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,
2379 unsigned int alloc_flags)
2381 struct free_area *area;
2383 int min_order = order;
2389 * Do not steal pages from freelists belonging to other pageblocks
2390 * i.e. orders < pageblock_order. If there are no local zones free,
2391 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2393 if (alloc_flags & ALLOC_NOFRAGMENT)
2394 min_order = pageblock_order;
2397 * Find the largest available free page in the other list. This roughly
2398 * approximates finding the pageblock with the most free pages, which
2399 * would be too costly to do exactly.
2401 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2403 area = &(zone->free_area[current_order]);
2404 fallback_mt = find_suitable_fallback(area, current_order,
2405 start_migratetype, false, &can_steal);
2406 if (fallback_mt == -1)
2410 * We cannot steal all free pages from the pageblock and the
2411 * requested migratetype is movable. In that case it's better to
2412 * steal and split the smallest available page instead of the
2413 * largest available page, because even if the next movable
2414 * allocation falls back into a different pageblock than this
2415 * one, it won't cause permanent fragmentation.
2417 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2418 && current_order > order)
2427 for (current_order = order; current_order < MAX_ORDER;
2429 area = &(zone->free_area[current_order]);
2430 fallback_mt = find_suitable_fallback(area, current_order,
2431 start_migratetype, false, &can_steal);
2432 if (fallback_mt != -1)
2437 * This should not happen - we already found a suitable fallback
2438 * when looking for the largest page.
2440 VM_BUG_ON(current_order == MAX_ORDER);
2443 page = list_first_entry(&area->free_list[fallback_mt],
2446 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2448 trace_mm_page_alloc_extfrag(page, order, current_order,
2449 start_migratetype, fallback_mt);
2456 * Do the hard work of removing an element from the buddy allocator.
2457 * Call me with the zone->lock already held.
2459 static __always_inline struct page *
2460 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2461 unsigned int alloc_flags)
2466 page = __rmqueue_smallest(zone, order, migratetype);
2467 if (unlikely(!page)) {
2468 if (migratetype == MIGRATE_MOVABLE)
2469 page = __rmqueue_cma_fallback(zone, order);
2471 if (!page && __rmqueue_fallback(zone, order, migratetype,
2476 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2481 * Obtain a specified number of elements from the buddy allocator, all under
2482 * a single hold of the lock, for efficiency. Add them to the supplied list.
2483 * Returns the number of new pages which were placed at *list.
2485 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2486 unsigned long count, struct list_head *list,
2487 int migratetype, unsigned int alloc_flags)
2491 spin_lock(&zone->lock);
2492 for (i = 0; i < count; ++i) {
2493 struct page *page = __rmqueue(zone, order, migratetype,
2495 if (unlikely(page == NULL))
2498 if (unlikely(check_pcp_refill(page)))
2502 * Split buddy pages returned by expand() are received here in
2503 * physical page order. The page is added to the tail of
2504 * caller's list. From the callers perspective, the linked list
2505 * is ordered by page number under some conditions. This is
2506 * useful for IO devices that can forward direction from the
2507 * head, thus also in the physical page order. This is useful
2508 * for IO devices that can merge IO requests if the physical
2509 * pages are ordered properly.
2511 list_add_tail(&page->lru, list);
2513 if (is_migrate_cma(get_pcppage_migratetype(page)))
2514 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2519 * i pages were removed from the buddy list even if some leak due
2520 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2521 * on i. Do not confuse with 'alloced' which is the number of
2522 * pages added to the pcp list.
2524 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2525 spin_unlock(&zone->lock);
2531 * Called from the vmstat counter updater to drain pagesets of this
2532 * currently executing processor on remote nodes after they have
2535 * Note that this function must be called with the thread pinned to
2536 * a single processor.
2538 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2540 unsigned long flags;
2541 int to_drain, batch;
2543 local_irq_save(flags);
2544 batch = READ_ONCE(pcp->batch);
2545 to_drain = min(pcp->count, batch);
2547 free_pcppages_bulk(zone, to_drain, pcp);
2548 local_irq_restore(flags);
2553 * Drain pcplists of the indicated processor and zone.
2555 * The processor must either be the current processor and the
2556 * thread pinned to the current processor or a processor that
2559 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2561 unsigned long flags;
2562 struct per_cpu_pageset *pset;
2563 struct per_cpu_pages *pcp;
2565 local_irq_save(flags);
2566 pset = per_cpu_ptr(zone->pageset, cpu);
2570 free_pcppages_bulk(zone, pcp->count, pcp);
2571 local_irq_restore(flags);
2575 * Drain pcplists of all zones on the indicated processor.
2577 * The processor must either be the current processor and the
2578 * thread pinned to the current processor or a processor that
2581 static void drain_pages(unsigned int cpu)
2585 for_each_populated_zone(zone) {
2586 drain_pages_zone(cpu, zone);
2591 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2593 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2594 * the single zone's pages.
2596 void drain_local_pages(struct zone *zone)
2598 int cpu = smp_processor_id();
2601 drain_pages_zone(cpu, zone);
2606 static void drain_local_pages_wq(struct work_struct *work)
2609 * drain_all_pages doesn't use proper cpu hotplug protection so
2610 * we can race with cpu offline when the WQ can move this from
2611 * a cpu pinned worker to an unbound one. We can operate on a different
2612 * cpu which is allright but we also have to make sure to not move to
2616 drain_local_pages(NULL);
2621 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2623 * When zone parameter is non-NULL, spill just the single zone's pages.
2625 * Note that this can be extremely slow as the draining happens in a workqueue.
2627 void drain_all_pages(struct zone *zone)
2632 * Allocate in the BSS so we wont require allocation in
2633 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2635 static cpumask_t cpus_with_pcps;
2638 * Make sure nobody triggers this path before mm_percpu_wq is fully
2641 if (WARN_ON_ONCE(!mm_percpu_wq))
2645 * Do not drain if one is already in progress unless it's specific to
2646 * a zone. Such callers are primarily CMA and memory hotplug and need
2647 * the drain to be complete when the call returns.
2649 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2652 mutex_lock(&pcpu_drain_mutex);
2656 * We don't care about racing with CPU hotplug event
2657 * as offline notification will cause the notified
2658 * cpu to drain that CPU pcps and on_each_cpu_mask
2659 * disables preemption as part of its processing
2661 for_each_online_cpu(cpu) {
2662 struct per_cpu_pageset *pcp;
2664 bool has_pcps = false;
2667 pcp = per_cpu_ptr(zone->pageset, cpu);
2671 for_each_populated_zone(z) {
2672 pcp = per_cpu_ptr(z->pageset, cpu);
2673 if (pcp->pcp.count) {
2681 cpumask_set_cpu(cpu, &cpus_with_pcps);
2683 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2686 for_each_cpu(cpu, &cpus_with_pcps) {
2687 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2688 INIT_WORK(work, drain_local_pages_wq);
2689 queue_work_on(cpu, mm_percpu_wq, work);
2691 for_each_cpu(cpu, &cpus_with_pcps)
2692 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2694 mutex_unlock(&pcpu_drain_mutex);
2697 #ifdef CONFIG_HIBERNATION
2700 * Touch the watchdog for every WD_PAGE_COUNT pages.
2702 #define WD_PAGE_COUNT (128*1024)
2704 void mark_free_pages(struct zone *zone)
2706 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2707 unsigned long flags;
2708 unsigned int order, t;
2711 if (zone_is_empty(zone))
2714 spin_lock_irqsave(&zone->lock, flags);
2716 max_zone_pfn = zone_end_pfn(zone);
2717 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2718 if (pfn_valid(pfn)) {
2719 page = pfn_to_page(pfn);
2721 if (!--page_count) {
2722 touch_nmi_watchdog();
2723 page_count = WD_PAGE_COUNT;
2726 if (page_zone(page) != zone)
2729 if (!swsusp_page_is_forbidden(page))
2730 swsusp_unset_page_free(page);
2733 for_each_migratetype_order(order, t) {
2734 list_for_each_entry(page,
2735 &zone->free_area[order].free_list[t], lru) {
2738 pfn = page_to_pfn(page);
2739 for (i = 0; i < (1UL << order); i++) {
2740 if (!--page_count) {
2741 touch_nmi_watchdog();
2742 page_count = WD_PAGE_COUNT;
2744 swsusp_set_page_free(pfn_to_page(pfn + i));
2748 spin_unlock_irqrestore(&zone->lock, flags);
2750 #endif /* CONFIG_PM */
2752 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2756 if (!free_pcp_prepare(page))
2759 migratetype = get_pfnblock_migratetype(page, pfn);
2760 set_pcppage_migratetype(page, migratetype);
2764 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2766 struct zone *zone = page_zone(page);
2767 struct per_cpu_pages *pcp;
2770 migratetype = get_pcppage_migratetype(page);
2771 __count_vm_event(PGFREE);
2774 * We only track unmovable, reclaimable and movable on pcp lists.
2775 * Free ISOLATE pages back to the allocator because they are being
2776 * offlined but treat HIGHATOMIC as movable pages so we can get those
2777 * areas back if necessary. Otherwise, we may have to free
2778 * excessively into the page allocator
2780 if (migratetype >= MIGRATE_PCPTYPES) {
2781 if (unlikely(is_migrate_isolate(migratetype))) {
2782 free_one_page(zone, page, pfn, 0, migratetype);
2785 migratetype = MIGRATE_MOVABLE;
2788 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2789 list_add(&page->lru, &pcp->lists[migratetype]);
2791 if (pcp->count >= pcp->high) {
2792 unsigned long batch = READ_ONCE(pcp->batch);
2793 free_pcppages_bulk(zone, batch, pcp);
2798 * Free a 0-order page
2800 void free_unref_page(struct page *page)
2802 unsigned long flags;
2803 unsigned long pfn = page_to_pfn(page);
2805 if (!free_unref_page_prepare(page, pfn))
2808 local_irq_save(flags);
2809 free_unref_page_commit(page, pfn);
2810 local_irq_restore(flags);
2814 * Free a list of 0-order pages
2816 void free_unref_page_list(struct list_head *list)
2818 struct page *page, *next;
2819 unsigned long flags, pfn;
2820 int batch_count = 0;
2822 /* Prepare pages for freeing */
2823 list_for_each_entry_safe(page, next, list, lru) {
2824 pfn = page_to_pfn(page);
2825 if (!free_unref_page_prepare(page, pfn))
2826 list_del(&page->lru);
2827 set_page_private(page, pfn);
2830 local_irq_save(flags);
2831 list_for_each_entry_safe(page, next, list, lru) {
2832 unsigned long pfn = page_private(page);
2834 set_page_private(page, 0);
2835 trace_mm_page_free_batched(page);
2836 free_unref_page_commit(page, pfn);
2839 * Guard against excessive IRQ disabled times when we get
2840 * a large list of pages to free.
2842 if (++batch_count == SWAP_CLUSTER_MAX) {
2843 local_irq_restore(flags);
2845 local_irq_save(flags);
2848 local_irq_restore(flags);
2852 * split_page takes a non-compound higher-order page, and splits it into
2853 * n (1<<order) sub-pages: page[0..n]
2854 * Each sub-page must be freed individually.
2856 * Note: this is probably too low level an operation for use in drivers.
2857 * Please consult with lkml before using this in your driver.
2859 void split_page(struct page *page, unsigned int order)
2863 VM_BUG_ON_PAGE(PageCompound(page), page);
2864 VM_BUG_ON_PAGE(!page_count(page), page);
2866 for (i = 1; i < (1 << order); i++)
2867 set_page_refcounted(page + i);
2868 split_page_owner(page, order);
2870 EXPORT_SYMBOL_GPL(split_page);
2872 int __isolate_free_page(struct page *page, unsigned int order)
2874 unsigned long watermark;
2878 BUG_ON(!PageBuddy(page));
2880 zone = page_zone(page);
2881 mt = get_pageblock_migratetype(page);
2883 if (!is_migrate_isolate(mt)) {
2885 * Obey watermarks as if the page was being allocated. We can
2886 * emulate a high-order watermark check with a raised order-0
2887 * watermark, because we already know our high-order page
2890 watermark = min_wmark_pages(zone) + (1UL << order);
2891 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2894 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2897 /* Remove page from free list */
2898 list_del(&page->lru);
2899 zone->free_area[order].nr_free--;
2900 rmv_page_order(page);
2903 * Set the pageblock if the isolated page is at least half of a
2906 if (order >= pageblock_order - 1) {
2907 struct page *endpage = page + (1 << order) - 1;
2908 for (; page < endpage; page += pageblock_nr_pages) {
2909 int mt = get_pageblock_migratetype(page);
2910 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2911 && !is_migrate_highatomic(mt))
2912 set_pageblock_migratetype(page,
2918 return 1UL << order;
2922 * Update NUMA hit/miss statistics
2924 * Must be called with interrupts disabled.
2926 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2929 enum numa_stat_item local_stat = NUMA_LOCAL;
2931 /* skip numa counters update if numa stats is disabled */
2932 if (!static_branch_likely(&vm_numa_stat_key))
2935 if (zone_to_nid(z) != numa_node_id())
2936 local_stat = NUMA_OTHER;
2938 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2939 __inc_numa_state(z, NUMA_HIT);
2941 __inc_numa_state(z, NUMA_MISS);
2942 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2944 __inc_numa_state(z, local_stat);
2948 /* Remove page from the per-cpu list, caller must protect the list */
2949 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2950 unsigned int alloc_flags,
2951 struct per_cpu_pages *pcp,
2952 struct list_head *list)
2957 if (list_empty(list)) {
2958 pcp->count += rmqueue_bulk(zone, 0,
2960 migratetype, alloc_flags);
2961 if (unlikely(list_empty(list)))
2965 page = list_first_entry(list, struct page, lru);
2966 list_del(&page->lru);
2968 } while (check_new_pcp(page));
2973 /* Lock and remove page from the per-cpu list */
2974 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2975 struct zone *zone, unsigned int order,
2976 gfp_t gfp_flags, int migratetype,
2977 unsigned int alloc_flags)
2979 struct per_cpu_pages *pcp;
2980 struct list_head *list;
2982 unsigned long flags;
2984 local_irq_save(flags);
2985 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2986 list = &pcp->lists[migratetype];
2987 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
2989 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2990 zone_statistics(preferred_zone, zone);
2992 local_irq_restore(flags);
2997 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3000 struct page *rmqueue(struct zone *preferred_zone,
3001 struct zone *zone, unsigned int order,
3002 gfp_t gfp_flags, unsigned int alloc_flags,
3005 unsigned long flags;
3008 if (likely(order == 0)) {
3009 page = rmqueue_pcplist(preferred_zone, zone, order,
3010 gfp_flags, migratetype, alloc_flags);
3015 * We most definitely don't want callers attempting to
3016 * allocate greater than order-1 page units with __GFP_NOFAIL.
3018 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3019 spin_lock_irqsave(&zone->lock, flags);
3023 if (alloc_flags & ALLOC_HARDER) {
3024 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3026 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3029 page = __rmqueue(zone, order, migratetype, alloc_flags);
3030 } while (page && check_new_pages(page, order));
3031 spin_unlock(&zone->lock);
3034 __mod_zone_freepage_state(zone, -(1 << order),
3035 get_pcppage_migratetype(page));
3037 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3038 zone_statistics(preferred_zone, zone);
3039 local_irq_restore(flags);
3042 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3046 local_irq_restore(flags);
3050 #ifdef CONFIG_FAIL_PAGE_ALLOC
3053 struct fault_attr attr;
3055 bool ignore_gfp_highmem;
3056 bool ignore_gfp_reclaim;
3058 } fail_page_alloc = {
3059 .attr = FAULT_ATTR_INITIALIZER,
3060 .ignore_gfp_reclaim = true,
3061 .ignore_gfp_highmem = true,
3065 static int __init setup_fail_page_alloc(char *str)
3067 return setup_fault_attr(&fail_page_alloc.attr, str);
3069 __setup("fail_page_alloc=", setup_fail_page_alloc);
3071 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3073 if (order < fail_page_alloc.min_order)
3075 if (gfp_mask & __GFP_NOFAIL)
3077 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3079 if (fail_page_alloc.ignore_gfp_reclaim &&
3080 (gfp_mask & __GFP_DIRECT_RECLAIM))
3083 return should_fail(&fail_page_alloc.attr, 1 << order);
3086 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3088 static int __init fail_page_alloc_debugfs(void)
3090 umode_t mode = S_IFREG | 0600;
3093 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3094 &fail_page_alloc.attr);
3096 return PTR_ERR(dir);
3098 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3099 &fail_page_alloc.ignore_gfp_reclaim))
3101 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3102 &fail_page_alloc.ignore_gfp_highmem))
3104 if (!debugfs_create_u32("min-order", mode, dir,
3105 &fail_page_alloc.min_order))
3110 debugfs_remove_recursive(dir);
3115 late_initcall(fail_page_alloc_debugfs);
3117 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3119 #else /* CONFIG_FAIL_PAGE_ALLOC */
3121 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3126 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3129 * Return true if free base pages are above 'mark'. For high-order checks it
3130 * will return true of the order-0 watermark is reached and there is at least
3131 * one free page of a suitable size. Checking now avoids taking the zone lock
3132 * to check in the allocation paths if no pages are free.
3134 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3135 int classzone_idx, unsigned int alloc_flags,
3140 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3142 /* free_pages may go negative - that's OK */
3143 free_pages -= (1 << order) - 1;
3145 if (alloc_flags & ALLOC_HIGH)
3149 * If the caller does not have rights to ALLOC_HARDER then subtract
3150 * the high-atomic reserves. This will over-estimate the size of the
3151 * atomic reserve but it avoids a search.
3153 if (likely(!alloc_harder)) {
3154 free_pages -= z->nr_reserved_highatomic;
3157 * OOM victims can try even harder than normal ALLOC_HARDER
3158 * users on the grounds that it's definitely going to be in
3159 * the exit path shortly and free memory. Any allocation it
3160 * makes during the free path will be small and short-lived.
3162 if (alloc_flags & ALLOC_OOM)
3170 /* If allocation can't use CMA areas don't use free CMA pages */
3171 if (!(alloc_flags & ALLOC_CMA))
3172 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3176 * Check watermarks for an order-0 allocation request. If these
3177 * are not met, then a high-order request also cannot go ahead
3178 * even if a suitable page happened to be free.
3180 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3183 /* If this is an order-0 request then the watermark is fine */
3187 /* For a high-order request, check at least one suitable page is free */
3188 for (o = order; o < MAX_ORDER; o++) {
3189 struct free_area *area = &z->free_area[o];
3195 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3196 if (!list_empty(&area->free_list[mt]))
3201 if ((alloc_flags & ALLOC_CMA) &&
3202 !list_empty(&area->free_list[MIGRATE_CMA])) {
3207 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3213 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3214 int classzone_idx, unsigned int alloc_flags)
3216 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3217 zone_page_state(z, NR_FREE_PAGES));
3220 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3221 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3223 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3227 /* If allocation can't use CMA areas don't use free CMA pages */
3228 if (!(alloc_flags & ALLOC_CMA))
3229 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3233 * Fast check for order-0 only. If this fails then the reserves
3234 * need to be calculated. There is a corner case where the check
3235 * passes but only the high-order atomic reserve are free. If
3236 * the caller is !atomic then it'll uselessly search the free
3237 * list. That corner case is then slower but it is harmless.
3239 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3242 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3246 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3247 unsigned long mark, int classzone_idx)
3249 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3251 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3252 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3254 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3259 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3261 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3264 #else /* CONFIG_NUMA */
3265 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3269 #endif /* CONFIG_NUMA */
3271 #ifdef CONFIG_ZONE_DMA32
3273 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3274 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3275 * premature use of a lower zone may cause lowmem pressure problems that
3276 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3277 * probably too small. It only makes sense to spread allocations to avoid
3278 * fragmentation between the Normal and DMA32 zones.
3280 static inline unsigned int
3281 alloc_flags_nofragment(struct zone *zone)
3283 if (zone_idx(zone) != ZONE_NORMAL)
3287 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3288 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3289 * on UMA that if Normal is populated then so is DMA32.
3291 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3292 if (nr_online_nodes > 1 && !populated_zone(--zone))
3295 return ALLOC_NOFRAGMENT;
3298 static inline unsigned int
3299 alloc_flags_nofragment(struct zone *zone)
3306 * get_page_from_freelist goes through the zonelist trying to allocate
3309 static struct page *
3310 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3311 const struct alloc_context *ac)
3315 struct pglist_data *last_pgdat_dirty_limit = NULL;
3320 * Scan zonelist, looking for a zone with enough free.
3321 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3323 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3324 z = ac->preferred_zoneref;
3325 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3330 if (cpusets_enabled() &&
3331 (alloc_flags & ALLOC_CPUSET) &&
3332 !__cpuset_zone_allowed(zone, gfp_mask))
3335 * When allocating a page cache page for writing, we
3336 * want to get it from a node that is within its dirty
3337 * limit, such that no single node holds more than its
3338 * proportional share of globally allowed dirty pages.
3339 * The dirty limits take into account the node's
3340 * lowmem reserves and high watermark so that kswapd
3341 * should be able to balance it without having to
3342 * write pages from its LRU list.
3344 * XXX: For now, allow allocations to potentially
3345 * exceed the per-node dirty limit in the slowpath
3346 * (spread_dirty_pages unset) before going into reclaim,
3347 * which is important when on a NUMA setup the allowed
3348 * nodes are together not big enough to reach the
3349 * global limit. The proper fix for these situations
3350 * will require awareness of nodes in the
3351 * dirty-throttling and the flusher threads.
3353 if (ac->spread_dirty_pages) {
3354 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3357 if (!node_dirty_ok(zone->zone_pgdat)) {
3358 last_pgdat_dirty_limit = zone->zone_pgdat;
3363 if (no_fallback && nr_online_nodes > 1 &&
3364 zone != ac->preferred_zoneref->zone) {
3368 * If moving to a remote node, retry but allow
3369 * fragmenting fallbacks. Locality is more important
3370 * than fragmentation avoidance.
3372 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3373 if (zone_to_nid(zone) != local_nid) {
3374 alloc_flags &= ~ALLOC_NOFRAGMENT;
3379 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3380 if (!zone_watermark_fast(zone, order, mark,
3381 ac_classzone_idx(ac), alloc_flags)) {
3384 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3386 * Watermark failed for this zone, but see if we can
3387 * grow this zone if it contains deferred pages.
3389 if (static_branch_unlikely(&deferred_pages)) {
3390 if (_deferred_grow_zone(zone, order))
3394 /* Checked here to keep the fast path fast */
3395 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3396 if (alloc_flags & ALLOC_NO_WATERMARKS)
3399 if (node_reclaim_mode == 0 ||
3400 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3403 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3405 case NODE_RECLAIM_NOSCAN:
3408 case NODE_RECLAIM_FULL:
3409 /* scanned but unreclaimable */
3412 /* did we reclaim enough */
3413 if (zone_watermark_ok(zone, order, mark,
3414 ac_classzone_idx(ac), alloc_flags))
3422 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3423 gfp_mask, alloc_flags, ac->migratetype);
3425 prep_new_page(page, order, gfp_mask, alloc_flags);
3428 * If this is a high-order atomic allocation then check
3429 * if the pageblock should be reserved for the future
3431 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3432 reserve_highatomic_pageblock(page, zone, order);
3436 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3437 /* Try again if zone has deferred pages */
3438 if (static_branch_unlikely(&deferred_pages)) {
3439 if (_deferred_grow_zone(zone, order))
3447 * It's possible on a UMA machine to get through all zones that are
3448 * fragmented. If avoiding fragmentation, reset and try again.
3451 alloc_flags &= ~ALLOC_NOFRAGMENT;
3458 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3460 unsigned int filter = SHOW_MEM_FILTER_NODES;
3461 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3463 if (!__ratelimit(&show_mem_rs))
3467 * This documents exceptions given to allocations in certain
3468 * contexts that are allowed to allocate outside current's set
3471 if (!(gfp_mask & __GFP_NOMEMALLOC))
3472 if (tsk_is_oom_victim(current) ||
3473 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3474 filter &= ~SHOW_MEM_FILTER_NODES;
3475 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3476 filter &= ~SHOW_MEM_FILTER_NODES;
3478 show_mem(filter, nodemask);
3481 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3483 struct va_format vaf;
3485 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3486 DEFAULT_RATELIMIT_BURST);
3488 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3491 va_start(args, fmt);
3494 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3495 current->comm, &vaf, gfp_mask, &gfp_mask,
3496 nodemask_pr_args(nodemask));
3499 cpuset_print_current_mems_allowed();
3502 warn_alloc_show_mem(gfp_mask, nodemask);
3505 static inline struct page *
3506 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3507 unsigned int alloc_flags,
3508 const struct alloc_context *ac)
3512 page = get_page_from_freelist(gfp_mask, order,
3513 alloc_flags|ALLOC_CPUSET, ac);
3515 * fallback to ignore cpuset restriction if our nodes
3519 page = get_page_from_freelist(gfp_mask, order,
3525 static inline struct page *
3526 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3527 const struct alloc_context *ac, unsigned long *did_some_progress)
3529 struct oom_control oc = {
3530 .zonelist = ac->zonelist,
3531 .nodemask = ac->nodemask,
3533 .gfp_mask = gfp_mask,
3538 *did_some_progress = 0;
3541 * Acquire the oom lock. If that fails, somebody else is
3542 * making progress for us.
3544 if (!mutex_trylock(&oom_lock)) {
3545 *did_some_progress = 1;
3546 schedule_timeout_uninterruptible(1);
3551 * Go through the zonelist yet one more time, keep very high watermark
3552 * here, this is only to catch a parallel oom killing, we must fail if
3553 * we're still under heavy pressure. But make sure that this reclaim
3554 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3555 * allocation which will never fail due to oom_lock already held.
3557 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3558 ~__GFP_DIRECT_RECLAIM, order,
3559 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3563 /* Coredumps can quickly deplete all memory reserves */
3564 if (current->flags & PF_DUMPCORE)
3566 /* The OOM killer will not help higher order allocs */
3567 if (order > PAGE_ALLOC_COSTLY_ORDER)
3570 * We have already exhausted all our reclaim opportunities without any
3571 * success so it is time to admit defeat. We will skip the OOM killer
3572 * because it is very likely that the caller has a more reasonable
3573 * fallback than shooting a random task.
3575 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3577 /* The OOM killer does not needlessly kill tasks for lowmem */
3578 if (ac->high_zoneidx < ZONE_NORMAL)
3580 if (pm_suspended_storage())
3583 * XXX: GFP_NOFS allocations should rather fail than rely on
3584 * other request to make a forward progress.
3585 * We are in an unfortunate situation where out_of_memory cannot
3586 * do much for this context but let's try it to at least get
3587 * access to memory reserved if the current task is killed (see
3588 * out_of_memory). Once filesystems are ready to handle allocation
3589 * failures more gracefully we should just bail out here.
3592 /* The OOM killer may not free memory on a specific node */
3593 if (gfp_mask & __GFP_THISNODE)
3596 /* Exhausted what can be done so it's blame time */
3597 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3598 *did_some_progress = 1;
3601 * Help non-failing allocations by giving them access to memory
3604 if (gfp_mask & __GFP_NOFAIL)
3605 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3606 ALLOC_NO_WATERMARKS, ac);
3609 mutex_unlock(&oom_lock);
3614 * Maximum number of compaction retries wit a progress before OOM
3615 * killer is consider as the only way to move forward.
3617 #define MAX_COMPACT_RETRIES 16
3619 #ifdef CONFIG_COMPACTION
3620 /* Try memory compaction for high-order allocations before reclaim */
3621 static struct page *
3622 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3623 unsigned int alloc_flags, const struct alloc_context *ac,
3624 enum compact_priority prio, enum compact_result *compact_result)
3627 unsigned long pflags;
3628 unsigned int noreclaim_flag;
3633 psi_memstall_enter(&pflags);
3634 noreclaim_flag = memalloc_noreclaim_save();
3636 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3639 memalloc_noreclaim_restore(noreclaim_flag);
3640 psi_memstall_leave(&pflags);
3642 if (*compact_result <= COMPACT_INACTIVE)
3646 * At least in one zone compaction wasn't deferred or skipped, so let's
3647 * count a compaction stall
3649 count_vm_event(COMPACTSTALL);
3651 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3654 struct zone *zone = page_zone(page);
3656 zone->compact_blockskip_flush = false;
3657 compaction_defer_reset(zone, order, true);
3658 count_vm_event(COMPACTSUCCESS);
3663 * It's bad if compaction run occurs and fails. The most likely reason
3664 * is that pages exist, but not enough to satisfy watermarks.
3666 count_vm_event(COMPACTFAIL);
3674 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3675 enum compact_result compact_result,
3676 enum compact_priority *compact_priority,
3677 int *compaction_retries)
3679 int max_retries = MAX_COMPACT_RETRIES;
3682 int retries = *compaction_retries;
3683 enum compact_priority priority = *compact_priority;
3688 if (compaction_made_progress(compact_result))
3689 (*compaction_retries)++;
3692 * compaction considers all the zone as desperately out of memory
3693 * so it doesn't really make much sense to retry except when the
3694 * failure could be caused by insufficient priority
3696 if (compaction_failed(compact_result))
3697 goto check_priority;
3700 * make sure the compaction wasn't deferred or didn't bail out early
3701 * due to locks contention before we declare that we should give up.
3702 * But do not retry if the given zonelist is not suitable for
3705 if (compaction_withdrawn(compact_result)) {
3706 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3711 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3712 * costly ones because they are de facto nofail and invoke OOM
3713 * killer to move on while costly can fail and users are ready
3714 * to cope with that. 1/4 retries is rather arbitrary but we
3715 * would need much more detailed feedback from compaction to
3716 * make a better decision.
3718 if (order > PAGE_ALLOC_COSTLY_ORDER)
3720 if (*compaction_retries <= max_retries) {
3726 * Make sure there are attempts at the highest priority if we exhausted
3727 * all retries or failed at the lower priorities.
3730 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3731 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3733 if (*compact_priority > min_priority) {
3734 (*compact_priority)--;
3735 *compaction_retries = 0;
3739 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3743 static inline struct page *
3744 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3745 unsigned int alloc_flags, const struct alloc_context *ac,
3746 enum compact_priority prio, enum compact_result *compact_result)
3748 *compact_result = COMPACT_SKIPPED;
3753 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3754 enum compact_result compact_result,
3755 enum compact_priority *compact_priority,
3756 int *compaction_retries)
3761 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3765 * There are setups with compaction disabled which would prefer to loop
3766 * inside the allocator rather than hit the oom killer prematurely.
3767 * Let's give them a good hope and keep retrying while the order-0
3768 * watermarks are OK.
3770 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3772 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3773 ac_classzone_idx(ac), alloc_flags))
3778 #endif /* CONFIG_COMPACTION */
3780 #ifdef CONFIG_LOCKDEP
3781 static struct lockdep_map __fs_reclaim_map =
3782 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3784 static bool __need_fs_reclaim(gfp_t gfp_mask)
3786 gfp_mask = current_gfp_context(gfp_mask);
3788 /* no reclaim without waiting on it */
3789 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3792 /* this guy won't enter reclaim */
3793 if (current->flags & PF_MEMALLOC)
3796 /* We're only interested __GFP_FS allocations for now */
3797 if (!(gfp_mask & __GFP_FS))
3800 if (gfp_mask & __GFP_NOLOCKDEP)
3806 void __fs_reclaim_acquire(void)
3808 lock_map_acquire(&__fs_reclaim_map);
3811 void __fs_reclaim_release(void)
3813 lock_map_release(&__fs_reclaim_map);
3816 void fs_reclaim_acquire(gfp_t gfp_mask)
3818 if (__need_fs_reclaim(gfp_mask))
3819 __fs_reclaim_acquire();
3821 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3823 void fs_reclaim_release(gfp_t gfp_mask)
3825 if (__need_fs_reclaim(gfp_mask))
3826 __fs_reclaim_release();
3828 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3831 /* Perform direct synchronous page reclaim */
3833 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3834 const struct alloc_context *ac)
3836 struct reclaim_state reclaim_state;
3838 unsigned int noreclaim_flag;
3839 unsigned long pflags;
3843 /* We now go into synchronous reclaim */
3844 cpuset_memory_pressure_bump();
3845 psi_memstall_enter(&pflags);
3846 fs_reclaim_acquire(gfp_mask);
3847 noreclaim_flag = memalloc_noreclaim_save();
3848 reclaim_state.reclaimed_slab = 0;
3849 current->reclaim_state = &reclaim_state;
3851 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3854 current->reclaim_state = NULL;
3855 memalloc_noreclaim_restore(noreclaim_flag);
3856 fs_reclaim_release(gfp_mask);
3857 psi_memstall_leave(&pflags);
3864 /* The really slow allocator path where we enter direct reclaim */
3865 static inline struct page *
3866 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3867 unsigned int alloc_flags, const struct alloc_context *ac,
3868 unsigned long *did_some_progress)
3870 struct page *page = NULL;
3871 bool drained = false;
3873 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3874 if (unlikely(!(*did_some_progress)))
3878 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3881 * If an allocation failed after direct reclaim, it could be because
3882 * pages are pinned on the per-cpu lists or in high alloc reserves.
3883 * Shrink them them and try again
3885 if (!page && !drained) {
3886 unreserve_highatomic_pageblock(ac, false);
3887 drain_all_pages(NULL);
3895 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3896 const struct alloc_context *ac)
3900 pg_data_t *last_pgdat = NULL;
3901 enum zone_type high_zoneidx = ac->high_zoneidx;
3903 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3905 if (last_pgdat != zone->zone_pgdat)
3906 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3907 last_pgdat = zone->zone_pgdat;
3911 static inline unsigned int
3912 gfp_to_alloc_flags(gfp_t gfp_mask)
3914 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3916 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3917 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3920 * The caller may dip into page reserves a bit more if the caller
3921 * cannot run direct reclaim, or if the caller has realtime scheduling
3922 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3923 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3925 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3927 if (gfp_mask & __GFP_ATOMIC) {
3929 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3930 * if it can't schedule.
3932 if (!(gfp_mask & __GFP_NOMEMALLOC))
3933 alloc_flags |= ALLOC_HARDER;
3935 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3936 * comment for __cpuset_node_allowed().
3938 alloc_flags &= ~ALLOC_CPUSET;
3939 } else if (unlikely(rt_task(current)) && !in_interrupt())
3940 alloc_flags |= ALLOC_HARDER;
3943 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3944 alloc_flags |= ALLOC_CMA;
3949 static bool oom_reserves_allowed(struct task_struct *tsk)
3951 if (!tsk_is_oom_victim(tsk))
3955 * !MMU doesn't have oom reaper so give access to memory reserves
3956 * only to the thread with TIF_MEMDIE set
3958 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3965 * Distinguish requests which really need access to full memory
3966 * reserves from oom victims which can live with a portion of it
3968 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3970 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3972 if (gfp_mask & __GFP_MEMALLOC)
3973 return ALLOC_NO_WATERMARKS;
3974 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3975 return ALLOC_NO_WATERMARKS;
3976 if (!in_interrupt()) {
3977 if (current->flags & PF_MEMALLOC)
3978 return ALLOC_NO_WATERMARKS;
3979 else if (oom_reserves_allowed(current))
3986 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3988 return !!__gfp_pfmemalloc_flags(gfp_mask);
3992 * Checks whether it makes sense to retry the reclaim to make a forward progress
3993 * for the given allocation request.
3995 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3996 * without success, or when we couldn't even meet the watermark if we
3997 * reclaimed all remaining pages on the LRU lists.
3999 * Returns true if a retry is viable or false to enter the oom path.
4002 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4003 struct alloc_context *ac, int alloc_flags,
4004 bool did_some_progress, int *no_progress_loops)
4011 * Costly allocations might have made a progress but this doesn't mean
4012 * their order will become available due to high fragmentation so
4013 * always increment the no progress counter for them
4015 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4016 *no_progress_loops = 0;
4018 (*no_progress_loops)++;
4021 * Make sure we converge to OOM if we cannot make any progress
4022 * several times in the row.
4024 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4025 /* Before OOM, exhaust highatomic_reserve */
4026 return unreserve_highatomic_pageblock(ac, true);
4030 * Keep reclaiming pages while there is a chance this will lead
4031 * somewhere. If none of the target zones can satisfy our allocation
4032 * request even if all reclaimable pages are considered then we are
4033 * screwed and have to go OOM.
4035 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4037 unsigned long available;
4038 unsigned long reclaimable;
4039 unsigned long min_wmark = min_wmark_pages(zone);
4042 available = reclaimable = zone_reclaimable_pages(zone);
4043 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4046 * Would the allocation succeed if we reclaimed all
4047 * reclaimable pages?
4049 wmark = __zone_watermark_ok(zone, order, min_wmark,
4050 ac_classzone_idx(ac), alloc_flags, available);
4051 trace_reclaim_retry_zone(z, order, reclaimable,
4052 available, min_wmark, *no_progress_loops, wmark);
4055 * If we didn't make any progress and have a lot of
4056 * dirty + writeback pages then we should wait for
4057 * an IO to complete to slow down the reclaim and
4058 * prevent from pre mature OOM
4060 if (!did_some_progress) {
4061 unsigned long write_pending;
4063 write_pending = zone_page_state_snapshot(zone,
4064 NR_ZONE_WRITE_PENDING);
4066 if (2 * write_pending > reclaimable) {
4067 congestion_wait(BLK_RW_ASYNC, HZ/10);
4079 * Memory allocation/reclaim might be called from a WQ context and the
4080 * current implementation of the WQ concurrency control doesn't
4081 * recognize that a particular WQ is congested if the worker thread is
4082 * looping without ever sleeping. Therefore we have to do a short sleep
4083 * here rather than calling cond_resched().
4085 if (current->flags & PF_WQ_WORKER)
4086 schedule_timeout_uninterruptible(1);
4093 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4096 * It's possible that cpuset's mems_allowed and the nodemask from
4097 * mempolicy don't intersect. This should be normally dealt with by
4098 * policy_nodemask(), but it's possible to race with cpuset update in
4099 * such a way the check therein was true, and then it became false
4100 * before we got our cpuset_mems_cookie here.
4101 * This assumes that for all allocations, ac->nodemask can come only
4102 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4103 * when it does not intersect with the cpuset restrictions) or the
4104 * caller can deal with a violated nodemask.
4106 if (cpusets_enabled() && ac->nodemask &&
4107 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4108 ac->nodemask = NULL;
4113 * When updating a task's mems_allowed or mempolicy nodemask, it is
4114 * possible to race with parallel threads in such a way that our
4115 * allocation can fail while the mask is being updated. If we are about
4116 * to fail, check if the cpuset changed during allocation and if so,
4119 if (read_mems_allowed_retry(cpuset_mems_cookie))
4125 static inline struct page *
4126 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4127 struct alloc_context *ac)
4129 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4130 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4131 struct page *page = NULL;
4132 unsigned int alloc_flags;
4133 unsigned long did_some_progress;
4134 enum compact_priority compact_priority;
4135 enum compact_result compact_result;
4136 int compaction_retries;
4137 int no_progress_loops;
4138 unsigned int cpuset_mems_cookie;
4142 * We also sanity check to catch abuse of atomic reserves being used by
4143 * callers that are not in atomic context.
4145 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4146 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4147 gfp_mask &= ~__GFP_ATOMIC;
4150 compaction_retries = 0;
4151 no_progress_loops = 0;
4152 compact_priority = DEF_COMPACT_PRIORITY;
4153 cpuset_mems_cookie = read_mems_allowed_begin();
4156 * The fast path uses conservative alloc_flags to succeed only until
4157 * kswapd needs to be woken up, and to avoid the cost of setting up
4158 * alloc_flags precisely. So we do that now.
4160 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4163 * We need to recalculate the starting point for the zonelist iterator
4164 * because we might have used different nodemask in the fast path, or
4165 * there was a cpuset modification and we are retrying - otherwise we
4166 * could end up iterating over non-eligible zones endlessly.
4168 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4169 ac->high_zoneidx, ac->nodemask);
4170 if (!ac->preferred_zoneref->zone)
4173 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4174 wake_all_kswapds(order, gfp_mask, ac);
4177 * The adjusted alloc_flags might result in immediate success, so try
4180 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4185 * For costly allocations, try direct compaction first, as it's likely
4186 * that we have enough base pages and don't need to reclaim. For non-
4187 * movable high-order allocations, do that as well, as compaction will
4188 * try prevent permanent fragmentation by migrating from blocks of the
4190 * Don't try this for allocations that are allowed to ignore
4191 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4193 if (can_direct_reclaim &&
4195 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4196 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4197 page = __alloc_pages_direct_compact(gfp_mask, order,
4199 INIT_COMPACT_PRIORITY,
4205 * Checks for costly allocations with __GFP_NORETRY, which
4206 * includes THP page fault allocations
4208 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4210 * If compaction is deferred for high-order allocations,
4211 * it is because sync compaction recently failed. If
4212 * this is the case and the caller requested a THP
4213 * allocation, we do not want to heavily disrupt the
4214 * system, so we fail the allocation instead of entering
4217 if (compact_result == COMPACT_DEFERRED)
4221 * Looks like reclaim/compaction is worth trying, but
4222 * sync compaction could be very expensive, so keep
4223 * using async compaction.
4225 compact_priority = INIT_COMPACT_PRIORITY;
4230 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4231 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4232 wake_all_kswapds(order, gfp_mask, ac);
4234 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4236 alloc_flags = reserve_flags;
4239 * Reset the nodemask and zonelist iterators if memory policies can be
4240 * ignored. These allocations are high priority and system rather than
4243 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4244 ac->nodemask = NULL;
4245 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4246 ac->high_zoneidx, ac->nodemask);
4249 /* Attempt with potentially adjusted zonelist and alloc_flags */
4250 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4254 /* Caller is not willing to reclaim, we can't balance anything */
4255 if (!can_direct_reclaim)
4258 /* Avoid recursion of direct reclaim */
4259 if (current->flags & PF_MEMALLOC)
4262 /* Try direct reclaim and then allocating */
4263 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4264 &did_some_progress);
4268 /* Try direct compaction and then allocating */
4269 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4270 compact_priority, &compact_result);
4274 /* Do not loop if specifically requested */
4275 if (gfp_mask & __GFP_NORETRY)
4279 * Do not retry costly high order allocations unless they are
4280 * __GFP_RETRY_MAYFAIL
4282 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4285 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4286 did_some_progress > 0, &no_progress_loops))
4290 * It doesn't make any sense to retry for the compaction if the order-0
4291 * reclaim is not able to make any progress because the current
4292 * implementation of the compaction depends on the sufficient amount
4293 * of free memory (see __compaction_suitable)
4295 if (did_some_progress > 0 &&
4296 should_compact_retry(ac, order, alloc_flags,
4297 compact_result, &compact_priority,
4298 &compaction_retries))
4302 /* Deal with possible cpuset update races before we start OOM killing */
4303 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4306 /* Reclaim has failed us, start killing things */
4307 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4311 /* Avoid allocations with no watermarks from looping endlessly */
4312 if (tsk_is_oom_victim(current) &&
4313 (alloc_flags == ALLOC_OOM ||
4314 (gfp_mask & __GFP_NOMEMALLOC)))
4317 /* Retry as long as the OOM killer is making progress */
4318 if (did_some_progress) {
4319 no_progress_loops = 0;
4324 /* Deal with possible cpuset update races before we fail */
4325 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4329 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4332 if (gfp_mask & __GFP_NOFAIL) {
4334 * All existing users of the __GFP_NOFAIL are blockable, so warn
4335 * of any new users that actually require GFP_NOWAIT
4337 if (WARN_ON_ONCE(!can_direct_reclaim))
4341 * PF_MEMALLOC request from this context is rather bizarre
4342 * because we cannot reclaim anything and only can loop waiting
4343 * for somebody to do a work for us
4345 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4348 * non failing costly orders are a hard requirement which we
4349 * are not prepared for much so let's warn about these users
4350 * so that we can identify them and convert them to something
4353 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4356 * Help non-failing allocations by giving them access to memory
4357 * reserves but do not use ALLOC_NO_WATERMARKS because this
4358 * could deplete whole memory reserves which would just make
4359 * the situation worse
4361 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4369 warn_alloc(gfp_mask, ac->nodemask,
4370 "page allocation failure: order:%u", order);
4375 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4376 int preferred_nid, nodemask_t *nodemask,
4377 struct alloc_context *ac, gfp_t *alloc_mask,
4378 unsigned int *alloc_flags)
4380 ac->high_zoneidx = gfp_zone(gfp_mask);
4381 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4382 ac->nodemask = nodemask;
4383 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4385 if (cpusets_enabled()) {
4386 *alloc_mask |= __GFP_HARDWALL;
4388 ac->nodemask = &cpuset_current_mems_allowed;
4390 *alloc_flags |= ALLOC_CPUSET;
4393 fs_reclaim_acquire(gfp_mask);
4394 fs_reclaim_release(gfp_mask);
4396 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4398 if (should_fail_alloc_page(gfp_mask, order))
4401 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4402 *alloc_flags |= ALLOC_CMA;
4407 /* Determine whether to spread dirty pages and what the first usable zone */
4408 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4410 /* Dirty zone balancing only done in the fast path */
4411 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4414 * The preferred zone is used for statistics but crucially it is
4415 * also used as the starting point for the zonelist iterator. It
4416 * may get reset for allocations that ignore memory policies.
4418 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4419 ac->high_zoneidx, ac->nodemask);
4423 * This is the 'heart' of the zoned buddy allocator.
4426 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4427 nodemask_t *nodemask)
4430 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4431 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4432 struct alloc_context ac = { };
4435 * There are several places where we assume that the order value is sane
4436 * so bail out early if the request is out of bound.
4438 if (unlikely(order >= MAX_ORDER)) {
4439 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4443 gfp_mask &= gfp_allowed_mask;
4444 alloc_mask = gfp_mask;
4445 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4448 finalise_ac(gfp_mask, &ac);
4451 * Forbid the first pass from falling back to types that fragment
4452 * memory until all local zones are considered.
4454 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone);
4456 /* First allocation attempt */
4457 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4462 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4463 * resp. GFP_NOIO which has to be inherited for all allocation requests
4464 * from a particular context which has been marked by
4465 * memalloc_no{fs,io}_{save,restore}.
4467 alloc_mask = current_gfp_context(gfp_mask);
4468 ac.spread_dirty_pages = false;
4471 * Restore the original nodemask if it was potentially replaced with
4472 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4474 if (unlikely(ac.nodemask != nodemask))
4475 ac.nodemask = nodemask;
4477 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4480 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4481 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4482 __free_pages(page, order);
4486 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4490 EXPORT_SYMBOL(__alloc_pages_nodemask);
4493 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4494 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4495 * you need to access high mem.
4497 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4501 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4504 return (unsigned long) page_address(page);
4506 EXPORT_SYMBOL(__get_free_pages);
4508 unsigned long get_zeroed_page(gfp_t gfp_mask)
4510 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4512 EXPORT_SYMBOL(get_zeroed_page);
4514 static inline void free_the_page(struct page *page, unsigned int order)
4516 if (order == 0) /* Via pcp? */
4517 free_unref_page(page);
4519 __free_pages_ok(page, order);
4522 void __free_pages(struct page *page, unsigned int order)
4524 if (put_page_testzero(page))
4525 free_the_page(page, order);
4527 EXPORT_SYMBOL(__free_pages);
4529 void free_pages(unsigned long addr, unsigned int order)
4532 VM_BUG_ON(!virt_addr_valid((void *)addr));
4533 __free_pages(virt_to_page((void *)addr), order);
4537 EXPORT_SYMBOL(free_pages);
4541 * An arbitrary-length arbitrary-offset area of memory which resides
4542 * within a 0 or higher order page. Multiple fragments within that page
4543 * are individually refcounted, in the page's reference counter.
4545 * The page_frag functions below provide a simple allocation framework for
4546 * page fragments. This is used by the network stack and network device
4547 * drivers to provide a backing region of memory for use as either an
4548 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4550 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4553 struct page *page = NULL;
4554 gfp_t gfp = gfp_mask;
4556 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4557 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4559 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4560 PAGE_FRAG_CACHE_MAX_ORDER);
4561 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4563 if (unlikely(!page))
4564 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4566 nc->va = page ? page_address(page) : NULL;
4571 void __page_frag_cache_drain(struct page *page, unsigned int count)
4573 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4575 if (page_ref_sub_and_test(page, count))
4576 free_the_page(page, compound_order(page));
4578 EXPORT_SYMBOL(__page_frag_cache_drain);
4580 void *page_frag_alloc(struct page_frag_cache *nc,
4581 unsigned int fragsz, gfp_t gfp_mask)
4583 unsigned int size = PAGE_SIZE;
4587 if (unlikely(!nc->va)) {
4589 page = __page_frag_cache_refill(nc, gfp_mask);
4593 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4594 /* if size can vary use size else just use PAGE_SIZE */
4597 /* Even if we own the page, we do not use atomic_set().
4598 * This would break get_page_unless_zero() users.
4600 page_ref_add(page, size - 1);
4602 /* reset page count bias and offset to start of new frag */
4603 nc->pfmemalloc = page_is_pfmemalloc(page);
4604 nc->pagecnt_bias = size;
4608 offset = nc->offset - fragsz;
4609 if (unlikely(offset < 0)) {
4610 page = virt_to_page(nc->va);
4612 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4615 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4616 /* if size can vary use size else just use PAGE_SIZE */
4619 /* OK, page count is 0, we can safely set it */
4620 set_page_count(page, size);
4622 /* reset page count bias and offset to start of new frag */
4623 nc->pagecnt_bias = size;
4624 offset = size - fragsz;
4628 nc->offset = offset;
4630 return nc->va + offset;
4632 EXPORT_SYMBOL(page_frag_alloc);
4635 * Frees a page fragment allocated out of either a compound or order 0 page.
4637 void page_frag_free(void *addr)
4639 struct page *page = virt_to_head_page(addr);
4641 if (unlikely(put_page_testzero(page)))
4642 free_the_page(page, compound_order(page));
4644 EXPORT_SYMBOL(page_frag_free);
4646 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4650 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4651 unsigned long used = addr + PAGE_ALIGN(size);
4653 split_page(virt_to_page((void *)addr), order);
4654 while (used < alloc_end) {
4659 return (void *)addr;
4663 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4664 * @size: the number of bytes to allocate
4665 * @gfp_mask: GFP flags for the allocation
4667 * This function is similar to alloc_pages(), except that it allocates the
4668 * minimum number of pages to satisfy the request. alloc_pages() can only
4669 * allocate memory in power-of-two pages.
4671 * This function is also limited by MAX_ORDER.
4673 * Memory allocated by this function must be released by free_pages_exact().
4675 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4677 unsigned int order = get_order(size);
4680 addr = __get_free_pages(gfp_mask, order);
4681 return make_alloc_exact(addr, order, size);
4683 EXPORT_SYMBOL(alloc_pages_exact);
4686 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4688 * @nid: the preferred node ID where memory should be allocated
4689 * @size: the number of bytes to allocate
4690 * @gfp_mask: GFP flags for the allocation
4692 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4695 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4697 unsigned int order = get_order(size);
4698 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4701 return make_alloc_exact((unsigned long)page_address(p), order, size);
4705 * free_pages_exact - release memory allocated via alloc_pages_exact()
4706 * @virt: the value returned by alloc_pages_exact.
4707 * @size: size of allocation, same value as passed to alloc_pages_exact().
4709 * Release the memory allocated by a previous call to alloc_pages_exact.
4711 void free_pages_exact(void *virt, size_t size)
4713 unsigned long addr = (unsigned long)virt;
4714 unsigned long end = addr + PAGE_ALIGN(size);
4716 while (addr < end) {
4721 EXPORT_SYMBOL(free_pages_exact);
4724 * nr_free_zone_pages - count number of pages beyond high watermark
4725 * @offset: The zone index of the highest zone
4727 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4728 * high watermark within all zones at or below a given zone index. For each
4729 * zone, the number of pages is calculated as:
4731 * nr_free_zone_pages = managed_pages - high_pages
4733 static unsigned long nr_free_zone_pages(int offset)
4738 /* Just pick one node, since fallback list is circular */
4739 unsigned long sum = 0;
4741 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4743 for_each_zone_zonelist(zone, z, zonelist, offset) {
4744 unsigned long size = zone_managed_pages(zone);
4745 unsigned long high = high_wmark_pages(zone);
4754 * nr_free_buffer_pages - count number of pages beyond high watermark
4756 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4757 * watermark within ZONE_DMA and ZONE_NORMAL.
4759 unsigned long nr_free_buffer_pages(void)
4761 return nr_free_zone_pages(gfp_zone(GFP_USER));
4763 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4766 * nr_free_pagecache_pages - count number of pages beyond high watermark
4768 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4769 * high watermark within all zones.
4771 unsigned long nr_free_pagecache_pages(void)
4773 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4776 static inline void show_node(struct zone *zone)
4778 if (IS_ENABLED(CONFIG_NUMA))
4779 printk("Node %d ", zone_to_nid(zone));
4782 long si_mem_available(void)
4785 unsigned long pagecache;
4786 unsigned long wmark_low = 0;
4787 unsigned long pages[NR_LRU_LISTS];
4788 unsigned long reclaimable;
4792 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4793 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4796 wmark_low += low_wmark_pages(zone);
4799 * Estimate the amount of memory available for userspace allocations,
4800 * without causing swapping.
4802 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4805 * Not all the page cache can be freed, otherwise the system will
4806 * start swapping. Assume at least half of the page cache, or the
4807 * low watermark worth of cache, needs to stay.
4809 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4810 pagecache -= min(pagecache / 2, wmark_low);
4811 available += pagecache;
4814 * Part of the reclaimable slab and other kernel memory consists of
4815 * items that are in use, and cannot be freed. Cap this estimate at the
4818 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4819 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4820 available += reclaimable - min(reclaimable / 2, wmark_low);
4826 EXPORT_SYMBOL_GPL(si_mem_available);
4828 void si_meminfo(struct sysinfo *val)
4830 val->totalram = totalram_pages();
4831 val->sharedram = global_node_page_state(NR_SHMEM);
4832 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4833 val->bufferram = nr_blockdev_pages();
4834 val->totalhigh = totalhigh_pages();
4835 val->freehigh = nr_free_highpages();
4836 val->mem_unit = PAGE_SIZE;
4839 EXPORT_SYMBOL(si_meminfo);
4842 void si_meminfo_node(struct sysinfo *val, int nid)
4844 int zone_type; /* needs to be signed */
4845 unsigned long managed_pages = 0;
4846 unsigned long managed_highpages = 0;
4847 unsigned long free_highpages = 0;
4848 pg_data_t *pgdat = NODE_DATA(nid);
4850 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4851 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
4852 val->totalram = managed_pages;
4853 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4854 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4855 #ifdef CONFIG_HIGHMEM
4856 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4857 struct zone *zone = &pgdat->node_zones[zone_type];
4859 if (is_highmem(zone)) {
4860 managed_highpages += zone_managed_pages(zone);
4861 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4864 val->totalhigh = managed_highpages;
4865 val->freehigh = free_highpages;
4867 val->totalhigh = managed_highpages;
4868 val->freehigh = free_highpages;
4870 val->mem_unit = PAGE_SIZE;
4875 * Determine whether the node should be displayed or not, depending on whether
4876 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4878 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4880 if (!(flags & SHOW_MEM_FILTER_NODES))
4884 * no node mask - aka implicit memory numa policy. Do not bother with
4885 * the synchronization - read_mems_allowed_begin - because we do not
4886 * have to be precise here.
4889 nodemask = &cpuset_current_mems_allowed;
4891 return !node_isset(nid, *nodemask);
4894 #define K(x) ((x) << (PAGE_SHIFT-10))
4896 static void show_migration_types(unsigned char type)
4898 static const char types[MIGRATE_TYPES] = {
4899 [MIGRATE_UNMOVABLE] = 'U',
4900 [MIGRATE_MOVABLE] = 'M',
4901 [MIGRATE_RECLAIMABLE] = 'E',
4902 [MIGRATE_HIGHATOMIC] = 'H',
4904 [MIGRATE_CMA] = 'C',
4906 #ifdef CONFIG_MEMORY_ISOLATION
4907 [MIGRATE_ISOLATE] = 'I',
4910 char tmp[MIGRATE_TYPES + 1];
4914 for (i = 0; i < MIGRATE_TYPES; i++) {
4915 if (type & (1 << i))
4920 printk(KERN_CONT "(%s) ", tmp);
4924 * Show free area list (used inside shift_scroll-lock stuff)
4925 * We also calculate the percentage fragmentation. We do this by counting the
4926 * memory on each free list with the exception of the first item on the list.
4929 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4932 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4934 unsigned long free_pcp = 0;
4939 for_each_populated_zone(zone) {
4940 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4943 for_each_online_cpu(cpu)
4944 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4947 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4948 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4949 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4950 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4951 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4952 " free:%lu free_pcp:%lu free_cma:%lu\n",
4953 global_node_page_state(NR_ACTIVE_ANON),
4954 global_node_page_state(NR_INACTIVE_ANON),
4955 global_node_page_state(NR_ISOLATED_ANON),
4956 global_node_page_state(NR_ACTIVE_FILE),
4957 global_node_page_state(NR_INACTIVE_FILE),
4958 global_node_page_state(NR_ISOLATED_FILE),
4959 global_node_page_state(NR_UNEVICTABLE),
4960 global_node_page_state(NR_FILE_DIRTY),
4961 global_node_page_state(NR_WRITEBACK),
4962 global_node_page_state(NR_UNSTABLE_NFS),
4963 global_node_page_state(NR_SLAB_RECLAIMABLE),
4964 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4965 global_node_page_state(NR_FILE_MAPPED),
4966 global_node_page_state(NR_SHMEM),
4967 global_zone_page_state(NR_PAGETABLE),
4968 global_zone_page_state(NR_BOUNCE),
4969 global_zone_page_state(NR_FREE_PAGES),
4971 global_zone_page_state(NR_FREE_CMA_PAGES));
4973 for_each_online_pgdat(pgdat) {
4974 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4978 " active_anon:%lukB"
4979 " inactive_anon:%lukB"
4980 " active_file:%lukB"
4981 " inactive_file:%lukB"
4982 " unevictable:%lukB"
4983 " isolated(anon):%lukB"
4984 " isolated(file):%lukB"
4989 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4991 " shmem_pmdmapped: %lukB"
4994 " writeback_tmp:%lukB"
4996 " all_unreclaimable? %s"
4999 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5000 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5001 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5002 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5003 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5004 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5005 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5006 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5007 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5008 K(node_page_state(pgdat, NR_WRITEBACK)),
5009 K(node_page_state(pgdat, NR_SHMEM)),
5010 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5011 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5012 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5014 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5016 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5017 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5018 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5022 for_each_populated_zone(zone) {
5025 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5029 for_each_online_cpu(cpu)
5030 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5039 " active_anon:%lukB"
5040 " inactive_anon:%lukB"
5041 " active_file:%lukB"
5042 " inactive_file:%lukB"
5043 " unevictable:%lukB"
5044 " writepending:%lukB"
5048 " kernel_stack:%lukB"
5056 K(zone_page_state(zone, NR_FREE_PAGES)),
5057 K(min_wmark_pages(zone)),
5058 K(low_wmark_pages(zone)),
5059 K(high_wmark_pages(zone)),
5060 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5061 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5062 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5063 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5064 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5065 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5066 K(zone->present_pages),
5067 K(zone_managed_pages(zone)),
5068 K(zone_page_state(zone, NR_MLOCK)),
5069 zone_page_state(zone, NR_KERNEL_STACK_KB),
5070 K(zone_page_state(zone, NR_PAGETABLE)),
5071 K(zone_page_state(zone, NR_BOUNCE)),
5073 K(this_cpu_read(zone->pageset->pcp.count)),
5074 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5075 printk("lowmem_reserve[]:");
5076 for (i = 0; i < MAX_NR_ZONES; i++)
5077 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5078 printk(KERN_CONT "\n");
5081 for_each_populated_zone(zone) {
5083 unsigned long nr[MAX_ORDER], flags, total = 0;
5084 unsigned char types[MAX_ORDER];
5086 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5089 printk(KERN_CONT "%s: ", zone->name);
5091 spin_lock_irqsave(&zone->lock, flags);
5092 for (order = 0; order < MAX_ORDER; order++) {
5093 struct free_area *area = &zone->free_area[order];
5096 nr[order] = area->nr_free;
5097 total += nr[order] << order;
5100 for (type = 0; type < MIGRATE_TYPES; type++) {
5101 if (!list_empty(&area->free_list[type]))
5102 types[order] |= 1 << type;
5105 spin_unlock_irqrestore(&zone->lock, flags);
5106 for (order = 0; order < MAX_ORDER; order++) {
5107 printk(KERN_CONT "%lu*%lukB ",
5108 nr[order], K(1UL) << order);
5110 show_migration_types(types[order]);
5112 printk(KERN_CONT "= %lukB\n", K(total));
5115 hugetlb_show_meminfo();
5117 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5119 show_swap_cache_info();
5122 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5124 zoneref->zone = zone;
5125 zoneref->zone_idx = zone_idx(zone);
5129 * Builds allocation fallback zone lists.
5131 * Add all populated zones of a node to the zonelist.
5133 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5136 enum zone_type zone_type = MAX_NR_ZONES;
5141 zone = pgdat->node_zones + zone_type;
5142 if (managed_zone(zone)) {
5143 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5144 check_highest_zone(zone_type);
5146 } while (zone_type);
5153 static int __parse_numa_zonelist_order(char *s)
5156 * We used to support different zonlists modes but they turned
5157 * out to be just not useful. Let's keep the warning in place
5158 * if somebody still use the cmd line parameter so that we do
5159 * not fail it silently
5161 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5162 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5168 static __init int setup_numa_zonelist_order(char *s)
5173 return __parse_numa_zonelist_order(s);
5175 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5177 char numa_zonelist_order[] = "Node";
5180 * sysctl handler for numa_zonelist_order
5182 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5183 void __user *buffer, size_t *length,
5190 return proc_dostring(table, write, buffer, length, ppos);
5191 str = memdup_user_nul(buffer, 16);
5193 return PTR_ERR(str);
5195 ret = __parse_numa_zonelist_order(str);
5201 #define MAX_NODE_LOAD (nr_online_nodes)
5202 static int node_load[MAX_NUMNODES];
5205 * find_next_best_node - find the next node that should appear in a given node's fallback list
5206 * @node: node whose fallback list we're appending
5207 * @used_node_mask: nodemask_t of already used nodes
5209 * We use a number of factors to determine which is the next node that should
5210 * appear on a given node's fallback list. The node should not have appeared
5211 * already in @node's fallback list, and it should be the next closest node
5212 * according to the distance array (which contains arbitrary distance values
5213 * from each node to each node in the system), and should also prefer nodes
5214 * with no CPUs, since presumably they'll have very little allocation pressure
5215 * on them otherwise.
5216 * It returns -1 if no node is found.
5218 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5221 int min_val = INT_MAX;
5222 int best_node = NUMA_NO_NODE;
5223 const struct cpumask *tmp = cpumask_of_node(0);
5225 /* Use the local node if we haven't already */
5226 if (!node_isset(node, *used_node_mask)) {
5227 node_set(node, *used_node_mask);
5231 for_each_node_state(n, N_MEMORY) {
5233 /* Don't want a node to appear more than once */
5234 if (node_isset(n, *used_node_mask))
5237 /* Use the distance array to find the distance */
5238 val = node_distance(node, n);
5240 /* Penalize nodes under us ("prefer the next node") */
5243 /* Give preference to headless and unused nodes */
5244 tmp = cpumask_of_node(n);
5245 if (!cpumask_empty(tmp))
5246 val += PENALTY_FOR_NODE_WITH_CPUS;
5248 /* Slight preference for less loaded node */
5249 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5250 val += node_load[n];
5252 if (val < min_val) {
5259 node_set(best_node, *used_node_mask);
5266 * Build zonelists ordered by node and zones within node.
5267 * This results in maximum locality--normal zone overflows into local
5268 * DMA zone, if any--but risks exhausting DMA zone.
5270 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5273 struct zoneref *zonerefs;
5276 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5278 for (i = 0; i < nr_nodes; i++) {
5281 pg_data_t *node = NODE_DATA(node_order[i]);
5283 nr_zones = build_zonerefs_node(node, zonerefs);
5284 zonerefs += nr_zones;
5286 zonerefs->zone = NULL;
5287 zonerefs->zone_idx = 0;
5291 * Build gfp_thisnode zonelists
5293 static void build_thisnode_zonelists(pg_data_t *pgdat)
5295 struct zoneref *zonerefs;
5298 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5299 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5300 zonerefs += nr_zones;
5301 zonerefs->zone = NULL;
5302 zonerefs->zone_idx = 0;
5306 * Build zonelists ordered by zone and nodes within zones.
5307 * This results in conserving DMA zone[s] until all Normal memory is
5308 * exhausted, but results in overflowing to remote node while memory
5309 * may still exist in local DMA zone.
5312 static void build_zonelists(pg_data_t *pgdat)
5314 static int node_order[MAX_NUMNODES];
5315 int node, load, nr_nodes = 0;
5316 nodemask_t used_mask;
5317 int local_node, prev_node;
5319 /* NUMA-aware ordering of nodes */
5320 local_node = pgdat->node_id;
5321 load = nr_online_nodes;
5322 prev_node = local_node;
5323 nodes_clear(used_mask);
5325 memset(node_order, 0, sizeof(node_order));
5326 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5328 * We don't want to pressure a particular node.
5329 * So adding penalty to the first node in same
5330 * distance group to make it round-robin.
5332 if (node_distance(local_node, node) !=
5333 node_distance(local_node, prev_node))
5334 node_load[node] = load;
5336 node_order[nr_nodes++] = node;
5341 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5342 build_thisnode_zonelists(pgdat);
5345 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5347 * Return node id of node used for "local" allocations.
5348 * I.e., first node id of first zone in arg node's generic zonelist.
5349 * Used for initializing percpu 'numa_mem', which is used primarily
5350 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5352 int local_memory_node(int node)
5356 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5357 gfp_zone(GFP_KERNEL),
5359 return zone_to_nid(z->zone);
5363 static void setup_min_unmapped_ratio(void);
5364 static void setup_min_slab_ratio(void);
5365 #else /* CONFIG_NUMA */
5367 static void build_zonelists(pg_data_t *pgdat)
5369 int node, local_node;
5370 struct zoneref *zonerefs;
5373 local_node = pgdat->node_id;
5375 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5376 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5377 zonerefs += nr_zones;
5380 * Now we build the zonelist so that it contains the zones
5381 * of all the other nodes.
5382 * We don't want to pressure a particular node, so when
5383 * building the zones for node N, we make sure that the
5384 * zones coming right after the local ones are those from
5385 * node N+1 (modulo N)
5387 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5388 if (!node_online(node))
5390 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5391 zonerefs += nr_zones;
5393 for (node = 0; node < local_node; node++) {
5394 if (!node_online(node))
5396 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5397 zonerefs += nr_zones;
5400 zonerefs->zone = NULL;
5401 zonerefs->zone_idx = 0;
5404 #endif /* CONFIG_NUMA */
5407 * Boot pageset table. One per cpu which is going to be used for all
5408 * zones and all nodes. The parameters will be set in such a way
5409 * that an item put on a list will immediately be handed over to
5410 * the buddy list. This is safe since pageset manipulation is done
5411 * with interrupts disabled.
5413 * The boot_pagesets must be kept even after bootup is complete for
5414 * unused processors and/or zones. They do play a role for bootstrapping
5415 * hotplugged processors.
5417 * zoneinfo_show() and maybe other functions do
5418 * not check if the processor is online before following the pageset pointer.
5419 * Other parts of the kernel may not check if the zone is available.
5421 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5422 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5423 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5425 static void __build_all_zonelists(void *data)
5428 int __maybe_unused cpu;
5429 pg_data_t *self = data;
5430 static DEFINE_SPINLOCK(lock);
5435 memset(node_load, 0, sizeof(node_load));
5439 * This node is hotadded and no memory is yet present. So just
5440 * building zonelists is fine - no need to touch other nodes.
5442 if (self && !node_online(self->node_id)) {
5443 build_zonelists(self);
5445 for_each_online_node(nid) {
5446 pg_data_t *pgdat = NODE_DATA(nid);
5448 build_zonelists(pgdat);
5451 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5453 * We now know the "local memory node" for each node--
5454 * i.e., the node of the first zone in the generic zonelist.
5455 * Set up numa_mem percpu variable for on-line cpus. During
5456 * boot, only the boot cpu should be on-line; we'll init the
5457 * secondary cpus' numa_mem as they come on-line. During
5458 * node/memory hotplug, we'll fixup all on-line cpus.
5460 for_each_online_cpu(cpu)
5461 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5468 static noinline void __init
5469 build_all_zonelists_init(void)
5473 __build_all_zonelists(NULL);
5476 * Initialize the boot_pagesets that are going to be used
5477 * for bootstrapping processors. The real pagesets for
5478 * each zone will be allocated later when the per cpu
5479 * allocator is available.
5481 * boot_pagesets are used also for bootstrapping offline
5482 * cpus if the system is already booted because the pagesets
5483 * are needed to initialize allocators on a specific cpu too.
5484 * F.e. the percpu allocator needs the page allocator which
5485 * needs the percpu allocator in order to allocate its pagesets
5486 * (a chicken-egg dilemma).
5488 for_each_possible_cpu(cpu)
5489 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5491 mminit_verify_zonelist();
5492 cpuset_init_current_mems_allowed();
5496 * unless system_state == SYSTEM_BOOTING.
5498 * __ref due to call of __init annotated helper build_all_zonelists_init
5499 * [protected by SYSTEM_BOOTING].
5501 void __ref build_all_zonelists(pg_data_t *pgdat)
5503 if (system_state == SYSTEM_BOOTING) {
5504 build_all_zonelists_init();
5506 __build_all_zonelists(pgdat);
5507 /* cpuset refresh routine should be here */
5509 vm_total_pages = nr_free_pagecache_pages();
5511 * Disable grouping by mobility if the number of pages in the
5512 * system is too low to allow the mechanism to work. It would be
5513 * more accurate, but expensive to check per-zone. This check is
5514 * made on memory-hotadd so a system can start with mobility
5515 * disabled and enable it later
5517 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5518 page_group_by_mobility_disabled = 1;
5520 page_group_by_mobility_disabled = 0;
5522 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5524 page_group_by_mobility_disabled ? "off" : "on",
5527 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5531 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5532 static bool __meminit
5533 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5535 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5536 static struct memblock_region *r;
5538 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5539 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5540 for_each_memblock(memory, r) {
5541 if (*pfn < memblock_region_memory_end_pfn(r))
5545 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5546 memblock_is_mirror(r)) {
5547 *pfn = memblock_region_memory_end_pfn(r);
5556 * Initially all pages are reserved - free ones are freed
5557 * up by memblock_free_all() once the early boot process is
5558 * done. Non-atomic initialization, single-pass.
5560 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5561 unsigned long start_pfn, enum memmap_context context,
5562 struct vmem_altmap *altmap)
5564 unsigned long pfn, end_pfn = start_pfn + size;
5567 if (highest_memmap_pfn < end_pfn - 1)
5568 highest_memmap_pfn = end_pfn - 1;
5570 #ifdef CONFIG_ZONE_DEVICE
5572 * Honor reservation requested by the driver for this ZONE_DEVICE
5573 * memory. We limit the total number of pages to initialize to just
5574 * those that might contain the memory mapping. We will defer the
5575 * ZONE_DEVICE page initialization until after we have released
5578 if (zone == ZONE_DEVICE) {
5582 if (start_pfn == altmap->base_pfn)
5583 start_pfn += altmap->reserve;
5584 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5588 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5590 * There can be holes in boot-time mem_map[]s handed to this
5591 * function. They do not exist on hotplugged memory.
5593 if (context == MEMMAP_EARLY) {
5594 if (!early_pfn_valid(pfn))
5596 if (!early_pfn_in_nid(pfn, nid))
5598 if (overlap_memmap_init(zone, &pfn))
5600 if (defer_init(nid, pfn, end_pfn))
5604 page = pfn_to_page(pfn);
5605 __init_single_page(page, pfn, zone, nid);
5606 if (context == MEMMAP_HOTPLUG)
5607 __SetPageReserved(page);
5610 * Mark the block movable so that blocks are reserved for
5611 * movable at startup. This will force kernel allocations
5612 * to reserve their blocks rather than leaking throughout
5613 * the address space during boot when many long-lived
5614 * kernel allocations are made.
5616 * bitmap is created for zone's valid pfn range. but memmap
5617 * can be created for invalid pages (for alignment)
5618 * check here not to call set_pageblock_migratetype() against
5621 if (!(pfn & (pageblock_nr_pages - 1))) {
5622 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5626 #ifdef CONFIG_SPARSEMEM
5628 * If the zone does not span the rest of the section then
5629 * we should at least initialize those pages. Otherwise we
5630 * could blow up on a poisoned page in some paths which depend
5631 * on full sections being initialized (e.g. memory hotplug).
5633 while (end_pfn % PAGES_PER_SECTION) {
5634 __init_single_page(pfn_to_page(end_pfn), end_pfn, zone, nid);
5640 #ifdef CONFIG_ZONE_DEVICE
5641 void __ref memmap_init_zone_device(struct zone *zone,
5642 unsigned long start_pfn,
5644 struct dev_pagemap *pgmap)
5646 unsigned long pfn, end_pfn = start_pfn + size;
5647 struct pglist_data *pgdat = zone->zone_pgdat;
5648 unsigned long zone_idx = zone_idx(zone);
5649 unsigned long start = jiffies;
5650 int nid = pgdat->node_id;
5652 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5656 * The call to memmap_init_zone should have already taken care
5657 * of the pages reserved for the memmap, so we can just jump to
5658 * the end of that region and start processing the device pages.
5660 if (pgmap->altmap_valid) {
5661 struct vmem_altmap *altmap = &pgmap->altmap;
5663 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5664 size = end_pfn - start_pfn;
5667 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5668 struct page *page = pfn_to_page(pfn);
5670 __init_single_page(page, pfn, zone_idx, nid);
5673 * Mark page reserved as it will need to wait for onlining
5674 * phase for it to be fully associated with a zone.
5676 * We can use the non-atomic __set_bit operation for setting
5677 * the flag as we are still initializing the pages.
5679 __SetPageReserved(page);
5682 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5683 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5684 * page is ever freed or placed on a driver-private list.
5686 page->pgmap = pgmap;
5690 * Mark the block movable so that blocks are reserved for
5691 * movable at startup. This will force kernel allocations
5692 * to reserve their blocks rather than leaking throughout
5693 * the address space during boot when many long-lived
5694 * kernel allocations are made.
5696 * bitmap is created for zone's valid pfn range. but memmap
5697 * can be created for invalid pages (for alignment)
5698 * check here not to call set_pageblock_migratetype() against
5701 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5702 * because this is done early in sparse_add_one_section
5704 if (!(pfn & (pageblock_nr_pages - 1))) {
5705 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5710 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5711 size, jiffies_to_msecs(jiffies - start));
5715 static void __meminit zone_init_free_lists(struct zone *zone)
5717 unsigned int order, t;
5718 for_each_migratetype_order(order, t) {
5719 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5720 zone->free_area[order].nr_free = 0;
5724 void __meminit __weak memmap_init(unsigned long size, int nid,
5725 unsigned long zone, unsigned long start_pfn)
5727 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5730 static int zone_batchsize(struct zone *zone)
5736 * The per-cpu-pages pools are set to around 1000th of the
5739 batch = zone_managed_pages(zone) / 1024;
5740 /* But no more than a meg. */
5741 if (batch * PAGE_SIZE > 1024 * 1024)
5742 batch = (1024 * 1024) / PAGE_SIZE;
5743 batch /= 4; /* We effectively *= 4 below */
5748 * Clamp the batch to a 2^n - 1 value. Having a power
5749 * of 2 value was found to be more likely to have
5750 * suboptimal cache aliasing properties in some cases.
5752 * For example if 2 tasks are alternately allocating
5753 * batches of pages, one task can end up with a lot
5754 * of pages of one half of the possible page colors
5755 * and the other with pages of the other colors.
5757 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5762 /* The deferral and batching of frees should be suppressed under NOMMU
5765 * The problem is that NOMMU needs to be able to allocate large chunks
5766 * of contiguous memory as there's no hardware page translation to
5767 * assemble apparent contiguous memory from discontiguous pages.
5769 * Queueing large contiguous runs of pages for batching, however,
5770 * causes the pages to actually be freed in smaller chunks. As there
5771 * can be a significant delay between the individual batches being
5772 * recycled, this leads to the once large chunks of space being
5773 * fragmented and becoming unavailable for high-order allocations.
5780 * pcp->high and pcp->batch values are related and dependent on one another:
5781 * ->batch must never be higher then ->high.
5782 * The following function updates them in a safe manner without read side
5785 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5786 * those fields changing asynchronously (acording the the above rule).
5788 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5789 * outside of boot time (or some other assurance that no concurrent updaters
5792 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5793 unsigned long batch)
5795 /* start with a fail safe value for batch */
5799 /* Update high, then batch, in order */
5806 /* a companion to pageset_set_high() */
5807 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5809 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5812 static void pageset_init(struct per_cpu_pageset *p)
5814 struct per_cpu_pages *pcp;
5817 memset(p, 0, sizeof(*p));
5820 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5821 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5824 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5827 pageset_set_batch(p, batch);
5831 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5832 * to the value high for the pageset p.
5834 static void pageset_set_high(struct per_cpu_pageset *p,
5837 unsigned long batch = max(1UL, high / 4);
5838 if ((high / 4) > (PAGE_SHIFT * 8))
5839 batch = PAGE_SHIFT * 8;
5841 pageset_update(&p->pcp, high, batch);
5844 static void pageset_set_high_and_batch(struct zone *zone,
5845 struct per_cpu_pageset *pcp)
5847 if (percpu_pagelist_fraction)
5848 pageset_set_high(pcp,
5849 (zone_managed_pages(zone) /
5850 percpu_pagelist_fraction));
5852 pageset_set_batch(pcp, zone_batchsize(zone));
5855 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5857 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5860 pageset_set_high_and_batch(zone, pcp);
5863 void __meminit setup_zone_pageset(struct zone *zone)
5866 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5867 for_each_possible_cpu(cpu)
5868 zone_pageset_init(zone, cpu);
5872 * Allocate per cpu pagesets and initialize them.
5873 * Before this call only boot pagesets were available.
5875 void __init setup_per_cpu_pageset(void)
5877 struct pglist_data *pgdat;
5880 for_each_populated_zone(zone)
5881 setup_zone_pageset(zone);
5883 for_each_online_pgdat(pgdat)
5884 pgdat->per_cpu_nodestats =
5885 alloc_percpu(struct per_cpu_nodestat);
5888 static __meminit void zone_pcp_init(struct zone *zone)
5891 * per cpu subsystem is not up at this point. The following code
5892 * relies on the ability of the linker to provide the
5893 * offset of a (static) per cpu variable into the per cpu area.
5895 zone->pageset = &boot_pageset;
5897 if (populated_zone(zone))
5898 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5899 zone->name, zone->present_pages,
5900 zone_batchsize(zone));
5903 void __meminit init_currently_empty_zone(struct zone *zone,
5904 unsigned long zone_start_pfn,
5907 struct pglist_data *pgdat = zone->zone_pgdat;
5908 int zone_idx = zone_idx(zone) + 1;
5910 if (zone_idx > pgdat->nr_zones)
5911 pgdat->nr_zones = zone_idx;
5913 zone->zone_start_pfn = zone_start_pfn;
5915 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5916 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5918 (unsigned long)zone_idx(zone),
5919 zone_start_pfn, (zone_start_pfn + size));
5921 zone_init_free_lists(zone);
5922 zone->initialized = 1;
5925 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5926 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5929 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5931 int __meminit __early_pfn_to_nid(unsigned long pfn,
5932 struct mminit_pfnnid_cache *state)
5934 unsigned long start_pfn, end_pfn;
5937 if (state->last_start <= pfn && pfn < state->last_end)
5938 return state->last_nid;
5940 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5942 state->last_start = start_pfn;
5943 state->last_end = end_pfn;
5944 state->last_nid = nid;
5949 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5952 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5953 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5954 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5956 * If an architecture guarantees that all ranges registered contain no holes
5957 * and may be freed, this this function may be used instead of calling
5958 * memblock_free_early_nid() manually.
5960 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5962 unsigned long start_pfn, end_pfn;
5965 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5966 start_pfn = min(start_pfn, max_low_pfn);
5967 end_pfn = min(end_pfn, max_low_pfn);
5969 if (start_pfn < end_pfn)
5970 memblock_free_early_nid(PFN_PHYS(start_pfn),
5971 (end_pfn - start_pfn) << PAGE_SHIFT,
5977 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5978 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5980 * If an architecture guarantees that all ranges registered contain no holes and may
5981 * be freed, this function may be used instead of calling memory_present() manually.
5983 void __init sparse_memory_present_with_active_regions(int nid)
5985 unsigned long start_pfn, end_pfn;
5988 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5989 memory_present(this_nid, start_pfn, end_pfn);
5993 * get_pfn_range_for_nid - Return the start and end page frames for a node
5994 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5995 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5996 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5998 * It returns the start and end page frame of a node based on information
5999 * provided by memblock_set_node(). If called for a node
6000 * with no available memory, a warning is printed and the start and end
6003 void __meminit get_pfn_range_for_nid(unsigned int nid,
6004 unsigned long *start_pfn, unsigned long *end_pfn)
6006 unsigned long this_start_pfn, this_end_pfn;
6012 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6013 *start_pfn = min(*start_pfn, this_start_pfn);
6014 *end_pfn = max(*end_pfn, this_end_pfn);
6017 if (*start_pfn == -1UL)
6022 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6023 * assumption is made that zones within a node are ordered in monotonic
6024 * increasing memory addresses so that the "highest" populated zone is used
6026 static void __init find_usable_zone_for_movable(void)
6029 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6030 if (zone_index == ZONE_MOVABLE)
6033 if (arch_zone_highest_possible_pfn[zone_index] >
6034 arch_zone_lowest_possible_pfn[zone_index])
6038 VM_BUG_ON(zone_index == -1);
6039 movable_zone = zone_index;
6043 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6044 * because it is sized independent of architecture. Unlike the other zones,
6045 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6046 * in each node depending on the size of each node and how evenly kernelcore
6047 * is distributed. This helper function adjusts the zone ranges
6048 * provided by the architecture for a given node by using the end of the
6049 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6050 * zones within a node are in order of monotonic increases memory addresses
6052 static void __meminit adjust_zone_range_for_zone_movable(int nid,
6053 unsigned long zone_type,
6054 unsigned long node_start_pfn,
6055 unsigned long node_end_pfn,
6056 unsigned long *zone_start_pfn,
6057 unsigned long *zone_end_pfn)
6059 /* Only adjust if ZONE_MOVABLE is on this node */
6060 if (zone_movable_pfn[nid]) {
6061 /* Size ZONE_MOVABLE */
6062 if (zone_type == ZONE_MOVABLE) {
6063 *zone_start_pfn = zone_movable_pfn[nid];
6064 *zone_end_pfn = min(node_end_pfn,
6065 arch_zone_highest_possible_pfn[movable_zone]);
6067 /* Adjust for ZONE_MOVABLE starting within this range */
6068 } else if (!mirrored_kernelcore &&
6069 *zone_start_pfn < zone_movable_pfn[nid] &&
6070 *zone_end_pfn > zone_movable_pfn[nid]) {
6071 *zone_end_pfn = zone_movable_pfn[nid];
6073 /* Check if this whole range is within ZONE_MOVABLE */
6074 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6075 *zone_start_pfn = *zone_end_pfn;
6080 * Return the number of pages a zone spans in a node, including holes
6081 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6083 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
6084 unsigned long zone_type,
6085 unsigned long node_start_pfn,
6086 unsigned long node_end_pfn,
6087 unsigned long *zone_start_pfn,
6088 unsigned long *zone_end_pfn,
6089 unsigned long *ignored)
6091 /* When hotadd a new node from cpu_up(), the node should be empty */
6092 if (!node_start_pfn && !node_end_pfn)
6095 /* Get the start and end of the zone */
6096 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6097 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6098 adjust_zone_range_for_zone_movable(nid, zone_type,
6099 node_start_pfn, node_end_pfn,
6100 zone_start_pfn, zone_end_pfn);
6102 /* Check that this node has pages within the zone's required range */
6103 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6106 /* Move the zone boundaries inside the node if necessary */
6107 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6108 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6110 /* Return the spanned pages */
6111 return *zone_end_pfn - *zone_start_pfn;
6115 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6116 * then all holes in the requested range will be accounted for.
6118 unsigned long __meminit __absent_pages_in_range(int nid,
6119 unsigned long range_start_pfn,
6120 unsigned long range_end_pfn)
6122 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6123 unsigned long start_pfn, end_pfn;
6126 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6127 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6128 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6129 nr_absent -= end_pfn - start_pfn;
6135 * absent_pages_in_range - Return number of page frames in holes within a range
6136 * @start_pfn: The start PFN to start searching for holes
6137 * @end_pfn: The end PFN to stop searching for holes
6139 * It returns the number of pages frames in memory holes within a range.
6141 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6142 unsigned long end_pfn)
6144 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6147 /* Return the number of page frames in holes in a zone on a node */
6148 static unsigned long __meminit zone_absent_pages_in_node(int nid,
6149 unsigned long zone_type,
6150 unsigned long node_start_pfn,
6151 unsigned long node_end_pfn,
6152 unsigned long *ignored)
6154 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6155 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6156 unsigned long zone_start_pfn, zone_end_pfn;
6157 unsigned long nr_absent;
6159 /* When hotadd a new node from cpu_up(), the node should be empty */
6160 if (!node_start_pfn && !node_end_pfn)
6163 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6164 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6166 adjust_zone_range_for_zone_movable(nid, zone_type,
6167 node_start_pfn, node_end_pfn,
6168 &zone_start_pfn, &zone_end_pfn);
6169 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6172 * ZONE_MOVABLE handling.
6173 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6176 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6177 unsigned long start_pfn, end_pfn;
6178 struct memblock_region *r;
6180 for_each_memblock(memory, r) {
6181 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6182 zone_start_pfn, zone_end_pfn);
6183 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6184 zone_start_pfn, zone_end_pfn);
6186 if (zone_type == ZONE_MOVABLE &&
6187 memblock_is_mirror(r))
6188 nr_absent += end_pfn - start_pfn;
6190 if (zone_type == ZONE_NORMAL &&
6191 !memblock_is_mirror(r))
6192 nr_absent += end_pfn - start_pfn;
6199 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6200 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6201 unsigned long zone_type,
6202 unsigned long node_start_pfn,
6203 unsigned long node_end_pfn,
6204 unsigned long *zone_start_pfn,
6205 unsigned long *zone_end_pfn,
6206 unsigned long *zones_size)
6210 *zone_start_pfn = node_start_pfn;
6211 for (zone = 0; zone < zone_type; zone++)
6212 *zone_start_pfn += zones_size[zone];
6214 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6216 return zones_size[zone_type];
6219 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6220 unsigned long zone_type,
6221 unsigned long node_start_pfn,
6222 unsigned long node_end_pfn,
6223 unsigned long *zholes_size)
6228 return zholes_size[zone_type];
6231 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6233 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6234 unsigned long node_start_pfn,
6235 unsigned long node_end_pfn,
6236 unsigned long *zones_size,
6237 unsigned long *zholes_size)
6239 unsigned long realtotalpages = 0, totalpages = 0;
6242 for (i = 0; i < MAX_NR_ZONES; i++) {
6243 struct zone *zone = pgdat->node_zones + i;
6244 unsigned long zone_start_pfn, zone_end_pfn;
6245 unsigned long size, real_size;
6247 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6253 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6254 node_start_pfn, node_end_pfn,
6257 zone->zone_start_pfn = zone_start_pfn;
6259 zone->zone_start_pfn = 0;
6260 zone->spanned_pages = size;
6261 zone->present_pages = real_size;
6264 realtotalpages += real_size;
6267 pgdat->node_spanned_pages = totalpages;
6268 pgdat->node_present_pages = realtotalpages;
6269 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6273 #ifndef CONFIG_SPARSEMEM
6275 * Calculate the size of the zone->blockflags rounded to an unsigned long
6276 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6277 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6278 * round what is now in bits to nearest long in bits, then return it in
6281 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6283 unsigned long usemapsize;
6285 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6286 usemapsize = roundup(zonesize, pageblock_nr_pages);
6287 usemapsize = usemapsize >> pageblock_order;
6288 usemapsize *= NR_PAGEBLOCK_BITS;
6289 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6291 return usemapsize / 8;
6294 static void __ref setup_usemap(struct pglist_data *pgdat,
6296 unsigned long zone_start_pfn,
6297 unsigned long zonesize)
6299 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6300 zone->pageblock_flags = NULL;
6302 zone->pageblock_flags =
6303 memblock_alloc_node_nopanic(usemapsize,
6307 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6308 unsigned long zone_start_pfn, unsigned long zonesize) {}
6309 #endif /* CONFIG_SPARSEMEM */
6311 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6313 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6314 void __init set_pageblock_order(void)
6318 /* Check that pageblock_nr_pages has not already been setup */
6319 if (pageblock_order)
6322 if (HPAGE_SHIFT > PAGE_SHIFT)
6323 order = HUGETLB_PAGE_ORDER;
6325 order = MAX_ORDER - 1;
6328 * Assume the largest contiguous order of interest is a huge page.
6329 * This value may be variable depending on boot parameters on IA64 and
6332 pageblock_order = order;
6334 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6337 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6338 * is unused as pageblock_order is set at compile-time. See
6339 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6342 void __init set_pageblock_order(void)
6346 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6348 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6349 unsigned long present_pages)
6351 unsigned long pages = spanned_pages;
6354 * Provide a more accurate estimation if there are holes within
6355 * the zone and SPARSEMEM is in use. If there are holes within the
6356 * zone, each populated memory region may cost us one or two extra
6357 * memmap pages due to alignment because memmap pages for each
6358 * populated regions may not be naturally aligned on page boundary.
6359 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6361 if (spanned_pages > present_pages + (present_pages >> 4) &&
6362 IS_ENABLED(CONFIG_SPARSEMEM))
6363 pages = present_pages;
6365 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6368 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6369 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6371 spin_lock_init(&pgdat->split_queue_lock);
6372 INIT_LIST_HEAD(&pgdat->split_queue);
6373 pgdat->split_queue_len = 0;
6376 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6379 #ifdef CONFIG_COMPACTION
6380 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6382 init_waitqueue_head(&pgdat->kcompactd_wait);
6385 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6388 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6390 pgdat_resize_init(pgdat);
6392 pgdat_init_split_queue(pgdat);
6393 pgdat_init_kcompactd(pgdat);
6395 init_waitqueue_head(&pgdat->kswapd_wait);
6396 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6398 pgdat_page_ext_init(pgdat);
6399 spin_lock_init(&pgdat->lru_lock);
6400 lruvec_init(node_lruvec(pgdat));
6403 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6404 unsigned long remaining_pages)
6406 atomic_long_set(&zone->managed_pages, remaining_pages);
6407 zone_set_nid(zone, nid);
6408 zone->name = zone_names[idx];
6409 zone->zone_pgdat = NODE_DATA(nid);
6410 spin_lock_init(&zone->lock);
6411 zone_seqlock_init(zone);
6412 zone_pcp_init(zone);
6416 * Set up the zone data structures
6417 * - init pgdat internals
6418 * - init all zones belonging to this node
6420 * NOTE: this function is only called during memory hotplug
6422 #ifdef CONFIG_MEMORY_HOTPLUG
6423 void __ref free_area_init_core_hotplug(int nid)
6426 pg_data_t *pgdat = NODE_DATA(nid);
6428 pgdat_init_internals(pgdat);
6429 for (z = 0; z < MAX_NR_ZONES; z++)
6430 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6435 * Set up the zone data structures:
6436 * - mark all pages reserved
6437 * - mark all memory queues empty
6438 * - clear the memory bitmaps
6440 * NOTE: pgdat should get zeroed by caller.
6441 * NOTE: this function is only called during early init.
6443 static void __init free_area_init_core(struct pglist_data *pgdat)
6446 int nid = pgdat->node_id;
6448 pgdat_init_internals(pgdat);
6449 pgdat->per_cpu_nodestats = &boot_nodestats;
6451 for (j = 0; j < MAX_NR_ZONES; j++) {
6452 struct zone *zone = pgdat->node_zones + j;
6453 unsigned long size, freesize, memmap_pages;
6454 unsigned long zone_start_pfn = zone->zone_start_pfn;
6456 size = zone->spanned_pages;
6457 freesize = zone->present_pages;
6460 * Adjust freesize so that it accounts for how much memory
6461 * is used by this zone for memmap. This affects the watermark
6462 * and per-cpu initialisations
6464 memmap_pages = calc_memmap_size(size, freesize);
6465 if (!is_highmem_idx(j)) {
6466 if (freesize >= memmap_pages) {
6467 freesize -= memmap_pages;
6470 " %s zone: %lu pages used for memmap\n",
6471 zone_names[j], memmap_pages);
6473 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6474 zone_names[j], memmap_pages, freesize);
6477 /* Account for reserved pages */
6478 if (j == 0 && freesize > dma_reserve) {
6479 freesize -= dma_reserve;
6480 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6481 zone_names[0], dma_reserve);
6484 if (!is_highmem_idx(j))
6485 nr_kernel_pages += freesize;
6486 /* Charge for highmem memmap if there are enough kernel pages */
6487 else if (nr_kernel_pages > memmap_pages * 2)
6488 nr_kernel_pages -= memmap_pages;
6489 nr_all_pages += freesize;
6492 * Set an approximate value for lowmem here, it will be adjusted
6493 * when the bootmem allocator frees pages into the buddy system.
6494 * And all highmem pages will be managed by the buddy system.
6496 zone_init_internals(zone, j, nid, freesize);
6501 set_pageblock_order();
6502 setup_usemap(pgdat, zone, zone_start_pfn, size);
6503 init_currently_empty_zone(zone, zone_start_pfn, size);
6504 memmap_init(size, nid, j, zone_start_pfn);
6508 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6509 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6511 unsigned long __maybe_unused start = 0;
6512 unsigned long __maybe_unused offset = 0;
6514 /* Skip empty nodes */
6515 if (!pgdat->node_spanned_pages)
6518 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6519 offset = pgdat->node_start_pfn - start;
6520 /* ia64 gets its own node_mem_map, before this, without bootmem */
6521 if (!pgdat->node_mem_map) {
6522 unsigned long size, end;
6526 * The zone's endpoints aren't required to be MAX_ORDER
6527 * aligned but the node_mem_map endpoints must be in order
6528 * for the buddy allocator to function correctly.
6530 end = pgdat_end_pfn(pgdat);
6531 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6532 size = (end - start) * sizeof(struct page);
6533 map = memblock_alloc_node_nopanic(size, pgdat->node_id);
6534 pgdat->node_mem_map = map + offset;
6536 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6537 __func__, pgdat->node_id, (unsigned long)pgdat,
6538 (unsigned long)pgdat->node_mem_map);
6539 #ifndef CONFIG_NEED_MULTIPLE_NODES
6541 * With no DISCONTIG, the global mem_map is just set as node 0's
6543 if (pgdat == NODE_DATA(0)) {
6544 mem_map = NODE_DATA(0)->node_mem_map;
6545 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6546 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6548 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6553 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6554 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6556 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6557 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6560 * We start only with one section of pages, more pages are added as
6561 * needed until the rest of deferred pages are initialized.
6563 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6564 pgdat->node_spanned_pages);
6565 pgdat->first_deferred_pfn = ULONG_MAX;
6568 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6571 void __init free_area_init_node(int nid, unsigned long *zones_size,
6572 unsigned long node_start_pfn,
6573 unsigned long *zholes_size)
6575 pg_data_t *pgdat = NODE_DATA(nid);
6576 unsigned long start_pfn = 0;
6577 unsigned long end_pfn = 0;
6579 /* pg_data_t should be reset to zero when it's allocated */
6580 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6582 pgdat->node_id = nid;
6583 pgdat->node_start_pfn = node_start_pfn;
6584 pgdat->per_cpu_nodestats = NULL;
6585 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6586 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6587 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6588 (u64)start_pfn << PAGE_SHIFT,
6589 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6591 start_pfn = node_start_pfn;
6593 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6594 zones_size, zholes_size);
6596 alloc_node_mem_map(pgdat);
6597 pgdat_set_deferred_range(pgdat);
6599 free_area_init_core(pgdat);
6602 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6604 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6607 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6612 for (pfn = spfn; pfn < epfn; pfn++) {
6613 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6614 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6615 + pageblock_nr_pages - 1;
6618 mm_zero_struct_page(pfn_to_page(pfn));
6626 * Only struct pages that are backed by physical memory are zeroed and
6627 * initialized by going through __init_single_page(). But, there are some
6628 * struct pages which are reserved in memblock allocator and their fields
6629 * may be accessed (for example page_to_pfn() on some configuration accesses
6630 * flags). We must explicitly zero those struct pages.
6632 * This function also addresses a similar issue where struct pages are left
6633 * uninitialized because the physical address range is not covered by
6634 * memblock.memory or memblock.reserved. That could happen when memblock
6635 * layout is manually configured via memmap=.
6637 void __init zero_resv_unavail(void)
6639 phys_addr_t start, end;
6641 phys_addr_t next = 0;
6644 * Loop through unavailable ranges not covered by memblock.memory.
6647 for_each_mem_range(i, &memblock.memory, NULL,
6648 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6650 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6653 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6656 * Struct pages that do not have backing memory. This could be because
6657 * firmware is using some of this memory, or for some other reasons.
6660 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6662 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6664 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6666 #if MAX_NUMNODES > 1
6668 * Figure out the number of possible node ids.
6670 void __init setup_nr_node_ids(void)
6672 unsigned int highest;
6674 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6675 nr_node_ids = highest + 1;
6680 * node_map_pfn_alignment - determine the maximum internode alignment
6682 * This function should be called after node map is populated and sorted.
6683 * It calculates the maximum power of two alignment which can distinguish
6686 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6687 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6688 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6689 * shifted, 1GiB is enough and this function will indicate so.
6691 * This is used to test whether pfn -> nid mapping of the chosen memory
6692 * model has fine enough granularity to avoid incorrect mapping for the
6693 * populated node map.
6695 * Returns the determined alignment in pfn's. 0 if there is no alignment
6696 * requirement (single node).
6698 unsigned long __init node_map_pfn_alignment(void)
6700 unsigned long accl_mask = 0, last_end = 0;
6701 unsigned long start, end, mask;
6705 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6706 if (!start || last_nid < 0 || last_nid == nid) {
6713 * Start with a mask granular enough to pin-point to the
6714 * start pfn and tick off bits one-by-one until it becomes
6715 * too coarse to separate the current node from the last.
6717 mask = ~((1 << __ffs(start)) - 1);
6718 while (mask && last_end <= (start & (mask << 1)))
6721 /* accumulate all internode masks */
6725 /* convert mask to number of pages */
6726 return ~accl_mask + 1;
6729 /* Find the lowest pfn for a node */
6730 static unsigned long __init find_min_pfn_for_node(int nid)
6732 unsigned long min_pfn = ULONG_MAX;
6733 unsigned long start_pfn;
6736 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6737 min_pfn = min(min_pfn, start_pfn);
6739 if (min_pfn == ULONG_MAX) {
6740 pr_warn("Could not find start_pfn for node %d\n", nid);
6748 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6750 * It returns the minimum PFN based on information provided via
6751 * memblock_set_node().
6753 unsigned long __init find_min_pfn_with_active_regions(void)
6755 return find_min_pfn_for_node(MAX_NUMNODES);
6759 * early_calculate_totalpages()
6760 * Sum pages in active regions for movable zone.
6761 * Populate N_MEMORY for calculating usable_nodes.
6763 static unsigned long __init early_calculate_totalpages(void)
6765 unsigned long totalpages = 0;
6766 unsigned long start_pfn, end_pfn;
6769 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6770 unsigned long pages = end_pfn - start_pfn;
6772 totalpages += pages;
6774 node_set_state(nid, N_MEMORY);
6780 * Find the PFN the Movable zone begins in each node. Kernel memory
6781 * is spread evenly between nodes as long as the nodes have enough
6782 * memory. When they don't, some nodes will have more kernelcore than
6785 static void __init find_zone_movable_pfns_for_nodes(void)
6788 unsigned long usable_startpfn;
6789 unsigned long kernelcore_node, kernelcore_remaining;
6790 /* save the state before borrow the nodemask */
6791 nodemask_t saved_node_state = node_states[N_MEMORY];
6792 unsigned long totalpages = early_calculate_totalpages();
6793 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6794 struct memblock_region *r;
6796 /* Need to find movable_zone earlier when movable_node is specified. */
6797 find_usable_zone_for_movable();
6800 * If movable_node is specified, ignore kernelcore and movablecore
6803 if (movable_node_is_enabled()) {
6804 for_each_memblock(memory, r) {
6805 if (!memblock_is_hotpluggable(r))
6810 usable_startpfn = PFN_DOWN(r->base);
6811 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6812 min(usable_startpfn, zone_movable_pfn[nid]) :
6820 * If kernelcore=mirror is specified, ignore movablecore option
6822 if (mirrored_kernelcore) {
6823 bool mem_below_4gb_not_mirrored = false;
6825 for_each_memblock(memory, r) {
6826 if (memblock_is_mirror(r))
6831 usable_startpfn = memblock_region_memory_base_pfn(r);
6833 if (usable_startpfn < 0x100000) {
6834 mem_below_4gb_not_mirrored = true;
6838 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6839 min(usable_startpfn, zone_movable_pfn[nid]) :
6843 if (mem_below_4gb_not_mirrored)
6844 pr_warn("This configuration results in unmirrored kernel memory.");
6850 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6851 * amount of necessary memory.
6853 if (required_kernelcore_percent)
6854 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6856 if (required_movablecore_percent)
6857 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6861 * If movablecore= was specified, calculate what size of
6862 * kernelcore that corresponds so that memory usable for
6863 * any allocation type is evenly spread. If both kernelcore
6864 * and movablecore are specified, then the value of kernelcore
6865 * will be used for required_kernelcore if it's greater than
6866 * what movablecore would have allowed.
6868 if (required_movablecore) {
6869 unsigned long corepages;
6872 * Round-up so that ZONE_MOVABLE is at least as large as what
6873 * was requested by the user
6875 required_movablecore =
6876 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6877 required_movablecore = min(totalpages, required_movablecore);
6878 corepages = totalpages - required_movablecore;
6880 required_kernelcore = max(required_kernelcore, corepages);
6884 * If kernelcore was not specified or kernelcore size is larger
6885 * than totalpages, there is no ZONE_MOVABLE.
6887 if (!required_kernelcore || required_kernelcore >= totalpages)
6890 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6891 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6894 /* Spread kernelcore memory as evenly as possible throughout nodes */
6895 kernelcore_node = required_kernelcore / usable_nodes;
6896 for_each_node_state(nid, N_MEMORY) {
6897 unsigned long start_pfn, end_pfn;
6900 * Recalculate kernelcore_node if the division per node
6901 * now exceeds what is necessary to satisfy the requested
6902 * amount of memory for the kernel
6904 if (required_kernelcore < kernelcore_node)
6905 kernelcore_node = required_kernelcore / usable_nodes;
6908 * As the map is walked, we track how much memory is usable
6909 * by the kernel using kernelcore_remaining. When it is
6910 * 0, the rest of the node is usable by ZONE_MOVABLE
6912 kernelcore_remaining = kernelcore_node;
6914 /* Go through each range of PFNs within this node */
6915 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6916 unsigned long size_pages;
6918 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6919 if (start_pfn >= end_pfn)
6922 /* Account for what is only usable for kernelcore */
6923 if (start_pfn < usable_startpfn) {
6924 unsigned long kernel_pages;
6925 kernel_pages = min(end_pfn, usable_startpfn)
6928 kernelcore_remaining -= min(kernel_pages,
6929 kernelcore_remaining);
6930 required_kernelcore -= min(kernel_pages,
6931 required_kernelcore);
6933 /* Continue if range is now fully accounted */
6934 if (end_pfn <= usable_startpfn) {
6937 * Push zone_movable_pfn to the end so
6938 * that if we have to rebalance
6939 * kernelcore across nodes, we will
6940 * not double account here
6942 zone_movable_pfn[nid] = end_pfn;
6945 start_pfn = usable_startpfn;
6949 * The usable PFN range for ZONE_MOVABLE is from
6950 * start_pfn->end_pfn. Calculate size_pages as the
6951 * number of pages used as kernelcore
6953 size_pages = end_pfn - start_pfn;
6954 if (size_pages > kernelcore_remaining)
6955 size_pages = kernelcore_remaining;
6956 zone_movable_pfn[nid] = start_pfn + size_pages;
6959 * Some kernelcore has been met, update counts and
6960 * break if the kernelcore for this node has been
6963 required_kernelcore -= min(required_kernelcore,
6965 kernelcore_remaining -= size_pages;
6966 if (!kernelcore_remaining)
6972 * If there is still required_kernelcore, we do another pass with one
6973 * less node in the count. This will push zone_movable_pfn[nid] further
6974 * along on the nodes that still have memory until kernelcore is
6978 if (usable_nodes && required_kernelcore > usable_nodes)
6982 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6983 for (nid = 0; nid < MAX_NUMNODES; nid++)
6984 zone_movable_pfn[nid] =
6985 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6988 /* restore the node_state */
6989 node_states[N_MEMORY] = saved_node_state;
6992 /* Any regular or high memory on that node ? */
6993 static void check_for_memory(pg_data_t *pgdat, int nid)
6995 enum zone_type zone_type;
6997 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6998 struct zone *zone = &pgdat->node_zones[zone_type];
6999 if (populated_zone(zone)) {
7000 if (IS_ENABLED(CONFIG_HIGHMEM))
7001 node_set_state(nid, N_HIGH_MEMORY);
7002 if (zone_type <= ZONE_NORMAL)
7003 node_set_state(nid, N_NORMAL_MEMORY);
7010 * free_area_init_nodes - Initialise all pg_data_t and zone data
7011 * @max_zone_pfn: an array of max PFNs for each zone
7013 * This will call free_area_init_node() for each active node in the system.
7014 * Using the page ranges provided by memblock_set_node(), the size of each
7015 * zone in each node and their holes is calculated. If the maximum PFN
7016 * between two adjacent zones match, it is assumed that the zone is empty.
7017 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7018 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7019 * starts where the previous one ended. For example, ZONE_DMA32 starts
7020 * at arch_max_dma_pfn.
7022 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7024 unsigned long start_pfn, end_pfn;
7027 /* Record where the zone boundaries are */
7028 memset(arch_zone_lowest_possible_pfn, 0,
7029 sizeof(arch_zone_lowest_possible_pfn));
7030 memset(arch_zone_highest_possible_pfn, 0,
7031 sizeof(arch_zone_highest_possible_pfn));
7033 start_pfn = find_min_pfn_with_active_regions();
7035 for (i = 0; i < MAX_NR_ZONES; i++) {
7036 if (i == ZONE_MOVABLE)
7039 end_pfn = max(max_zone_pfn[i], start_pfn);
7040 arch_zone_lowest_possible_pfn[i] = start_pfn;
7041 arch_zone_highest_possible_pfn[i] = end_pfn;
7043 start_pfn = end_pfn;
7046 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7047 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7048 find_zone_movable_pfns_for_nodes();
7050 /* Print out the zone ranges */
7051 pr_info("Zone ranges:\n");
7052 for (i = 0; i < MAX_NR_ZONES; i++) {
7053 if (i == ZONE_MOVABLE)
7055 pr_info(" %-8s ", zone_names[i]);
7056 if (arch_zone_lowest_possible_pfn[i] ==
7057 arch_zone_highest_possible_pfn[i])
7060 pr_cont("[mem %#018Lx-%#018Lx]\n",
7061 (u64)arch_zone_lowest_possible_pfn[i]
7063 ((u64)arch_zone_highest_possible_pfn[i]
7064 << PAGE_SHIFT) - 1);
7067 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7068 pr_info("Movable zone start for each node\n");
7069 for (i = 0; i < MAX_NUMNODES; i++) {
7070 if (zone_movable_pfn[i])
7071 pr_info(" Node %d: %#018Lx\n", i,
7072 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7075 /* Print out the early node map */
7076 pr_info("Early memory node ranges\n");
7077 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7078 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7079 (u64)start_pfn << PAGE_SHIFT,
7080 ((u64)end_pfn << PAGE_SHIFT) - 1);
7082 /* Initialise every node */
7083 mminit_verify_pageflags_layout();
7084 setup_nr_node_ids();
7085 zero_resv_unavail();
7086 for_each_online_node(nid) {
7087 pg_data_t *pgdat = NODE_DATA(nid);
7088 free_area_init_node(nid, NULL,
7089 find_min_pfn_for_node(nid), NULL);
7091 /* Any memory on that node */
7092 if (pgdat->node_present_pages)
7093 node_set_state(nid, N_MEMORY);
7094 check_for_memory(pgdat, nid);
7098 static int __init cmdline_parse_core(char *p, unsigned long *core,
7099 unsigned long *percent)
7101 unsigned long long coremem;
7107 /* Value may be a percentage of total memory, otherwise bytes */
7108 coremem = simple_strtoull(p, &endptr, 0);
7109 if (*endptr == '%') {
7110 /* Paranoid check for percent values greater than 100 */
7111 WARN_ON(coremem > 100);
7115 coremem = memparse(p, &p);
7116 /* Paranoid check that UL is enough for the coremem value */
7117 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7119 *core = coremem >> PAGE_SHIFT;
7126 * kernelcore=size sets the amount of memory for use for allocations that
7127 * cannot be reclaimed or migrated.
7129 static int __init cmdline_parse_kernelcore(char *p)
7131 /* parse kernelcore=mirror */
7132 if (parse_option_str(p, "mirror")) {
7133 mirrored_kernelcore = true;
7137 return cmdline_parse_core(p, &required_kernelcore,
7138 &required_kernelcore_percent);
7142 * movablecore=size sets the amount of memory for use for allocations that
7143 * can be reclaimed or migrated.
7145 static int __init cmdline_parse_movablecore(char *p)
7147 return cmdline_parse_core(p, &required_movablecore,
7148 &required_movablecore_percent);
7151 early_param("kernelcore", cmdline_parse_kernelcore);
7152 early_param("movablecore", cmdline_parse_movablecore);
7154 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7156 void adjust_managed_page_count(struct page *page, long count)
7158 atomic_long_add(count, &page_zone(page)->managed_pages);
7159 totalram_pages_add(count);
7160 #ifdef CONFIG_HIGHMEM
7161 if (PageHighMem(page))
7162 totalhigh_pages_add(count);
7165 EXPORT_SYMBOL(adjust_managed_page_count);
7167 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
7170 unsigned long pages = 0;
7172 start = (void *)PAGE_ALIGN((unsigned long)start);
7173 end = (void *)((unsigned long)end & PAGE_MASK);
7174 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7175 struct page *page = virt_to_page(pos);
7176 void *direct_map_addr;
7179 * 'direct_map_addr' might be different from 'pos'
7180 * because some architectures' virt_to_page()
7181 * work with aliases. Getting the direct map
7182 * address ensures that we get a _writeable_
7183 * alias for the memset().
7185 direct_map_addr = page_address(page);
7186 if ((unsigned int)poison <= 0xFF)
7187 memset(direct_map_addr, poison, PAGE_SIZE);
7189 free_reserved_page(page);
7193 pr_info("Freeing %s memory: %ldK\n",
7194 s, pages << (PAGE_SHIFT - 10));
7198 EXPORT_SYMBOL(free_reserved_area);
7200 #ifdef CONFIG_HIGHMEM
7201 void free_highmem_page(struct page *page)
7203 __free_reserved_page(page);
7204 totalram_pages_inc();
7205 atomic_long_inc(&page_zone(page)->managed_pages);
7206 totalhigh_pages_inc();
7211 void __init mem_init_print_info(const char *str)
7213 unsigned long physpages, codesize, datasize, rosize, bss_size;
7214 unsigned long init_code_size, init_data_size;
7216 physpages = get_num_physpages();
7217 codesize = _etext - _stext;
7218 datasize = _edata - _sdata;
7219 rosize = __end_rodata - __start_rodata;
7220 bss_size = __bss_stop - __bss_start;
7221 init_data_size = __init_end - __init_begin;
7222 init_code_size = _einittext - _sinittext;
7225 * Detect special cases and adjust section sizes accordingly:
7226 * 1) .init.* may be embedded into .data sections
7227 * 2) .init.text.* may be out of [__init_begin, __init_end],
7228 * please refer to arch/tile/kernel/vmlinux.lds.S.
7229 * 3) .rodata.* may be embedded into .text or .data sections.
7231 #define adj_init_size(start, end, size, pos, adj) \
7233 if (start <= pos && pos < end && size > adj) \
7237 adj_init_size(__init_begin, __init_end, init_data_size,
7238 _sinittext, init_code_size);
7239 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7240 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7241 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7242 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7244 #undef adj_init_size
7246 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7247 #ifdef CONFIG_HIGHMEM
7251 nr_free_pages() << (PAGE_SHIFT - 10),
7252 physpages << (PAGE_SHIFT - 10),
7253 codesize >> 10, datasize >> 10, rosize >> 10,
7254 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7255 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7256 totalcma_pages << (PAGE_SHIFT - 10),
7257 #ifdef CONFIG_HIGHMEM
7258 totalhigh_pages() << (PAGE_SHIFT - 10),
7260 str ? ", " : "", str ? str : "");
7264 * set_dma_reserve - set the specified number of pages reserved in the first zone
7265 * @new_dma_reserve: The number of pages to mark reserved
7267 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7268 * In the DMA zone, a significant percentage may be consumed by kernel image
7269 * and other unfreeable allocations which can skew the watermarks badly. This
7270 * function may optionally be used to account for unfreeable pages in the
7271 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7272 * smaller per-cpu batchsize.
7274 void __init set_dma_reserve(unsigned long new_dma_reserve)
7276 dma_reserve = new_dma_reserve;
7279 void __init free_area_init(unsigned long *zones_size)
7281 zero_resv_unavail();
7282 free_area_init_node(0, zones_size,
7283 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7286 static int page_alloc_cpu_dead(unsigned int cpu)
7289 lru_add_drain_cpu(cpu);
7293 * Spill the event counters of the dead processor
7294 * into the current processors event counters.
7295 * This artificially elevates the count of the current
7298 vm_events_fold_cpu(cpu);
7301 * Zero the differential counters of the dead processor
7302 * so that the vm statistics are consistent.
7304 * This is only okay since the processor is dead and cannot
7305 * race with what we are doing.
7307 cpu_vm_stats_fold(cpu);
7311 void __init page_alloc_init(void)
7315 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7316 "mm/page_alloc:dead", NULL,
7317 page_alloc_cpu_dead);
7322 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7323 * or min_free_kbytes changes.
7325 static void calculate_totalreserve_pages(void)
7327 struct pglist_data *pgdat;
7328 unsigned long reserve_pages = 0;
7329 enum zone_type i, j;
7331 for_each_online_pgdat(pgdat) {
7333 pgdat->totalreserve_pages = 0;
7335 for (i = 0; i < MAX_NR_ZONES; i++) {
7336 struct zone *zone = pgdat->node_zones + i;
7338 unsigned long managed_pages = zone_managed_pages(zone);
7340 /* Find valid and maximum lowmem_reserve in the zone */
7341 for (j = i; j < MAX_NR_ZONES; j++) {
7342 if (zone->lowmem_reserve[j] > max)
7343 max = zone->lowmem_reserve[j];
7346 /* we treat the high watermark as reserved pages. */
7347 max += high_wmark_pages(zone);
7349 if (max > managed_pages)
7350 max = managed_pages;
7352 pgdat->totalreserve_pages += max;
7354 reserve_pages += max;
7357 totalreserve_pages = reserve_pages;
7361 * setup_per_zone_lowmem_reserve - called whenever
7362 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7363 * has a correct pages reserved value, so an adequate number of
7364 * pages are left in the zone after a successful __alloc_pages().
7366 static void setup_per_zone_lowmem_reserve(void)
7368 struct pglist_data *pgdat;
7369 enum zone_type j, idx;
7371 for_each_online_pgdat(pgdat) {
7372 for (j = 0; j < MAX_NR_ZONES; j++) {
7373 struct zone *zone = pgdat->node_zones + j;
7374 unsigned long managed_pages = zone_managed_pages(zone);
7376 zone->lowmem_reserve[j] = 0;
7380 struct zone *lower_zone;
7383 lower_zone = pgdat->node_zones + idx;
7385 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7386 sysctl_lowmem_reserve_ratio[idx] = 0;
7387 lower_zone->lowmem_reserve[j] = 0;
7389 lower_zone->lowmem_reserve[j] =
7390 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7392 managed_pages += zone_managed_pages(lower_zone);
7397 /* update totalreserve_pages */
7398 calculate_totalreserve_pages();
7401 static void __setup_per_zone_wmarks(void)
7403 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7404 unsigned long lowmem_pages = 0;
7406 unsigned long flags;
7408 /* Calculate total number of !ZONE_HIGHMEM pages */
7409 for_each_zone(zone) {
7410 if (!is_highmem(zone))
7411 lowmem_pages += zone_managed_pages(zone);
7414 for_each_zone(zone) {
7417 spin_lock_irqsave(&zone->lock, flags);
7418 tmp = (u64)pages_min * zone_managed_pages(zone);
7419 do_div(tmp, lowmem_pages);
7420 if (is_highmem(zone)) {
7422 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7423 * need highmem pages, so cap pages_min to a small
7426 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7427 * deltas control asynch page reclaim, and so should
7428 * not be capped for highmem.
7430 unsigned long min_pages;
7432 min_pages = zone_managed_pages(zone) / 1024;
7433 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7434 zone->_watermark[WMARK_MIN] = min_pages;
7437 * If it's a lowmem zone, reserve a number of pages
7438 * proportionate to the zone's size.
7440 zone->_watermark[WMARK_MIN] = tmp;
7444 * Set the kswapd watermarks distance according to the
7445 * scale factor in proportion to available memory, but
7446 * ensure a minimum size on small systems.
7448 tmp = max_t(u64, tmp >> 2,
7449 mult_frac(zone_managed_pages(zone),
7450 watermark_scale_factor, 10000));
7452 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7453 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7455 spin_unlock_irqrestore(&zone->lock, flags);
7458 /* update totalreserve_pages */
7459 calculate_totalreserve_pages();
7463 * setup_per_zone_wmarks - called when min_free_kbytes changes
7464 * or when memory is hot-{added|removed}
7466 * Ensures that the watermark[min,low,high] values for each zone are set
7467 * correctly with respect to min_free_kbytes.
7469 void setup_per_zone_wmarks(void)
7471 static DEFINE_SPINLOCK(lock);
7474 __setup_per_zone_wmarks();
7479 * Initialise min_free_kbytes.
7481 * For small machines we want it small (128k min). For large machines
7482 * we want it large (64MB max). But it is not linear, because network
7483 * bandwidth does not increase linearly with machine size. We use
7485 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7486 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7502 int __meminit init_per_zone_wmark_min(void)
7504 unsigned long lowmem_kbytes;
7505 int new_min_free_kbytes;
7507 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7508 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7510 if (new_min_free_kbytes > user_min_free_kbytes) {
7511 min_free_kbytes = new_min_free_kbytes;
7512 if (min_free_kbytes < 128)
7513 min_free_kbytes = 128;
7514 if (min_free_kbytes > 65536)
7515 min_free_kbytes = 65536;
7517 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7518 new_min_free_kbytes, user_min_free_kbytes);
7520 setup_per_zone_wmarks();
7521 refresh_zone_stat_thresholds();
7522 setup_per_zone_lowmem_reserve();
7525 setup_min_unmapped_ratio();
7526 setup_min_slab_ratio();
7531 core_initcall(init_per_zone_wmark_min)
7534 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7535 * that we can call two helper functions whenever min_free_kbytes
7538 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7539 void __user *buffer, size_t *length, loff_t *ppos)
7543 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7548 user_min_free_kbytes = min_free_kbytes;
7549 setup_per_zone_wmarks();
7554 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7555 void __user *buffer, size_t *length, loff_t *ppos)
7559 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7564 setup_per_zone_wmarks();
7570 static void setup_min_unmapped_ratio(void)
7575 for_each_online_pgdat(pgdat)
7576 pgdat->min_unmapped_pages = 0;
7579 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7580 sysctl_min_unmapped_ratio) / 100;
7584 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7585 void __user *buffer, size_t *length, loff_t *ppos)
7589 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7593 setup_min_unmapped_ratio();
7598 static void setup_min_slab_ratio(void)
7603 for_each_online_pgdat(pgdat)
7604 pgdat->min_slab_pages = 0;
7607 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7608 sysctl_min_slab_ratio) / 100;
7611 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7612 void __user *buffer, size_t *length, loff_t *ppos)
7616 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7620 setup_min_slab_ratio();
7627 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7628 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7629 * whenever sysctl_lowmem_reserve_ratio changes.
7631 * The reserve ratio obviously has absolutely no relation with the
7632 * minimum watermarks. The lowmem reserve ratio can only make sense
7633 * if in function of the boot time zone sizes.
7635 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7636 void __user *buffer, size_t *length, loff_t *ppos)
7638 proc_dointvec_minmax(table, write, buffer, length, ppos);
7639 setup_per_zone_lowmem_reserve();
7644 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7645 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7646 * pagelist can have before it gets flushed back to buddy allocator.
7648 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7649 void __user *buffer, size_t *length, loff_t *ppos)
7652 int old_percpu_pagelist_fraction;
7655 mutex_lock(&pcp_batch_high_lock);
7656 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7658 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7659 if (!write || ret < 0)
7662 /* Sanity checking to avoid pcp imbalance */
7663 if (percpu_pagelist_fraction &&
7664 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7665 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7671 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7674 for_each_populated_zone(zone) {
7677 for_each_possible_cpu(cpu)
7678 pageset_set_high_and_batch(zone,
7679 per_cpu_ptr(zone->pageset, cpu));
7682 mutex_unlock(&pcp_batch_high_lock);
7687 int hashdist = HASHDIST_DEFAULT;
7689 static int __init set_hashdist(char *str)
7693 hashdist = simple_strtoul(str, &str, 0);
7696 __setup("hashdist=", set_hashdist);
7699 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7701 * Returns the number of pages that arch has reserved but
7702 * is not known to alloc_large_system_hash().
7704 static unsigned long __init arch_reserved_kernel_pages(void)
7711 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7712 * machines. As memory size is increased the scale is also increased but at
7713 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7714 * quadruples the scale is increased by one, which means the size of hash table
7715 * only doubles, instead of quadrupling as well.
7716 * Because 32-bit systems cannot have large physical memory, where this scaling
7717 * makes sense, it is disabled on such platforms.
7719 #if __BITS_PER_LONG > 32
7720 #define ADAPT_SCALE_BASE (64ul << 30)
7721 #define ADAPT_SCALE_SHIFT 2
7722 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7726 * allocate a large system hash table from bootmem
7727 * - it is assumed that the hash table must contain an exact power-of-2
7728 * quantity of entries
7729 * - limit is the number of hash buckets, not the total allocation size
7731 void *__init alloc_large_system_hash(const char *tablename,
7732 unsigned long bucketsize,
7733 unsigned long numentries,
7736 unsigned int *_hash_shift,
7737 unsigned int *_hash_mask,
7738 unsigned long low_limit,
7739 unsigned long high_limit)
7741 unsigned long long max = high_limit;
7742 unsigned long log2qty, size;
7746 /* allow the kernel cmdline to have a say */
7748 /* round applicable memory size up to nearest megabyte */
7749 numentries = nr_kernel_pages;
7750 numentries -= arch_reserved_kernel_pages();
7752 /* It isn't necessary when PAGE_SIZE >= 1MB */
7753 if (PAGE_SHIFT < 20)
7754 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7756 #if __BITS_PER_LONG > 32
7758 unsigned long adapt;
7760 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7761 adapt <<= ADAPT_SCALE_SHIFT)
7766 /* limit to 1 bucket per 2^scale bytes of low memory */
7767 if (scale > PAGE_SHIFT)
7768 numentries >>= (scale - PAGE_SHIFT);
7770 numentries <<= (PAGE_SHIFT - scale);
7772 /* Make sure we've got at least a 0-order allocation.. */
7773 if (unlikely(flags & HASH_SMALL)) {
7774 /* Makes no sense without HASH_EARLY */
7775 WARN_ON(!(flags & HASH_EARLY));
7776 if (!(numentries >> *_hash_shift)) {
7777 numentries = 1UL << *_hash_shift;
7778 BUG_ON(!numentries);
7780 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7781 numentries = PAGE_SIZE / bucketsize;
7783 numentries = roundup_pow_of_two(numentries);
7785 /* limit allocation size to 1/16 total memory by default */
7787 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7788 do_div(max, bucketsize);
7790 max = min(max, 0x80000000ULL);
7792 if (numentries < low_limit)
7793 numentries = low_limit;
7794 if (numentries > max)
7797 log2qty = ilog2(numentries);
7799 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7801 size = bucketsize << log2qty;
7802 if (flags & HASH_EARLY) {
7803 if (flags & HASH_ZERO)
7804 table = memblock_alloc_nopanic(size,
7807 table = memblock_alloc_raw(size,
7809 } else if (hashdist) {
7810 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7813 * If bucketsize is not a power-of-two, we may free
7814 * some pages at the end of hash table which
7815 * alloc_pages_exact() automatically does
7817 if (get_order(size) < MAX_ORDER) {
7818 table = alloc_pages_exact(size, gfp_flags);
7819 kmemleak_alloc(table, size, 1, gfp_flags);
7822 } while (!table && size > PAGE_SIZE && --log2qty);
7825 panic("Failed to allocate %s hash table\n", tablename);
7827 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7828 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7831 *_hash_shift = log2qty;
7833 *_hash_mask = (1 << log2qty) - 1;
7839 * This function checks whether pageblock includes unmovable pages or not.
7840 * If @count is not zero, it is okay to include less @count unmovable pages
7842 * PageLRU check without isolation or lru_lock could race so that
7843 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7844 * check without lock_page also may miss some movable non-lru pages at
7845 * race condition. So you can't expect this function should be exact.
7847 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7848 int migratetype, int flags)
7850 unsigned long pfn, iter, found;
7853 * TODO we could make this much more efficient by not checking every
7854 * page in the range if we know all of them are in MOVABLE_ZONE and
7855 * that the movable zone guarantees that pages are migratable but
7856 * the later is not the case right now unfortunatelly. E.g. movablecore
7857 * can still lead to having bootmem allocations in zone_movable.
7861 * CMA allocations (alloc_contig_range) really need to mark isolate
7862 * CMA pageblocks even when they are not movable in fact so consider
7863 * them movable here.
7865 if (is_migrate_cma(migratetype) &&
7866 is_migrate_cma(get_pageblock_migratetype(page)))
7869 pfn = page_to_pfn(page);
7870 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7871 unsigned long check = pfn + iter;
7873 if (!pfn_valid_within(check))
7876 page = pfn_to_page(check);
7878 if (PageReserved(page))
7882 * If the zone is movable and we have ruled out all reserved
7883 * pages then it should be reasonably safe to assume the rest
7886 if (zone_idx(zone) == ZONE_MOVABLE)
7890 * Hugepages are not in LRU lists, but they're movable.
7891 * We need not scan over tail pages bacause we don't
7892 * handle each tail page individually in migration.
7894 if (PageHuge(page)) {
7895 struct page *head = compound_head(page);
7896 unsigned int skip_pages;
7898 if (!hugepage_migration_supported(page_hstate(head)))
7901 skip_pages = (1 << compound_order(head)) - (page - head);
7902 iter += skip_pages - 1;
7907 * We can't use page_count without pin a page
7908 * because another CPU can free compound page.
7909 * This check already skips compound tails of THP
7910 * because their page->_refcount is zero at all time.
7912 if (!page_ref_count(page)) {
7913 if (PageBuddy(page))
7914 iter += (1 << page_order(page)) - 1;
7919 * The HWPoisoned page may be not in buddy system, and
7920 * page_count() is not 0.
7922 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
7925 if (__PageMovable(page))
7931 * If there are RECLAIMABLE pages, we need to check
7932 * it. But now, memory offline itself doesn't call
7933 * shrink_node_slabs() and it still to be fixed.
7936 * If the page is not RAM, page_count()should be 0.
7937 * we don't need more check. This is an _used_ not-movable page.
7939 * The problematic thing here is PG_reserved pages. PG_reserved
7940 * is set to both of a memory hole page and a _used_ kernel
7948 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7949 if (flags & REPORT_FAILURE)
7950 dump_page(pfn_to_page(pfn+iter), "unmovable page");
7954 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7956 static unsigned long pfn_max_align_down(unsigned long pfn)
7958 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7959 pageblock_nr_pages) - 1);
7962 static unsigned long pfn_max_align_up(unsigned long pfn)
7964 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7965 pageblock_nr_pages));
7968 /* [start, end) must belong to a single zone. */
7969 static int __alloc_contig_migrate_range(struct compact_control *cc,
7970 unsigned long start, unsigned long end)
7972 /* This function is based on compact_zone() from compaction.c. */
7973 unsigned long nr_reclaimed;
7974 unsigned long pfn = start;
7975 unsigned int tries = 0;
7980 while (pfn < end || !list_empty(&cc->migratepages)) {
7981 if (fatal_signal_pending(current)) {
7986 if (list_empty(&cc->migratepages)) {
7987 cc->nr_migratepages = 0;
7988 pfn = isolate_migratepages_range(cc, pfn, end);
7994 } else if (++tries == 5) {
7995 ret = ret < 0 ? ret : -EBUSY;
7999 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8001 cc->nr_migratepages -= nr_reclaimed;
8003 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8004 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8007 putback_movable_pages(&cc->migratepages);
8014 * alloc_contig_range() -- tries to allocate given range of pages
8015 * @start: start PFN to allocate
8016 * @end: one-past-the-last PFN to allocate
8017 * @migratetype: migratetype of the underlaying pageblocks (either
8018 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8019 * in range must have the same migratetype and it must
8020 * be either of the two.
8021 * @gfp_mask: GFP mask to use during compaction
8023 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8024 * aligned. The PFN range must belong to a single zone.
8026 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8027 * pageblocks in the range. Once isolated, the pageblocks should not
8028 * be modified by others.
8030 * Returns zero on success or negative error code. On success all
8031 * pages which PFN is in [start, end) are allocated for the caller and
8032 * need to be freed with free_contig_range().
8034 int alloc_contig_range(unsigned long start, unsigned long end,
8035 unsigned migratetype, gfp_t gfp_mask)
8037 unsigned long outer_start, outer_end;
8041 struct compact_control cc = {
8042 .nr_migratepages = 0,
8044 .zone = page_zone(pfn_to_page(start)),
8045 .mode = MIGRATE_SYNC,
8046 .ignore_skip_hint = true,
8047 .no_set_skip_hint = true,
8048 .gfp_mask = current_gfp_context(gfp_mask),
8050 INIT_LIST_HEAD(&cc.migratepages);
8053 * What we do here is we mark all pageblocks in range as
8054 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8055 * have different sizes, and due to the way page allocator
8056 * work, we align the range to biggest of the two pages so
8057 * that page allocator won't try to merge buddies from
8058 * different pageblocks and change MIGRATE_ISOLATE to some
8059 * other migration type.
8061 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8062 * migrate the pages from an unaligned range (ie. pages that
8063 * we are interested in). This will put all the pages in
8064 * range back to page allocator as MIGRATE_ISOLATE.
8066 * When this is done, we take the pages in range from page
8067 * allocator removing them from the buddy system. This way
8068 * page allocator will never consider using them.
8070 * This lets us mark the pageblocks back as
8071 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8072 * aligned range but not in the unaligned, original range are
8073 * put back to page allocator so that buddy can use them.
8076 ret = start_isolate_page_range(pfn_max_align_down(start),
8077 pfn_max_align_up(end), migratetype, 0);
8082 * In case of -EBUSY, we'd like to know which page causes problem.
8083 * So, just fall through. test_pages_isolated() has a tracepoint
8084 * which will report the busy page.
8086 * It is possible that busy pages could become available before
8087 * the call to test_pages_isolated, and the range will actually be
8088 * allocated. So, if we fall through be sure to clear ret so that
8089 * -EBUSY is not accidentally used or returned to caller.
8091 ret = __alloc_contig_migrate_range(&cc, start, end);
8092 if (ret && ret != -EBUSY)
8097 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8098 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8099 * more, all pages in [start, end) are free in page allocator.
8100 * What we are going to do is to allocate all pages from
8101 * [start, end) (that is remove them from page allocator).
8103 * The only problem is that pages at the beginning and at the
8104 * end of interesting range may be not aligned with pages that
8105 * page allocator holds, ie. they can be part of higher order
8106 * pages. Because of this, we reserve the bigger range and
8107 * once this is done free the pages we are not interested in.
8109 * We don't have to hold zone->lock here because the pages are
8110 * isolated thus they won't get removed from buddy.
8113 lru_add_drain_all();
8114 drain_all_pages(cc.zone);
8117 outer_start = start;
8118 while (!PageBuddy(pfn_to_page(outer_start))) {
8119 if (++order >= MAX_ORDER) {
8120 outer_start = start;
8123 outer_start &= ~0UL << order;
8126 if (outer_start != start) {
8127 order = page_order(pfn_to_page(outer_start));
8130 * outer_start page could be small order buddy page and
8131 * it doesn't include start page. Adjust outer_start
8132 * in this case to report failed page properly
8133 * on tracepoint in test_pages_isolated()
8135 if (outer_start + (1UL << order) <= start)
8136 outer_start = start;
8139 /* Make sure the range is really isolated. */
8140 if (test_pages_isolated(outer_start, end, false)) {
8141 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8142 __func__, outer_start, end);
8147 /* Grab isolated pages from freelists. */
8148 outer_end = isolate_freepages_range(&cc, outer_start, end);
8154 /* Free head and tail (if any) */
8155 if (start != outer_start)
8156 free_contig_range(outer_start, start - outer_start);
8157 if (end != outer_end)
8158 free_contig_range(end, outer_end - end);
8161 undo_isolate_page_range(pfn_max_align_down(start),
8162 pfn_max_align_up(end), migratetype);
8166 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8168 unsigned int count = 0;
8170 for (; nr_pages--; pfn++) {
8171 struct page *page = pfn_to_page(pfn);
8173 count += page_count(page) != 1;
8176 WARN(count != 0, "%d pages are still in use!\n", count);
8180 #ifdef CONFIG_MEMORY_HOTPLUG
8182 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8183 * page high values need to be recalulated.
8185 void __meminit zone_pcp_update(struct zone *zone)
8188 mutex_lock(&pcp_batch_high_lock);
8189 for_each_possible_cpu(cpu)
8190 pageset_set_high_and_batch(zone,
8191 per_cpu_ptr(zone->pageset, cpu));
8192 mutex_unlock(&pcp_batch_high_lock);
8196 void zone_pcp_reset(struct zone *zone)
8198 unsigned long flags;
8200 struct per_cpu_pageset *pset;
8202 /* avoid races with drain_pages() */
8203 local_irq_save(flags);
8204 if (zone->pageset != &boot_pageset) {
8205 for_each_online_cpu(cpu) {
8206 pset = per_cpu_ptr(zone->pageset, cpu);
8207 drain_zonestat(zone, pset);
8209 free_percpu(zone->pageset);
8210 zone->pageset = &boot_pageset;
8212 local_irq_restore(flags);
8215 #ifdef CONFIG_MEMORY_HOTREMOVE
8217 * All pages in the range must be in a single zone and isolated
8218 * before calling this.
8221 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8225 unsigned int order, i;
8227 unsigned long flags;
8228 /* find the first valid pfn */
8229 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8234 offline_mem_sections(pfn, end_pfn);
8235 zone = page_zone(pfn_to_page(pfn));
8236 spin_lock_irqsave(&zone->lock, flags);
8238 while (pfn < end_pfn) {
8239 if (!pfn_valid(pfn)) {
8243 page = pfn_to_page(pfn);
8245 * The HWPoisoned page may be not in buddy system, and
8246 * page_count() is not 0.
8248 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8250 SetPageReserved(page);
8254 BUG_ON(page_count(page));
8255 BUG_ON(!PageBuddy(page));
8256 order = page_order(page);
8257 #ifdef CONFIG_DEBUG_VM
8258 pr_info("remove from free list %lx %d %lx\n",
8259 pfn, 1 << order, end_pfn);
8261 list_del(&page->lru);
8262 rmv_page_order(page);
8263 zone->free_area[order].nr_free--;
8264 for (i = 0; i < (1 << order); i++)
8265 SetPageReserved((page+i));
8266 pfn += (1 << order);
8268 spin_unlock_irqrestore(&zone->lock, flags);
8272 bool is_free_buddy_page(struct page *page)
8274 struct zone *zone = page_zone(page);
8275 unsigned long pfn = page_to_pfn(page);
8276 unsigned long flags;
8279 spin_lock_irqsave(&zone->lock, flags);
8280 for (order = 0; order < MAX_ORDER; order++) {
8281 struct page *page_head = page - (pfn & ((1 << order) - 1));
8283 if (PageBuddy(page_head) && page_order(page_head) >= order)
8286 spin_unlock_irqrestore(&zone->lock, flags);
8288 return order < MAX_ORDER;
8291 #ifdef CONFIG_MEMORY_FAILURE
8293 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8294 * test is performed under the zone lock to prevent a race against page
8297 bool set_hwpoison_free_buddy_page(struct page *page)
8299 struct zone *zone = page_zone(page);
8300 unsigned long pfn = page_to_pfn(page);
8301 unsigned long flags;
8303 bool hwpoisoned = false;
8305 spin_lock_irqsave(&zone->lock, flags);
8306 for (order = 0; order < MAX_ORDER; order++) {
8307 struct page *page_head = page - (pfn & ((1 << order) - 1));
8309 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8310 if (!TestSetPageHWPoison(page))
8315 spin_unlock_irqrestore(&zone->lock, flags);