_zone(GFP_KERNEL),
5275 return zone_to_nid(z->zone);
5279 static void setup_min_unmapped_ratio(void);
5280 static void setup_min_slab_ratio(void);
5281 #else /* CONFIG_NUMA */
5283 static void build_zonelists(pg_data_t *pgdat)
5285 int node, local_node;
5286 struct zoneref *zonerefs;
5289 local_node = pgdat->node_id;
5291 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5292 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5293 zonerefs += nr_zones;
5296 * Now we build the zonelist so that it contains the zones
5297 * of all the other nodes.
5298 * We don't want to pressure a particular node, so when
5299 * building the zones for node N, we make sure that the
5300 * zones coming right after the local ones are those from
5301 * node N+1 (modulo N)
5303 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5304 if (!node_online(node))
5306 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5307 zonerefs += nr_zones;
5309 for (node = 0; node < local_node; node++) {
5310 if (!node_online(node))
5312 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5313 zonerefs += nr_zones;
5316 zonerefs->zone = NULL;
5317 zonerefs->zone_idx = 0;
5320 #endif /* CONFIG_NUMA */
5323 * Boot pageset table. One per cpu which is going to be used for all
5324 * zones and all nodes. The parameters will be set in such a way
5325 * that an item put on a list will immediately be handed over to
5326 * the buddy list. This is safe since pageset manipulation is done
5327 * with interrupts disabled.
5329 * The boot_pagesets must be kept even after bootup is complete for
5330 * unused processors and/or zones. They do play a role for bootstrapping
5331 * hotplugged processors.
5333 * zoneinfo_show() and maybe other functions do
5334 * not check if the processor is online before following the pageset pointer.
5335 * Other parts of the kernel may not check if the zone is available.
5337 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5338 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5339 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5341 static void __build_all_zonelists(void *data)
5344 int __maybe_unused cpu;
5345 pg_data_t *self = data;
5346 static DEFINE_SPINLOCK(lock);
5351 memset(node_load, 0, sizeof(node_load));
5355 * This node is hotadded and no memory is yet present. So just
5356 * building zonelists is fine - no need to touch other nodes.
5358 if (self && !node_online(self->node_id)) {
5359 build_zonelists(self);
5361 for_each_online_node(nid) {
5362 pg_data_t *pgdat = NODE_DATA(nid);
5364 build_zonelists(pgdat);
5367 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5369 * We now know the "local memory node" for each node--
5370 * i.e., the node of the first zone in the generic zonelist.
5371 * Set up numa_mem percpu variable for on-line cpus. During
5372 * boot, only the boot cpu should be on-line; we'll init the
5373 * secondary cpus' numa_mem as they come on-line. During
5374 * node/memory hotplug, we'll fixup all on-line cpus.
5376 for_each_online_cpu(cpu)
5377 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5384 static noinline void __init
5385 build_all_zonelists_init(void)
5389 __build_all_zonelists(NULL);
5392 * Initialize the boot_pagesets that are going to be used
5393 * for bootstrapping processors. The real pagesets for
5394 * each zone will be allocated later when the per cpu
5395 * allocator is available.
5397 * boot_pagesets are used also for bootstrapping offline
5398 * cpus if the system is already booted because the pagesets
5399 * are needed to initialize allocators on a specific cpu too.
5400 * F.e. the percpu allocator needs the page allocator which
5401 * needs the percpu allocator in order to allocate its pagesets
5402 * (a chicken-egg dilemma).
5404 for_each_possible_cpu(cpu)
5405 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5407 mminit_verify_zonelist();
5408 cpuset_init_current_mems_allowed();
5412 * unless system_state == SYSTEM_BOOTING.
5414 * __ref due to call of __init annotated helper build_all_zonelists_init
5415 * [protected by SYSTEM_BOOTING].
5417 void __ref build_all_zonelists(pg_data_t *pgdat)
5419 if (system_state == SYSTEM_BOOTING) {
5420 build_all_zonelists_init();
5422 __build_all_zonelists(pgdat);
5423 /* cpuset refresh routine should be here */
5425 vm_total_pages = nr_free_pagecache_pages();
5427 * Disable grouping by mobility if the number of pages in the
5428 * system is too low to allow the mechanism to work. It would be
5429 * more accurate, but expensive to check per-zone. This check is
5430 * made on memory-hotadd so a system can start with mobility
5431 * disabled and enable it later
5433 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5434 page_group_by_mobility_disabled = 1;
5436 page_group_by_mobility_disabled = 0;
5438 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5440 page_group_by_mobility_disabled ? "off" : "on",
5443 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5447 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5448 static bool __meminit
5449 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5451 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5452 static struct memblock_region *r;
5454 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5455 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5456 for_each_memblock(memory, r) {
5457 if (*pfn < memblock_region_memory_end_pfn(r))
5461 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5462 memblock_is_mirror(r)) {
5463 *pfn = memblock_region_memory_end_pfn(r);
5472 * Initially all pages are reserved - free ones are freed
5473 * up by memblock_free_all() once the early boot process is
5474 * done. Non-atomic initialization, single-pass.
5476 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5477 unsigned long start_pfn, enum memmap_context context,
5478 struct vmem_altmap *altmap)
5480 unsigned long pfn, end_pfn = start_pfn + size;
5483 if (highest_memmap_pfn < end_pfn - 1)
5484 highest_memmap_pfn = end_pfn - 1;
5486 #ifdef CONFIG_ZONE_DEVICE
5488 * Honor reservation requested by the driver for this ZONE_DEVICE
5489 * memory. We limit the total number of pages to initialize to just
5490 * those that might contain the memory mapping. We will defer the
5491 * ZONE_DEVICE page initialization until after we have released
5494 if (zone == ZONE_DEVICE) {
5498 if (start_pfn == altmap->base_pfn)
5499 start_pfn += altmap->reserve;
5500 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
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) {
5510 if (!early_pfn_valid(pfn))
5512 if (!early_pfn_in_nid(pfn, nid))
5514 if (overlap_memmap_init(zone, &pfn))
5516 if (defer_init(nid, pfn, end_pfn))
5520 page = pfn_to_page(pfn);
5521 __init_single_page(page, pfn, zone, nid);
5522 if (context == MEMMAP_HOTPLUG)
5523 __SetPageReserved(page);
5526 * Mark the block movable so that blocks are reserved for
5527 * movable at startup. This will force kernel allocations
5528 * to reserve their blocks rather than leaking throughout
5529 * the address space during boot when many long-lived
5530 * kernel allocations are made.
5532 * bitmap is created for zone's valid pfn range. but memmap
5533 * can be created for invalid pages (for alignment)
5534 * check here not to call set_pageblock_migratetype() against
5537 if (!(pfn & (pageblock_nr_pages - 1))) {
5538 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5542 #ifdef CONFIG_SPARSEMEM
5544 * If the zone does not span the rest of the section then
5545 * we should at least initialize those pages. Otherwise we
5546 * could blow up on a poisoned page in some paths which depend
5547 * on full sections being initialized (e.g. memory hotplug).
5549 while (end_pfn % PAGES_PER_SECTION) {
5550 __init_single_page(pfn_to_page(end_pfn), end_pfn, zone, nid);
5556 #ifdef CONFIG_ZONE_DEVICE
5557 void __ref memmap_init_zone_device(struct zone *zone,
5558 unsigned long start_pfn,
5560 struct dev_pagemap *pgmap)
5562 unsigned long pfn, end_pfn = start_pfn + size;
5563 struct pglist_data *pgdat = zone->zone_pgdat;
5564 unsigned long zone_idx = zone_idx(zone);
5565 unsigned long start = jiffies;
5566 int nid = pgdat->node_id;
5568 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5572 * The call to memmap_init_zone should have already taken care
5573 * of the pages reserved for the memmap, so we can just jump to
5574 * the end of that region and start processing the device pages.
5576 if (pgmap->altmap_valid) {
5577 struct vmem_altmap *altmap = &pgmap->altmap;
5579 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5580 size = end_pfn - start_pfn;
5583 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5584 struct page *page = pfn_to_page(pfn);
5586 __init_single_page(page, pfn, zone_idx, nid);
5589 * Mark page reserved as it will need to wait for onlining
5590 * phase for it to be fully associated with a zone.
5592 * We can use the non-atomic __set_bit operation for setting
5593 * the flag as we are still initializing the pages.
5595 __SetPageReserved(page);
5598 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5599 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5600 * page is ever freed or placed on a driver-private list.
5602 page->pgmap = pgmap;
5606 * Mark the block movable so that blocks are reserved for
5607 * movable at startup. This will force kernel allocations
5608 * to reserve their blocks rather than leaking throughout
5609 * the address space during boot when many long-lived
5610 * kernel allocations are made.
5612 * bitmap is created for zone's valid pfn range. but memmap
5613 * can be created for invalid pages (for alignment)
5614 * check here not to call set_pageblock_migratetype() against
5617 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5618 * because this is done early in sparse_add_one_section
5620 if (!(pfn & (pageblock_nr_pages - 1))) {
5621 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5626 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5627 size, jiffies_to_msecs(jiffies - start));
5631 static void __meminit zone_init_free_lists(struct zone *zone)
5633 unsigned int order, t;
5634 for_each_migratetype_order(order, t) {
5635 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5636 zone->free_area[order].nr_free = 0;
5640 void __meminit __weak memmap_init(unsigned long size, int nid,
5641 unsigned long zone, unsigned long start_pfn)
5643 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5646 static int zone_batchsize(struct zone *zone)
5652 * The per-cpu-pages pools are set to around 1000th of the
5655 batch = zone_managed_pages(zone) / 1024;
5656 /* But no more than a meg. */
5657 if (batch * PAGE_SIZE > 1024 * 1024)
5658 batch = (1024 * 1024) / PAGE_SIZE;
5659 batch /= 4; /* We effectively *= 4 below */
5664 * Clamp the batch to a 2^n - 1 value. Having a power
5665 * of 2 value was found to be more likely to have
5666 * suboptimal cache aliasing properties in some cases.
5668 * For example if 2 tasks are alternately allocating
5669 * batches of pages, one task can end up with a lot
5670 * of pages of one half of the possible page colors
5671 * and the other with pages of the other colors.
5673 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5678 /* The deferral and batching of frees should be suppressed under NOMMU
5681 * The problem is that NOMMU needs to be able to allocate large chunks
5682 * of contiguous memory as there's no hardware page translation to
5683 * assemble apparent contiguous memory from discontiguous pages.
5685 * Queueing large contiguous runs of pages for batching, however,
5686 * causes the pages to actually be freed in smaller chunks. As there
5687 * can be a significant delay between the individual batches being
5688 * recycled, this leads to the once large chunks of space being
5689 * fragmented and becoming unavailable for high-order allocations.
5696 * pcp->high and pcp->batch values are related and dependent on one another:
5697 * ->batch must never be higher then ->high.
5698 * The following function updates them in a safe manner without read side
5701 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5702 * those fields changing asynchronously (acording the the above rule).
5704 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5705 * outside of boot time (or some other assurance that no concurrent updaters
5708 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5709 unsigned long batch)
5711 /* start with a fail safe value for batch */
5715 /* Update high, then batch, in order */
5722 /* a companion to pageset_set_high() */
5723 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5725 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5728 static void pageset_init(struct per_cpu_pageset *p)
5730 struct per_cpu_pages *pcp;
5733 memset(p, 0, sizeof(*p));
5736 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5737 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5740 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5743 pageset_set_batch(p, batch);
5747 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5748 * to the value high for the pageset p.
5750 static void pageset_set_high(struct per_cpu_pageset *p,
5753 unsigned long batch = max(1UL, high / 4);
5754 if ((high / 4) > (PAGE_SHIFT * 8))
5755 batch = PAGE_SHIFT * 8;
5757 pageset_update(&p->pcp, high, batch);
5760 static void pageset_set_high_and_batch(struct zone *zone,
5761 struct per_cpu_pageset *pcp)
5763 if (percpu_pagelist_fraction)
5764 pageset_set_high(pcp,
5765 (zone_managed_pages(zone) /
5766 percpu_pagelist_fraction));
5768 pageset_set_batch(pcp, zone_batchsize(zone));
5771 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5773 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5776 pageset_set_high_and_batch(zone, pcp);
5779 void __meminit setup_zone_pageset(struct zone *zone)
5782 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5783 for_each_possible_cpu(cpu)
5784 zone_pageset_init(zone, cpu);
5788 * Allocate per cpu pagesets and initialize them.
5789 * Before this call only boot pagesets were available.
5791 void __init setup_per_cpu_pageset(void)
5793 struct pglist_data *pgdat;
5796 for_each_populated_zone(zone)
5797 setup_zone_pageset(zone);
5799 for_each_online_pgdat(pgdat)
5800 pgdat->per_cpu_nodestats =
5801 alloc_percpu(struct per_cpu_nodestat);
5804 static __meminit void zone_pcp_init(struct zone *zone)
5807 * per cpu subsystem is not up at this point. The following code
5808 * relies on the ability of the linker to provide the
5809 * offset of a (static) per cpu variable into the per cpu area.
5811 zone->pageset = &boot_pageset;
5813 if (populated_zone(zone))
5814 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5815 zone->name, zone->present_pages,
5816 zone_batchsize(zone));
5819 void __meminit init_currently_empty_zone(struct zone *zone,
5820 unsigned long zone_start_pfn,
5823 struct pglist_data *pgdat = zone->zone_pgdat;
5824 int zone_idx = zone_idx(zone) + 1;
5826 if (zone_idx > pgdat->nr_zones)
5827 pgdat->nr_zones = zone_idx;
5829 zone->zone_start_pfn = zone_start_pfn;
5831 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5832 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5834 (unsigned long)zone_idx(zone),
5835 zone_start_pfn, (zone_start_pfn + size));
5837 zone_init_free_lists(zone);
5838 zone->initialized = 1;
5841 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5842 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5845 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5847 int __meminit __early_pfn_to_nid(unsigned long pfn,
5848 struct mminit_pfnnid_cache *state)
5850 unsigned long start_pfn, end_pfn;
5853 if (state->last_start <= pfn && pfn < state->last_end)
5854 return state->last_nid;
5856 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5858 state->last_start = start_pfn;
5859 state->last_end = end_pfn;
5860 state->last_nid = nid;
5865 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5868 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5869 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5870 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5872 * If an architecture guarantees that all ranges registered contain no holes
5873 * and may be freed, this this function may be used instead of calling
5874 * memblock_free_early_nid() manually.
5876 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5878 unsigned long start_pfn, end_pfn;
5881 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5882 start_pfn = min(start_pfn, max_low_pfn);
5883 end_pfn = min(end_pfn, max_low_pfn);
5885 if (start_pfn < end_pfn)
5886 memblock_free_early_nid(PFN_PHYS(start_pfn),
5887 (end_pfn - start_pfn) << PAGE_SHIFT,
5893 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5894 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5896 * If an architecture guarantees that all ranges registered contain no holes and may
5897 * be freed, this function may be used instead of calling memory_present() manually.
5899 void __init sparse_memory_present_with_active_regions(int nid)
5901 unsigned long start_pfn, end_pfn;
5904 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5905 memory_present(this_nid, start_pfn, end_pfn);
5909 * get_pfn_range_for_nid - Return the start and end page frames for a node
5910 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5911 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5912 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5914 * It returns the start and end page frame of a node based on information
5915 * provided by memblock_set_node(). If called for a node
5916 * with no available memory, a warning is printed and the start and end
5919 void __meminit get_pfn_range_for_nid(unsigned int nid,
5920 unsigned long *start_pfn, unsigned long *end_pfn)
5922 unsigned long this_start_pfn, this_end_pfn;
5928 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5929 *start_pfn = min(*start_pfn, this_start_pfn);
5930 *end_pfn = max(*end_pfn, this_end_pfn);
5933 if (*start_pfn == -1UL)
5938 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5939 * assumption is made that zones within a node are ordered in monotonic
5940 * increasing memory addresses so that the "highest" populated zone is used
5942 static void __init find_usable_zone_for_movable(void)
5945 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5946 if (zone_index == ZONE_MOVABLE)
5949 if (arch_zone_highest_possible_pfn[zone_index] >
5950 arch_zone_lowest_possible_pfn[zone_index])
5954 VM_BUG_ON(zone_index == -1);
5955 movable_zone = zone_index;
5959 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5960 * because it is sized independent of architecture. Unlike the other zones,
5961 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5962 * in each node depending on the size of each node and how evenly kernelcore
5963 * is distributed. This helper function adjusts the zone ranges
5964 * provided by the architecture for a given node by using the end of the
5965 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5966 * zones within a node are in order of monotonic increases memory addresses
5968 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5969 unsigned long zone_type,
5970 unsigned long node_start_pfn,
5971 unsigned long node_end_pfn,
5972 unsigned long *zone_start_pfn,
5973 unsigned long *zone_end_pfn)
5975 /* Only adjust if ZONE_MOVABLE is on this node */
5976 if (zone_movable_pfn[nid]) {
5977 /* Size ZONE_MOVABLE */
5978 if (zone_type == ZONE_MOVABLE) {
5979 *zone_start_pfn = zone_movable_pfn[nid];
5980 *zone_end_pfn = min(node_end_pfn,
5981 arch_zone_highest_possible_pfn[movable_zone]);
5983 /* Adjust for ZONE_MOVABLE starting within this range */
5984 } else if (!mirrored_kernelcore &&
5985 *zone_start_pfn < zone_movable_pfn[nid] &&
5986 *zone_end_pfn > zone_movable_pfn[nid]) {
5987 *zone_end_pfn = zone_movable_pfn[nid];
5989 /* Check if this whole range is within ZONE_MOVABLE */
5990 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5991 *zone_start_pfn = *zone_end_pfn;
5996 * Return the number of pages a zone spans in a node, including holes
5997 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5999 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
6000 unsigned long zone_type,
6001 unsigned long node_start_pfn,
6002 unsigned long node_end_pfn,
6003 unsigned long *zone_start_pfn,
6004 unsigned long *zone_end_pfn,
6005 unsigned long *ignored)
6007 /* When hotadd a new node from cpu_up(), the node should be empty */
6008 if (!node_start_pfn && !node_end_pfn)
6011 /* Get the start and end of the zone */
6012 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6013 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6014 adjust_zone_range_for_zone_movable(nid, zone_type,
6015 node_start_pfn, node_end_pfn,
6016 zone_start_pfn, zone_end_pfn);
6018 /* Check that this node has pages within the zone's required range */
6019 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6022 /* Move the zone boundaries inside the node if necessary */
6023 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6024 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6026 /* Return the spanned pages */
6027 return *zone_end_pfn - *zone_start_pfn;
6031 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6032 * then all holes in the requested range will be accounted for.
6034 unsigned long __meminit __absent_pages_in_range(int nid,
6035 unsigned long range_start_pfn,
6036 unsigned long range_end_pfn)
6038 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6039 unsigned long start_pfn, end_pfn;
6042 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6043 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6044 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6045 nr_absent -= end_pfn - start_pfn;
6051 * absent_pages_in_range - Return number of page frames in holes within a range
6052 * @start_pfn: The start PFN to start searching for holes
6053 * @end_pfn: The end PFN to stop searching for holes
6055 * It returns the number of pages frames in memory holes within a range.
6057 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6058 unsigned long end_pfn)
6060 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6063 /* Return the number of page frames in holes in a zone on a node */
6064 static unsigned long __meminit zone_absent_pages_in_node(int nid,
6065 unsigned long zone_type,
6066 unsigned long node_start_pfn,
6067 unsigned long node_end_pfn,
6068 unsigned long *ignored)
6070 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6071 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6072 unsigned long zone_start_pfn, zone_end_pfn;
6073 unsigned long nr_absent;
6075 /* When hotadd a new node from cpu_up(), the node should be empty */
6076 if (!node_start_pfn && !node_end_pfn)
6079 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6080 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6082 adjust_zone_range_for_zone_movable(nid, zone_type,
6083 node_start_pfn, node_end_pfn,
6084 &zone_start_pfn, &zone_end_pfn);
6085 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6088 * ZONE_MOVABLE handling.
6089 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6092 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6093 unsigned long start_pfn, end_pfn;
6094 struct memblock_region *r;
6096 for_each_memblock(memory, r) {
6097 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6098 zone_start_pfn, zone_end_pfn);
6099 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6100 zone_start_pfn, zone_end_pfn);
6102 if (zone_type == ZONE_MOVABLE &&
6103 memblock_is_mirror(r))
6104 nr_absent += end_pfn - start_pfn;
6106 if (zone_type == ZONE_NORMAL &&
6107 !memblock_is_mirror(r))
6108 nr_absent += end_pfn - start_pfn;
6115 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6116 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6117 unsigned long zone_type,
6118 unsigned long node_start_pfn,
6119 unsigned long node_end_pfn,
6120 unsigned long *zone_start_pfn,
6121 unsigned long *zone_end_pfn,
6122 unsigned long *zones_size)
6126 *zone_start_pfn = node_start_pfn;
6127 for (zone = 0; zone < zone_type; zone++)
6128 *zone_start_pfn += zones_size[zone];
6130 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6132 return zones_size[zone_type];
6135 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6136 unsigned long zone_type,
6137 unsigned long node_start_pfn,
6138 unsigned long node_end_pfn,
6139 unsigned long *zholes_size)
6144 return zholes_size[zone_type];
6147 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6149 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6150 unsigned long node_start_pfn,
6151 unsigned long node_end_pfn,
6152 unsigned long *zones_size,
6153 unsigned long *zholes_size)
6155 unsigned long realtotalpages = 0, totalpages = 0;
6158 for (i = 0; i < MAX_NR_ZONES; i++) {
6159 struct zone *zone = pgdat->node_zones + i;
6160 unsigned long zone_start_pfn, zone_end_pfn;
6161 unsigned long size, real_size;
6163 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6169 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6170 node_start_pfn, node_end_pfn,
6173 zone->zone_start_pfn = zone_start_pfn;
6175 zone->zone_start_pfn = 0;
6176 zone->spanned_pages = size;
6177 zone->present_pages = real_size;
6180 realtotalpages += real_size;
6183 pgdat->node_spanned_pages = totalpages;
6184 pgdat->node_present_pages = realtotalpages;
6185 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6189 #ifndef CONFIG_SPARSEMEM
6191 * Calculate the size of the zone->blockflags rounded to an unsigned long
6192 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6193 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6194 * round what is now in bits to nearest long in bits, then return it in
6197 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6199 unsigned long usemapsize;
6201 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6202 usemapsize = roundup(zonesize, pageblock_nr_pages);
6203 usemapsize = usemapsize >> pageblock_order;
6204 usemapsize *= NR_PAGEBLOCK_BITS;
6205 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6207 return usemapsize / 8;
6210 static void __ref setup_usemap(struct pglist_data *pgdat,
6212 unsigned long zone_start_pfn,
6213 unsigned long zonesize)
6215 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6216 zone->pageblock_flags = NULL;
6218 zone->pageblock_flags =
6219 memblock_alloc_node_nopanic(usemapsize,
6223 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6224 unsigned long zone_start_pfn, unsigned long zonesize) {}
6225 #endif /* CONFIG_SPARSEMEM */
6227 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6229 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6230 void __init set_pageblock_order(void)
6234 /* Check that pageblock_nr_pages has not already been setup */
6235 if (pageblock_order)
6238 if (HPAGE_SHIFT > PAGE_SHIFT)
6239 order = HUGETLB_PAGE_ORDER;
6241 order = MAX_ORDER - 1;
6244 * Assume the largest contiguous order of interest is a huge page.
6245 * This value may be variable depending on boot parameters on IA64 and
6248 pageblock_order = order;
6250 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6253 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6254 * is unused as pageblock_order is set at compile-time. See
6255 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6258 void __init set_pageblock_order(void)
6262 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6264 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6265 unsigned long present_pages)
6267 unsigned long pages = spanned_pages;
6270 * Provide a more accurate estimation if there are holes within
6271 * the zone and SPARSEMEM is in use. If there are holes within the
6272 * zone, each populated memory region may cost us one or two extra
6273 * memmap pages due to alignment because memmap pages for each
6274 * populated regions may not be naturally aligned on page boundary.
6275 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6277 if (spanned_pages > present_pages + (present_pages >> 4) &&
6278 IS_ENABLED(CONFIG_SPARSEMEM))
6279 pages = present_pages;
6281 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6284 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6285 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6287 spin_lock_init(&pgdat->split_queue_lock);
6288 INIT_LIST_HEAD(&pgdat->split_queue);
6289 pgdat->split_queue_len = 0;
6292 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6295 #ifdef CONFIG_COMPACTION
6296 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6298 init_waitqueue_head(&pgdat->kcompactd_wait);
6301 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6304 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6306 pgdat_resize_init(pgdat);
6308 pgdat_init_split_queue(pgdat);
6309 pgdat_init_kcompactd(pgdat);
6311 init_waitqueue_head(&pgdat->kswapd_wait);
6312 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6314 pgdat_page_ext_init(pgdat);
6315 spin_lock_init(&pgdat->lru_lock);
6316 lruvec_init(node_lruvec(pgdat));
6319 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6320 unsigned long remaining_pages)
6322 atomic_long_set(&zone->managed_pages, remaining_pages);
6323 zone_set_nid(zone, nid);
6324 zone->name = zone_names[idx];
6325 zone->zone_pgdat = NODE_DATA(nid);
6326 spin_lock_init(&zone->lock);
6327 zone_seqlock_init(zone);
6328 zone_pcp_init(zone);
6332 * Set up the zone data structures
6333 * - init pgdat internals
6334 * - init all zones belonging to this node
6336 * NOTE: this function is only called during memory hotplug
6338 #ifdef CONFIG_MEMORY_HOTPLUG
6339 void __ref free_area_init_core_hotplug(int nid)
6342 pg_data_t *pgdat = NODE_DATA(nid);
6344 pgdat_init_internals(pgdat);
6345 for (z = 0; z < MAX_NR_ZONES; z++)
6346 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6351 * Set up the zone data structures:
6352 * - mark all pages reserved
6353 * - mark all memory queues empty
6354 * - clear the memory bitmaps
6356 * NOTE: pgdat should get zeroed by caller.
6357 * NOTE: this function is only called during early init.
6359 static void __init free_area_init_core(struct pglist_data *pgdat)
6362 int nid = pgdat->node_id;
6364 pgdat_init_internals(pgdat);
6365 pgdat->per_cpu_nodestats = &boot_nodestats;
6367 for (j = 0; j < MAX_NR_ZONES; j++) {
6368 struct zone *zone = pgdat->node_zones + j;
6369 unsigned long size, freesize, memmap_pages;
6370 unsigned long zone_start_pfn = zone->zone_start_pfn;
6372 size = zone->spanned_pages;
6373 freesize = zone->present_pages;
6376 * Adjust freesize so that it accounts for how much memory
6377 * is used by this zone for memmap. This affects the watermark
6378 * and per-cpu initialisations
6380 memmap_pages = calc_memmap_size(size, freesize);
6381 if (!is_highmem_idx(j)) {
6382 if (freesize >= memmap_pages) {
6383 freesize -= memmap_pages;
6386 " %s zone: %lu pages used for memmap\n",
6387 zone_names[j], memmap_pages);
6389 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6390 zone_names[j], memmap_pages, freesize);
6393 /* Account for reserved pages */
6394 if (j == 0 && freesize > dma_reserve) {
6395 freesize -= dma_reserve;
6396 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6397 zone_names[0], dma_reserve);
6400 if (!is_highmem_idx(j))
6401 nr_kernel_pages += freesize;
6402 /* Charge for highmem memmap if there are enough kernel pages */
6403 else if (nr_kernel_pages > memmap_pages * 2)
6404 nr_kernel_pages -= memmap_pages;
6405 nr_all_pages += freesize;
6408 * Set an approximate value for lowmem here, it will be adjusted
6409 * when the bootmem allocator frees pages into the buddy system.
6410 * And all highmem pages will be managed by the buddy system.
6412 zone_init_internals(zone, j, nid, freesize);
6417 set_pageblock_order();
6418 setup_usemap(pgdat, zone, zone_start_pfn, size);
6419 init_currently_empty_zone(zone, zone_start_pfn, size);
6420 memmap_init(size, nid, j, zone_start_pfn);
6424 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6425 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6427 unsigned long __maybe_unused start = 0;
6428 unsigned long __maybe_unused offset = 0;
6430 /* Skip empty nodes */
6431 if (!pgdat->node_spanned_pages)
6434 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6435 offset = pgdat->node_start_pfn - start;
6436 /* ia64 gets its own node_mem_map, before this, without bootmem */
6437 if (!pgdat->node_mem_map) {
6438 unsigned long size, end;
6442 * The zone's endpoints aren't required to be MAX_ORDER
6443 * aligned but the node_mem_map endpoints must be in order
6444 * for the buddy allocator to function correctly.
6446 end = pgdat_end_pfn(pgdat);
6447 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6448 size = (end - start) * sizeof(struct page);
6449 map = memblock_alloc_node_nopanic(size, pgdat->node_id);
6450 pgdat->node_mem_map = map + offset;
6452 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6453 __func__, pgdat->node_id, (unsigned long)pgdat,
6454 (unsigned long)pgdat->node_mem_map);
6455 #ifndef CONFIG_NEED_MULTIPLE_NODES
6457 * With no DISCONTIG, the global mem_map is just set as node 0's
6459 if (pgdat == NODE_DATA(0)) {
6460 mem_map = NODE_DATA(0)->node_mem_map;
6461 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6462 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6464 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6469 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6470 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6472 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6473 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6476 * We start only with one section of pages, more pages are added as
6477 * needed until the rest of deferred pages are initialized.
6479 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6480 pgdat->node_spanned_pages);
6481 pgdat->first_deferred_pfn = ULONG_MAX;
6484 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6487 void __init free_area_init_node(int nid, unsigned long *zones_size,
6488 unsigned long node_start_pfn,
6489 unsigned long *zholes_size)
6491 pg_data_t *pgdat = NODE_DATA(nid);
6492 unsigned long start_pfn = 0;
6493 unsigned long end_pfn = 0;
6495 /* pg_data_t should be reset to zero when it's allocated */
6496 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6498 pgdat->node_id = nid;
6499 pgdat->node_start_pfn = node_start_pfn;
6500 pgdat->per_cpu_nodestats = NULL;
6501 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6502 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6503 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6504 (u64)start_pfn << PAGE_SHIFT,
6505 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6507 start_pfn = node_start_pfn;
6509 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6510 zones_size, zholes_size);
6512 alloc_node_mem_map(pgdat);
6513 pgdat_set_deferred_range(pgdat);
6515 free_area_init_core(pgdat);
6518 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6520 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6523 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6528 for (pfn = spfn; pfn < epfn; pfn++) {
6529 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6530 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6531 + pageblock_nr_pages - 1;
6534 mm_zero_struct_page(pfn_to_page(pfn));
6542 * Only struct pages that are backed by physical memory are zeroed and
6543 * initialized by going through __init_single_page(). But, there are some
6544 * struct pages which are reserved in memblock allocator and their fields
6545 * may be accessed (for example page_to_pfn() on some configuration accesses
6546 * flags). We must explicitly zero those struct pages.
6548 * This function also addresses a similar issue where struct pages are left
6549 * uninitialized because the physical address range is not covered by
6550 * memblock.memory or memblock.reserved. That could happen when memblock
6551 * layout is manually configured via memmap=.
6553 void __init zero_resv_unavail(void)
6555 phys_addr_t start, end;
6557 phys_addr_t next = 0;
6560 * Loop through unavailable ranges not covered by memblock.memory.
6563 for_each_mem_range(i, &memblock.memory, NULL,
6564 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6566 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6569 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6572 * Struct pages that do not have backing memory. This could be because
6573 * firmware is using some of this memory, or for some other reasons.
6576 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6578 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6580 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6582 #if MAX_NUMNODES > 1
6584 * Figure out the number of possible node ids.
6586 void __init setup_nr_node_ids(void)
6588 unsigned int highest;
6590 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6591 nr_node_ids = highest + 1;
6596 * node_map_pfn_alignment - determine the maximum internode alignment
6598 * This function should be called after node map is populated and sorted.
6599 * It calculates the maximum power of two alignment which can distinguish
6602 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6603 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6604 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6605 * shifted, 1GiB is enough and this function will indicate so.
6607 * This is used to test whether pfn -> nid mapping of the chosen memory
6608 * model has fine enough granularity to avoid incorrect mapping for the
6609 * populated node map.
6611 * Returns the determined alignment in pfn's. 0 if there is no alignment
6612 * requirement (single node).
6614 unsigned long __init node_map_pfn_alignment(void)
6616 unsigned long accl_mask = 0, last_end = 0;
6617 unsigned long start, end, mask;
6621 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6622 if (!start || last_nid < 0 || last_nid == nid) {
6629 * Start with a mask granular enough to pin-point to the
6630 * start pfn and tick off bits one-by-one until it becomes
6631 * too coarse to separate the current node from the last.
6633 mask = ~((1 << __ffs(start)) - 1);
6634 while (mask && last_end <= (start & (mask << 1)))
6637 /* accumulate all internode masks */
6641 /* convert mask to number of pages */
6642 return ~accl_mask + 1;
6645 /* Find the lowest pfn for a node */
6646 static unsigned long __init find_min_pfn_for_node(int nid)
6648 unsigned long min_pfn = ULONG_MAX;
6649 unsigned long start_pfn;
6652 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6653 min_pfn = min(min_pfn, start_pfn);
6655 if (min_pfn == ULONG_MAX) {
6656 pr_warn("Could not find start_pfn for node %d\n", nid);
6664 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6666 * It returns the minimum PFN based on information provided via
6667 * memblock_set_node().
6669 unsigned long __init find_min_pfn_with_active_regions(void)
6671 return find_min_pfn_for_node(MAX_NUMNODES);
6675 * early_calculate_totalpages()
6676 * Sum pages in active regions for movable zone.
6677 * Populate N_MEMORY for calculating usable_nodes.
6679 static unsigned long __init early_calculate_totalpages(void)
6681 unsigned long totalpages = 0;
6682 unsigned long start_pfn, end_pfn;
6685 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6686 unsigned long pages = end_pfn - start_pfn;
6688 totalpages += pages;
6690 node_set_state(nid, N_MEMORY);
6696 * Find the PFN the Movable zone begins in each node. Kernel memory
6697 * is spread evenly between nodes as long as the nodes have enough
6698 * memory. When they don't, some nodes will have more kernelcore than
6701 static void __init find_zone_movable_pfns_for_nodes(void)
6704 unsigned long usable_startpfn;
6705 unsigned long kernelcore_node, kernelcore_remaining;
6706 /* save the state before borrow the nodemask */
6707 nodemask_t saved_node_state = node_states[N_MEMORY];
6708 unsigned long totalpages = early_calculate_totalpages();
6709 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6710 struct memblock_region *r;
6712 /* Need to find movable_zone earlier when movable_node is specified. */
6713 find_usable_zone_for_movable();
6716 * If movable_node is specified, ignore kernelcore and movablecore
6719 if (movable_node_is_enabled()) {
6720 for_each_memblock(memory, r) {
6721 if (!memblock_is_hotpluggable(r))
6726 usable_startpfn = PFN_DOWN(r->base);
6727 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6728 min(usable_startpfn, zone_movable_pfn[nid]) :
6736 * If kernelcore=mirror is specified, ignore movablecore option
6738 if (mirrored_kernelcore) {
6739 bool mem_below_4gb_not_mirrored = false;
6741 for_each_memblock(memory, r) {
6742 if (memblock_is_mirror(r))
6747 usable_startpfn = memblock_region_memory_base_pfn(r);
6749 if (usable_startpfn < 0x100000) {
6750 mem_below_4gb_not_mirrored = true;
6754 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6755 min(usable_startpfn, zone_movable_pfn[nid]) :
6759 if (mem_below_4gb_not_mirrored)
6760 pr_warn("This configuration results in unmirrored kernel memory.");
6766 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6767 * amount of necessary memory.
6769 if (required_kernelcore_percent)
6770 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6772 if (required_movablecore_percent)
6773 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6777 * If movablecore= was specified, calculate what size of
6778 * kernelcore that corresponds so that memory usable for
6779 * any allocation type is evenly spread. If both kernelcore
6780 * and movablecore are specified, then the value of kernelcore
6781 * will be used for required_kernelcore if it's greater than
6782 * what movablecore would have allowed.
6784 if (required_movablecore) {
6785 unsigned long corepages;
6788 * Round-up so that ZONE_MOVABLE is at least as large as what
6789 * was requested by the user
6791 required_movablecore =
6792 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6793 required_movablecore = min(totalpages, required_movablecore);
6794 corepages = totalpages - required_movablecore;
6796 required_kernelcore = max(required_kernelcore, corepages);
6800 * If kernelcore was not specified or kernelcore size is larger
6801 * than totalpages, there is no ZONE_MOVABLE.
6803 if (!required_kernelcore || required_kernelcore >= totalpages)
6806 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6807 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6810 /* Spread kernelcore memory as evenly as possible throughout nodes */
6811 kernelcore_node = required_kernelcore / usable_nodes;
6812 for_each_node_state(nid, N_MEMORY) {
6813 unsigned long start_pfn, end_pfn;
6816 * Recalculate kernelcore_node if the division per node
6817 * now exceeds what is necessary to satisfy the requested
6818 * amount of memory for the kernel
6820 if (required_kernelcore < kernelcore_node)
6821 kernelcore_node = required_kernelcore / usable_nodes;
6824 * As the map is walked, we track how much memory is usable
6825 * by the kernel using kernelcore_remaining. When it is
6826 * 0, the rest of the node is usable by ZONE_MOVABLE
6828 kernelcore_remaining = kernelcore_node;
6830 /* Go through each range of PFNs within this node */
6831 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6832 unsigned long size_pages;
6834 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6835 if (start_pfn >= end_pfn)
6838 /* Account for what is only usable for kernelcore */
6839 if (start_pfn < usable_startpfn) {
6840 unsigned long kernel_pages;
6841 kernel_pages = min(end_pfn, usable_startpfn)
6844 kernelcore_remaining -= min(kernel_pages,
6845 kernelcore_remaining);
6846 required_kernelcore -= min(kernel_pages,
6847 required_kernelcore);
6849 /* Continue if range is now fully accounted */
6850 if (end_pfn <= usable_startpfn) {
6853 * Push zone_movable_pfn to the end so
6854 * that if we have to rebalance
6855 * kernelcore across nodes, we will
6856 * not double account here
6858 zone_movable_pfn[nid] = end_pfn;
6861 start_pfn = usable_startpfn;
6865 * The usable PFN range for ZONE_MOVABLE is from
6866 * start_pfn->end_pfn. Calculate size_pages as the
6867 * number of pages used as kernelcore
6869 size_pages = end_pfn - start_pfn;
6870 if (size_pages > kernelcore_remaining)
6871 size_pages = kernelcore_remaining;
6872 zone_movable_pfn[nid] = start_pfn + size_pages;
6875 * Some kernelcore has been met, update counts and
6876 * break if the kernelcore for this node has been
6879 required_kernelcore -= min(required_kernelcore,
6881 kernelcore_remaining -= size_pages;
6882 if (!kernelcore_remaining)
6888 * If there is still required_kernelcore, we do another pass with one
6889 * less node in the count. This will push zone_movable_pfn[nid] further
6890 * along on the nodes that still have memory until kernelcore is
6894 if (usable_nodes && required_kernelcore > usable_nodes)
6898 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6899 for (nid = 0; nid < MAX_NUMNODES; nid++)
6900 zone_movable_pfn[nid] =
6901 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6904 /* restore the node_state */
6905 node_states[N_MEMORY] = saved_node_state;
6908 /* Any regular or high memory on that node ? */
6909 static void check_for_memory(pg_data_t *pgdat, int nid)
6911 enum zone_type zone_type;
6913 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6914 struct zone *zone = &pgdat->node_zones[zone_type];
6915 if (populated_zone(zone)) {
6916 if (IS_ENABLED(CONFIG_HIGHMEM))
6917 node_set_state(nid, N_HIGH_MEMORY);
6918 if (zone_type <= ZONE_NORMAL)
6919 node_set_state(nid, N_NORMAL_MEMORY);
6926 * free_area_init_nodes - Initialise all pg_data_t and zone data
6927 * @max_zone_pfn: an array of max PFNs for each zone
6929 * This will call free_area_init_node() for each active node in the system.
6930 * Using the page ranges provided by memblock_set_node(), the size of each
6931 * zone in each node and their holes is calculated. If the maximum PFN
6932 * between two adjacent zones match, it is assumed that the zone is empty.
6933 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6934 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6935 * starts where the previous one ended. For example, ZONE_DMA32 starts
6936 * at arch_max_dma_pfn.
6938 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6940 unsigned long start_pfn, end_pfn;
6943 /* Record where the zone boundaries are */
6944 memset(arch_zone_lowest_possible_pfn, 0,
6945 sizeof(arch_zone_lowest_possible_pfn));
6946 memset(arch_zone_highest_possible_pfn, 0,
6947 sizeof(arch_zone_highest_possible_pfn));
6949 start_pfn = find_min_pfn_with_active_regions();
6951 for (i = 0; i < MAX_NR_ZONES; i++) {
6952 if (i == ZONE_MOVABLE)
6955 end_pfn = max(max_zone_pfn[i], start_pfn);
6956 arch_zone_lowest_possible_pfn[i] = start_pfn;
6957 arch_zone_highest_possible_pfn[i] = end_pfn;
6959 start_pfn = end_pfn;
6962 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6963 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6964 find_zone_movable_pfns_for_nodes();
6966 /* Print out the zone ranges */
6967 pr_info("Zone ranges:\n");
6968 for (i = 0; i < MAX_NR_ZONES; i++) {
6969 if (i == ZONE_MOVABLE)
6971 pr_info(" %-8s ", zone_names[i]);
6972 if (arch_zone_lowest_possible_pfn[i] ==
6973 arch_zone_highest_possible_pfn[i])
6976 pr_cont("[mem %#018Lx-%#018Lx]\n",
6977 (u64)arch_zone_lowest_possible_pfn[i]
6979 ((u64)arch_zone_highest_possible_pfn[i]
6980 << PAGE_SHIFT) - 1);
6983 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6984 pr_info("Movable zone start for each node\n");
6985 for (i = 0; i < MAX_NUMNODES; i++) {
6986 if (zone_movable_pfn[i])
6987 pr_info(" Node %d: %#018Lx\n", i,
6988 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6991 /* Print out the early node map */
6992 pr_info("Early memory node ranges\n");
6993 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6994 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6995 (u64)start_pfn << PAGE_SHIFT,
6996 ((u64)end_pfn << PAGE_SHIFT) - 1);
6998 /* Initialise every node */
6999 mminit_verify_pageflags_layout();
7000 setup_nr_node_ids();
7001 zero_resv_unavail();
7002 for_each_online_node(nid) {
7003 pg_data_t *pgdat = NODE_DATA(nid);
7004 free_area_init_node(nid, NULL,
7005 find_min_pfn_for_node(nid), NULL);
7007 /* Any memory on that node */
7008 if (pgdat->node_present_pages)
7009 node_set_state(nid, N_MEMORY);
7010 check_for_memory(pgdat, nid);
7014 static int __init cmdline_parse_core(char *p, unsigned long *core,
7015 unsigned long *percent)
7017 unsigned long long coremem;
7023 /* Value may be a percentage of total memory, otherwise bytes */
7024 coremem = simple_strtoull(p, &endptr, 0);
7025 if (*endptr == '%') {
7026 /* Paranoid check for percent values greater than 100 */
7027 WARN_ON(coremem > 100);
7031 coremem = memparse(p, &p);
7032 /* Paranoid check that UL is enough for the coremem value */
7033 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7035 *core = coremem >> PAGE_SHIFT;
7042 * kernelcore=size sets the amount of memory for use for allocations that
7043 * cannot be reclaimed or migrated.
7045 static int __init cmdline_parse_kernelcore(char *p)
7047 /* parse kernelcore=mirror */
7048 if (parse_option_str(p, "mirror")) {
7049 mirrored_kernelcore = true;
7053 return cmdline_parse_core(p, &required_kernelcore,
7054 &required_kernelcore_percent);
7058 * movablecore=size sets the amount of memory for use for allocations that
7059 * can be reclaimed or migrated.
7061 static int __init cmdline_parse_movablecore(char *p)
7063 return cmdline_parse_core(p, &required_movablecore,
7064 &required_movablecore_percent);
7067 early_param("kernelcore", cmdline_parse_kernelcore);
7068 early_param("movablecore", cmdline_parse_movablecore);
7070 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7072 void adjust_managed_page_count(struct page *page, long count)
7074 atomic_long_add(count, &page_zone(page)->managed_pages);
7075 totalram_pages_add(count);
7076 #ifdef CONFIG_HIGHMEM
7077 if (PageHighMem(page))
7078 totalhigh_pages_add(count);
7081 EXPORT_SYMBOL(adjust_managed_page_count);
7083 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
7086 unsigned long pages = 0;
7088 start = (void *)PAGE_ALIGN((unsigned long)start);
7089 end = (void *)((unsigned long)end & PAGE_MASK);
7090 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7091 struct page *page = virt_to_page(pos);
7092 void *direct_map_addr;
7095 * 'direct_map_addr' might be different from 'pos'
7096 * because some architectures' virt_to_page()
7097 * work with aliases. Getting the direct map
7098 * address ensures that we get a _writeable_
7099 * alias for the memset().
7101 direct_map_addr = page_address(page);
7102 if ((unsigned int)poison <= 0xFF)
7103 memset(direct_map_addr, poison, PAGE_SIZE);
7105 free_reserved_page(page);
7109 pr_info("Freeing %s memory: %ldK\n",
7110 s, pages << (PAGE_SHIFT - 10));
7114 EXPORT_SYMBOL(free_reserved_area);
7116 #ifdef CONFIG_HIGHMEM
7117 void free_highmem_page(struct page *page)
7119 __free_reserved_page(page);
7120 totalram_pages_inc();
7121 atomic_long_inc(&page_zone(page)->managed_pages);
7122 totalhigh_pages_inc();
7127 void __init mem_init_print_info(const char *str)
7129 unsigned long physpages, codesize, datasize, rosize, bss_size;
7130 unsigned long init_code_size, init_data_size;
7132 physpages = get_num_physpages();
7133 codesize = _etext - _stext;
7134 datasize = _edata - _sdata;
7135 rosize = __end_rodata - __start_rodata;
7136 bss_size = __bss_stop - __bss_start;
7137 init_data_size = __init_end - __init_begin;
7138 init_code_size = _einittext - _sinittext;
7141 * Detect special cases and adjust section sizes accordingly:
7142 * 1) .init.* may be embedded into .data sections
7143 * 2) .init.text.* may be out of [__init_begin, __init_end],
7144 * please refer to arch/tile/kernel/vmlinux.lds.S.
7145 * 3) .rodata.* may be embedded into .text or .data sections.
7147 #define adj_init_size(start, end, size, pos, adj) \
7149 if (start <= pos && pos < end && size > adj) \
7153 adj_init_size(__init_begin, __init_end, init_data_size,
7154 _sinittext, init_code_size);
7155 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7156 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7157 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7158 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7160 #undef adj_init_size
7162 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7163 #ifdef CONFIG_HIGHMEM
7167 nr_free_pages() << (PAGE_SHIFT - 10),
7168 physpages << (PAGE_SHIFT - 10),
7169 codesize >> 10, datasize >> 10, rosize >> 10,
7170 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7171 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7172 totalcma_pages << (PAGE_SHIFT - 10),
7173 #ifdef CONFIG_HIGHMEM
7174 totalhigh_pages() << (PAGE_SHIFT - 10),
7176 str ? ", " : "", str ? str : "");
7180 * set_dma_reserve - set the specified number of pages reserved in the first zone
7181 * @new_dma_reserve: The number of pages to mark reserved
7183 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7184 * In the DMA zone, a significant percentage may be consumed by kernel image
7185 * and other unfreeable allocations which can skew the watermarks badly. This
7186 * function may optionally be used to account for unfreeable pages in the
7187 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7188 * smaller per-cpu batchsize.
7190 void __init set_dma_reserve(unsigned long new_dma_reserve)
7192 dma_reserve = new_dma_reserve;
7195 void __init free_area_init(unsigned long *zones_size)
7197 zero_resv_unavail();
7198 free_area_init_node(0, zones_size,
7199 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7202 static int page_alloc_cpu_dead(unsigned int cpu)
7205 lru_add_drain_cpu(cpu);
7209 * Spill the event counters of the dead processor
7210 * into the current processors event counters.
7211 * This artificially elevates the count of the current
7214 vm_events_fold_cpu(cpu);
7217 * Zero the differential counters of the dead processor
7218 * so that the vm statistics are consistent.
7220 * This is only okay since the processor is dead and cannot
7221 * race with what we are doing.
7223 cpu_vm_stats_fold(cpu);
7227 void __init page_alloc_init(void)
7231 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7232 "mm/page_alloc:dead", NULL,
7233 page_alloc_cpu_dead);
7238 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7239 * or min_free_kbytes changes.
7241 static void calculate_totalreserve_pages(void)
7243 struct pglist_data *pgdat;
7244 unsigned long reserve_pages = 0;
7245 enum zone_type i, j;
7247 for_each_online_pgdat(pgdat) {
7249 pgdat->totalreserve_pages = 0;
7251 for (i = 0; i < MAX_NR_ZONES; i++) {
7252 struct zone *zone = pgdat->node_zones + i;
7254 unsigned long managed_pages = zone_managed_pages(zone);
7256 /* Find valid and maximum lowmem_reserve in the zone */
7257 for (j = i; j < MAX_NR_ZONES; j++) {
7258 if (zone->lowmem_reserve[j] > max)
7259 max = zone->lowmem_reserve[j];
7262 /* we treat the high watermark as reserved pages. */
7263 max += high_wmark_pages(zone);
7265 if (max > managed_pages)
7266 max = managed_pages;
7268 pgdat->totalreserve_pages += max;
7270 reserve_pages += max;
7273 totalreserve_pages = reserve_pages;
7277 * setup_per_zone_lowmem_reserve - called whenever
7278 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7279 * has a correct pages reserved value, so an adequate number of
7280 * pages are left in the zone after a successful __alloc_pages().
7282 static void setup_per_zone_lowmem_reserve(void)
7284 struct pglist_data *pgdat;
7285 enum zone_type j, idx;
7287 for_each_online_pgdat(pgdat) {
7288 for (j = 0; j < MAX_NR_ZONES; j++) {
7289 struct zone *zone = pgdat->node_zones + j;
7290 unsigned long managed_pages = zone_managed_pages(zone);
7292 zone->lowmem_reserve[j] = 0;
7296 struct zone *lower_zone;
7299 lower_zone = pgdat->node_zones + idx;
7301 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7302 sysctl_lowmem_reserve_ratio[idx] = 0;
7303 lower_zone->lowmem_reserve[j] = 0;
7305 lower_zone->lowmem_reserve[j] =
7306 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7308 managed_pages += zone_managed_pages(lower_zone);
7313 /* update totalreserve_pages */
7314 calculate_totalreserve_pages();
7317 static void __setup_per_zone_wmarks(void)
7319 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7320 unsigned long lowmem_pages = 0;
7322 unsigned long flags;
7324 /* Calculate total number of !ZONE_HIGHMEM pages */
7325 for_each_zone(zone) {
7326 if (!is_highmem(zone))
7327 lowmem_pages += zone_managed_pages(zone);
7330 for_each_zone(zone) {
7333 spin_lock_irqsave(&zone->lock, flags);
7334 tmp = (u64)pages_min * zone_managed_pages(zone);
7335 do_div(tmp, lowmem_pages);
7336 if (is_highmem(zone)) {
7338 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7339 * need highmem pages, so cap pages_min to a small
7342 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7343 * deltas control asynch page reclaim, and so should
7344 * not be capped for highmem.
7346 unsigned long min_pages;
7348 min_pages = zone_managed_pages(zone) / 1024;
7349 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7350 zone->watermark[WMARK_MIN] = min_pages;
7353 * If it's a lowmem zone, reserve a number of pages
7354 * proportionate to the zone's size.
7356 zone->watermark[WMARK_MIN] = tmp;
7360 * Set the kswapd watermarks distance according to the
7361 * scale factor in proportion to available memory, but
7362 * ensure a minimum size on small systems.
7364 tmp = max_t(u64, tmp >> 2,
7365 mult_frac(zone_managed_pages(zone),
7366 watermark_scale_factor, 10000));
7368 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7369 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7371 spin_unlock_irqrestore(&zone->lock, flags);
7374 /* update totalreserve_pages */
7375 calculate_totalreserve_pages();
7379 * setup_per_zone_wmarks - called when min_free_kbytes changes
7380 * or when memory is hot-{added|removed}
7382 * Ensures that the watermark[min,low,high] values for each zone are set
7383 * correctly with respect to min_free_kbytes.
7385 void setup_per_zone_wmarks(void)
7387 static DEFINE_SPINLOCK(lock);
7390 __setup_per_zone_wmarks();
7395 * Initialise min_free_kbytes.
7397 * For small machines we want it small (128k min). For large machines
7398 * we want it large (64MB max). But it is not linear, because network
7399 * bandwidth does not increase linearly with machine size. We use
7401 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7402 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7418 int __meminit init_per_zone_wmark_min(void)
7420 unsigned long lowmem_kbytes;
7421 int new_min_free_kbytes;
7423 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7424 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7426 if (new_min_free_kbytes > user_min_free_kbytes) {
7427 min_free_kbytes = new_min_free_kbytes;
7428 if (min_free_kbytes < 128)
7429 min_free_kbytes = 128;
7430 if (min_free_kbytes > 65536)
7431 min_free_kbytes = 65536;
7433 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7434 new_min_free_kbytes, user_min_free_kbytes);
7436 setup_per_zone_wmarks();
7437 refresh_zone_stat_thresholds();
7438 setup_per_zone_lowmem_reserve();
7441 setup_min_unmapped_ratio();
7442 setup_min_slab_ratio();
7447 core_initcall(init_per_zone_wmark_min)
7450 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7451 * that we can call two helper functions whenever min_free_kbytes
7454 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7455 void __user *buffer, size_t *length, loff_t *ppos)
7459 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7464 user_min_free_kbytes = min_free_kbytes;
7465 setup_per_zone_wmarks();
7470 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7471 void __user *buffer, size_t *length, loff_t *ppos)
7475 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7480 setup_per_zone_wmarks();
7486 static void setup_min_unmapped_ratio(void)
7491 for_each_online_pgdat(pgdat)
7492 pgdat->min_unmapped_pages = 0;
7495 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7496 sysctl_min_unmapped_ratio) / 100;
7500 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7501 void __user *buffer, size_t *length, loff_t *ppos)
7505 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7509 setup_min_unmapped_ratio();
7514 static void setup_min_slab_ratio(void)
7519 for_each_online_pgdat(pgdat)
7520 pgdat->min_slab_pages = 0;
7523 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7524 sysctl_min_slab_ratio) / 100;
7527 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7528 void __user *buffer, size_t *length, loff_t *ppos)
7532 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7536 setup_min_slab_ratio();
7543 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7544 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7545 * whenever sysctl_lowmem_reserve_ratio changes.
7547 * The reserve ratio obviously has absolutely no relation with the
7548 * minimum watermarks. The lowmem reserve ratio can only make sense
7549 * if in function of the boot time zone sizes.
7551 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7552 void __user *buffer, size_t *length, loff_t *ppos)
7554 proc_dointvec_minmax(table, write, buffer, length, ppos);
7555 setup_per_zone_lowmem_reserve();
7560 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7561 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7562 * pagelist can have before it gets flushed back to buddy allocator.
7564 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7565 void __user *buffer, size_t *length, loff_t *ppos)
7568 int old_percpu_pagelist_fraction;
7571 mutex_lock(&pcp_batch_high_lock);
7572 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7574 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7575 if (!write || ret < 0)
7578 /* Sanity checking to avoid pcp imbalance */
7579 if (percpu_pagelist_fraction &&
7580 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7581 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7587 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7590 for_each_populated_zone(zone) {
7593 for_each_possible_cpu(cpu)
7594 pageset_set_high_and_batch(zone,
7595 per_cpu_ptr(zone->pageset, cpu));
7598 mutex_unlock(&pcp_batch_high_lock);
7603 int hashdist = HASHDIST_DEFAULT;
7605 static int __init set_hashdist(char *str)
7609 hashdist = simple_strtoul(str, &str, 0);
7612 __setup("hashdist=", set_hashdist);
7615 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7617 * Returns the number of pages that arch has reserved but
7618 * is not known to alloc_large_system_hash().
7620 static unsigned long __init arch_reserved_kernel_pages(void)
7627 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7628 * machines. As memory size is increased the scale is also increased but at
7629 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7630 * quadruples the scale is increased by one, which means the size of hash table
7631 * only doubles, instead of quadrupling as well.
7632 * Because 32-bit systems cannot have large physical memory, where this scaling
7633 * makes sense, it is disabled on such platforms.
7635 #if __BITS_PER_LONG > 32
7636 #define ADAPT_SCALE_BASE (64ul << 30)
7637 #define ADAPT_SCALE_SHIFT 2
7638 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7642 * allocate a large system hash table from bootmem
7643 * - it is assumed that the hash table must contain an exact power-of-2
7644 * quantity of entries
7645 * - limit is the number of hash buckets, not the total allocation size
7647 void *__init alloc_large_system_hash(const char *tablename,
7648 unsigned long bucketsize,
7649 unsigned long numentries,
7652 unsigned int *_hash_shift,
7653 unsigned int *_hash_mask,
7654 unsigned long low_limit,
7655 unsigned long high_limit)
7657 unsigned long long max = high_limit;
7658 unsigned long log2qty, size;
7662 /* allow the kernel cmdline to have a say */
7664 /* round applicable memory size up to nearest megabyte */
7665 numentries = nr_kernel_pages;
7666 numentries -= arch_reserved_kernel_pages();
7668 /* It isn't necessary when PAGE_SIZE >= 1MB */
7669 if (PAGE_SHIFT < 20)
7670 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7672 #if __BITS_PER_LONG > 32
7674 unsigned long adapt;
7676 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7677 adapt <<= ADAPT_SCALE_SHIFT)
7682 /* limit to 1 bucket per 2^scale bytes of low memory */
7683 if (scale > PAGE_SHIFT)
7684 numentries >>= (scale - PAGE_SHIFT);
7686 numentries <<= (PAGE_SHIFT - scale);
7688 /* Make sure we've got at least a 0-order allocation.. */
7689 if (unlikely(flags & HASH_SMALL)) {
7690 /* Makes no sense without HASH_EARLY */
7691 WARN_ON(!(flags & HASH_EARLY));
7692 if (!(numentries >> *_hash_shift)) {
7693 numentries = 1UL << *_hash_shift;
7694 BUG_ON(!numentries);
7696 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7697 numentries = PAGE_SIZE / bucketsize;
7699 numentries = roundup_pow_of_two(numentries);
7701 /* limit allocation size to 1/16 total memory by default */
7703 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7704 do_div(max, bucketsize);
7706 max = min(max, 0x80000000ULL);
7708 if (numentries < low_limit)
7709 numentries = low_limit;
7710 if (numentries > max)
7713 log2qty = ilog2(numentries);
7715 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7717 size = bucketsize << log2qty;
7718 if (flags & HASH_EARLY) {
7719 if (flags & HASH_ZERO)
7720 table = memblock_alloc_nopanic(size,
7723 table = memblock_alloc_raw(size,
7725 } else if (hashdist) {
7726 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7729 * If bucketsize is not a power-of-two, we may free
7730 * some pages at the end of hash table which
7731 * alloc_pages_exact() automatically does
7733 if (get_order(size) < MAX_ORDER) {
7734 table = alloc_pages_exact(size, gfp_flags);
7735 kmemleak_alloc(table, size, 1, gfp_flags);
7738 } while (!table && size > PAGE_SIZE && --log2qty);
7741 panic("Failed to allocate %s hash table\n", tablename);
7743 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7744 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7747 *_hash_shift = log2qty;
7749 *_hash_mask = (1 << log2qty) - 1;
7755 * This function checks whether pageblock includes unmovable pages or not.
7756 * If @count is not zero, it is okay to include less @count unmovable pages
7758 * PageLRU check without isolation or lru_lock could race so that
7759 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7760 * check without lock_page also may miss some movable non-lru pages at
7761 * race condition. So you can't expect this function should be exact.
7763 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7764 int migratetype, int flags)
7766 unsigned long pfn, iter, found;
7769 * TODO we could make this much more efficient by not checking every
7770 * page in the range if we know all of them are in MOVABLE_ZONE and
7771 * that the movable zone guarantees that pages are migratable but
7772 * the later is not the case right now unfortunatelly. E.g. movablecore
7773 * can still lead to having bootmem allocations in zone_movable.
7777 * CMA allocations (alloc_contig_range) really need to mark isolate
7778 * CMA pageblocks even when they are not movable in fact so consider
7779 * them movable here.
7781 if (is_migrate_cma(migratetype) &&
7782 is_migrate_cma(get_pageblock_migratetype(page)))
7785 pfn = page_to_pfn(page);
7786 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7787 unsigned long check = pfn + iter;
7789 if (!pfn_valid_within(check))
7792 page = pfn_to_page(check);
7794 if (PageReserved(page))
7798 * If the zone is movable and we have ruled out all reserved
7799 * pages then it should be reasonably safe to assume the rest
7802 if (zone_idx(zone) == ZONE_MOVABLE)
7806 * Hugepages are not in LRU lists, but they're movable.
7807 * We need not scan over tail pages bacause we don't
7808 * handle each tail page individually in migration.
7810 if (PageHuge(page)) {
7811 struct page *head = compound_head(page);
7812 unsigned int skip_pages;
7814 if (!hugepage_migration_supported(page_hstate(head)))
7817 skip_pages = (1 << compound_order(head)) - (page - head);
7818 iter += skip_pages - 1;
7823 * We can't use page_count without pin a page
7824 * because another CPU can free compound page.
7825 * This check already skips compound tails of THP
7826 * because their page->_refcount is zero at all time.
7828 if (!page_ref_count(page)) {
7829 if (PageBuddy(page))
7830 iter += (1 << page_order(page)) - 1;
7835 * The HWPoisoned page may be not in buddy system, and
7836 * page_count() is not 0.
7838 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
7841 if (__PageMovable(page))
7847 * If there are RECLAIMABLE pages, we need to check
7848 * it. But now, memory offline itself doesn't call
7849 * shrink_node_slabs() and it still to be fixed.
7852 * If the page is not RAM, page_count()should be 0.
7853 * we don't need more check. This is an _used_ not-movable page.
7855 * The problematic thing here is PG_reserved pages. PG_reserved
7856 * is set to both of a memory hole page and a _used_ kernel
7864 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7865 if (flags & REPORT_FAILURE)
7866 dump_page(pfn_to_page(pfn+iter), "unmovable page");
7870 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7872 static unsigned long pfn_max_align_down(unsigned long pfn)
7874 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7875 pageblock_nr_pages) - 1);
7878 static unsigned long pfn_max_align_up(unsigned long pfn)
7880 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7881 pageblock_nr_pages));
7884 /* [start, end) must belong to a single zone. */
7885 static int __alloc_contig_migrate_range(struct compact_control *cc,
7886 unsigned long start, unsigned long end)
7888 /* This function is based on compact_zone() from compaction.c. */
7889 unsigned long nr_reclaimed;
7890 unsigned long pfn = start;
7891 unsigned int tries = 0;
7896 while (pfn < end || !list_empty(&cc->migratepages)) {
7897 if (fatal_signal_pending(current)) {
7902 if (list_empty(&cc->migratepages)) {
7903 cc->nr_migratepages = 0;
7904 pfn = isolate_migratepages_range(cc, pfn, end);
7910 } else if (++tries == 5) {
7911 ret = ret < 0 ? ret : -EBUSY;
7915 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7917 cc->nr_migratepages -= nr_reclaimed;
7919 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7920 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7923 putback_movable_pages(&cc->migratepages);
7930 * alloc_contig_range() -- tries to allocate given range of pages
7931 * @start: start PFN to allocate
7932 * @end: one-past-the-last PFN to allocate
7933 * @migratetype: migratetype of the underlaying pageblocks (either
7934 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7935 * in range must have the same migratetype and it must
7936 * be either of the two.
7937 * @gfp_mask: GFP mask to use during compaction
7939 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7940 * aligned. The PFN range must belong to a single zone.
7942 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7943 * pageblocks in the range. Once isolated, the pageblocks should not
7944 * be modified by others.
7946 * Returns zero on success or negative error code. On success all
7947 * pages which PFN is in [start, end) are allocated for the caller and
7948 * need to be freed with free_contig_range().
7950 int alloc_contig_range(unsigned long start, unsigned long end,
7951 unsigned migratetype, gfp_t gfp_mask)
7953 unsigned long outer_start, outer_end;
7957 struct compact_control cc = {
7958 .nr_migratepages = 0,
7960 .zone = page_zone(pfn_to_page(start)),
7961 .mode = MIGRATE_SYNC,
7962 .ignore_skip_hint = true,
7963 .no_set_skip_hint = true,
7964 .gfp_mask = current_gfp_context(gfp_mask),
7966 INIT_LIST_HEAD(&cc.migratepages);
7969 * What we do here is we mark all pageblocks in range as
7970 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7971 * have different sizes, and due to the way page allocator
7972 * work, we align the range to biggest of the two pages so
7973 * that page allocator won't try to merge buddies from
7974 * different pageblocks and change MIGRATE_ISOLATE to some
7975 * other migration type.
7977 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7978 * migrate the pages from an unaligned range (ie. pages that
7979 * we are interested in). This will put all the pages in
7980 * range back to page allocator as MIGRATE_ISOLATE.
7982 * When this is done, we take the pages in range from page
7983 * allocator removing them from the buddy system. This way
7984 * page allocator will never consider using them.
7986 * This lets us mark the pageblocks back as
7987 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7988 * aligned range but not in the unaligned, original range are
7989 * put back to page allocator so that buddy can use them.
7992 ret = start_isolate_page_range(pfn_max_align_down(start),
7993 pfn_max_align_up(end), migratetype, 0);
7998 * In case of -EBUSY, we'd like to know which page causes problem.
7999 * So, just fall through. test_pages_isolated() has a tracepoint
8000 * which will report the busy page.
8002 * It is possible that busy pages could become available before
8003 * the call to test_pages_isolated, and the range will actually be
8004 * allocated. So, if we fall through be sure to clear ret so that
8005 * -EBUSY is not accidentally used or returned to caller.
8007 ret = __alloc_contig_migrate_range(&cc, start, end);
8008 if (ret && ret != -EBUSY)
8013 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8014 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8015 * more, all pages in [start, end) are free in page allocator.
8016 * What we are going to do is to allocate all pages from
8017 * [start, end) (that is remove them from page allocator).
8019 * The only problem is that pages at the beginning and at the
8020 * end of interesting range may be not aligned with pages that
8021 * page allocator holds, ie. they can be part of higher order
8022 * pages. Because of this, we reserve the bigger range and
8023 * once this is done free the pages we are not interested in.
8025 * We don't have to hold zone->lock here because the pages are
8026 * isolated thus they won't get removed from buddy.
8029 lru_add_drain_all();
8030 drain_all_pages(cc.zone);
8033 outer_start = start;
8034 while (!PageBuddy(pfn_to_page(outer_start))) {
8035 if (++order >= MAX_ORDER) {
8036 outer_start = start;
8039 outer_start &= ~0UL << order;
8042 if (outer_start != start) {
8043 order = page_order(pfn_to_page(outer_start));
8046 * outer_start page could be small order buddy page and
8047 * it doesn't include start page. Adjust outer_start
8048 * in this case to report failed page properly
8049 * on tracepoint in test_pages_isolated()
8051 if (outer_start + (1UL << order) <= start)
8052 outer_start = start;
8055 /* Make sure the range is really isolated. */
8056 if (test_pages_isolated(outer_start, end, false)) {
8057 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8058 __func__, outer_start, end);
8063 /* Grab isolated pages from freelists. */
8064 outer_end = isolate_freepages_range(&cc, outer_start, end);
8070 /* Free head and tail (if any) */
8071 if (start != outer_start)
8072 free_contig_range(outer_start, start - outer_start);
8073 if (end != outer_end)
8074 free_contig_range(end, outer_end - end);
8077 undo_isolate_page_range(pfn_max_align_down(start),
8078 pfn_max_align_up(end), migratetype);
8082 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8084 unsigned int count = 0;
8086 for (; nr_pages--; pfn++) {
8087 struct page *page = pfn_to_page(pfn);
8089 count += page_count(page) != 1;
8092 WARN(count != 0, "%d pages are still in use!\n", count);
8096 #ifdef CONFIG_MEMORY_HOTPLUG
8098 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8099 * page high values need to be recalulated.
8101 void __meminit zone_pcp_update(struct zone *zone)
8104 mutex_lock(&pcp_batch_high_lock);
8105 for_each_possible_cpu(cpu)
8106 pageset_set_high_and_batch(zone,
8107 per_cpu_ptr(zone->pageset, cpu));
8108 mutex_unlock(&pcp_batch_high_lock);
8112 void zone_pcp_reset(struct zone *zone)
8114 unsigned long flags;
8116 struct per_cpu_pageset *pset;
8118 /* avoid races with drain_pages() */
8119 local_irq_save(flags);
8120 if (zone->pageset != &boot_pageset) {
8121 for_each_online_cpu(cpu) {
8122 pset = per_cpu_ptr(zone->pageset, cpu);
8123 drain_zonestat(zone, pset);
8125 free_percpu(zone->pageset);
8126 zone->pageset = &boot_pageset;
8128 local_irq_restore(flags);
8131 #ifdef CONFIG_MEMORY_HOTREMOVE
8133 * All pages in the range must be in a single zone and isolated
8134 * before calling this.
8137 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8141 unsigned int order, i;
8143 unsigned long flags;
8144 /* find the first valid pfn */
8145 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8150 offline_mem_sections(pfn, end_pfn);
8151 zone = page_zone(pfn_to_page(pfn));
8152 spin_lock_irqsave(&zone->lock, flags);
8154 while (pfn < end_pfn) {
8155 if (!pfn_valid(pfn)) {
8159 page = pfn_to_page(pfn);
8161 * The HWPoisoned page may be not in buddy system, and
8162 * page_count() is not 0.
8164 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8166 SetPageReserved(page);
8170 BUG_ON(page_count(page));
8171 BUG_ON(!PageBuddy(page));
8172 order = page_order(page);
8173 #ifdef CONFIG_DEBUG_VM
8174 pr_info("remove from free list %lx %d %lx\n",
8175 pfn, 1 << order, end_pfn);
8177 list_del(&page->lru);
8178 rmv_page_order(page);
8179 zone->free_area[order].nr_free--;
8180 for (i = 0; i < (1 << order); i++)
8181 SetPageReserved((page+i));
8182 pfn += (1 << order);
8184 spin_unlock_irqrestore(&zone->lock, flags);
8188 bool is_free_buddy_page(struct page *page)
8190 struct zone *zone = page_zone(page);
8191 unsigned long pfn = page_to_pfn(page);
8192 unsigned long flags;
8195 spin_lock_irqsave(&zone->lock, flags);
8196 for (order = 0; order < MAX_ORDER; order++) {
8197 struct page *page_head = page - (pfn & ((1 << order) - 1));
8199 if (PageBuddy(page_head) && page_order(page_head) >= order)
8202 spin_unlock_irqrestore(&zone->lock, flags);
8204 return order < MAX_ORDER;
8207 #ifdef CONFIG_MEMORY_FAILURE
8209 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8210 * test is performed under the zone lock to prevent a race against page
8213 bool set_hwpoison_free_buddy_page(struct page *page)
8215 struct zone *zone = page_zone(page);
8216 unsigned long pfn = page_to_pfn(page);
8217 unsigned long flags;
8219 bool hwpoisoned = false;
8221 spin_lock_irqsave(&zone->lock, flags);
8222 for (order = 0; order < MAX_ORDER; order++) {
8223 struct page *page_head = page - (pfn & ((1 << order) - 1));
8225 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8226 if (!TestSetPageHWPoison(page))
8231 spin_unlock_irqrestore(&zone->lock, flags);