2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
68 #include <asm/sections.h>
69 #include <asm/tlbflush.h>
70 #include <asm/div64.h>
73 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
74 static DEFINE_MUTEX(pcp_batch_high_lock);
75 #define MIN_PERCPU_PAGELIST_FRACTION (8)
77 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
78 DEFINE_PER_CPU(int, numa_node);
79 EXPORT_PER_CPU_SYMBOL(numa_node);
82 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
84 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
85 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
86 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
87 * defined in <linux/topology.h>.
89 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
90 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
91 int _node_numa_mem_[MAX_NUMNODES];
94 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
95 volatile unsigned long latent_entropy __latent_entropy;
96 EXPORT_SYMBOL(latent_entropy);
100 * Array of node states.
102 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
103 [N_POSSIBLE] = NODE_MASK_ALL,
104 [N_ONLINE] = { { [0] = 1UL } },
106 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
107 #ifdef CONFIG_HIGHMEM
108 [N_HIGH_MEMORY] = { { [0] = 1UL } },
110 #ifdef CONFIG_MOVABLE_NODE
111 [N_MEMORY] = { { [0] = 1UL } },
113 [N_CPU] = { { [0] = 1UL } },
116 EXPORT_SYMBOL(node_states);
118 /* Protect totalram_pages and zone->managed_pages */
119 static DEFINE_SPINLOCK(managed_page_count_lock);
121 unsigned long totalram_pages __read_mostly;
122 unsigned long totalreserve_pages __read_mostly;
123 unsigned long totalcma_pages __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 char * const migratetype_names[MIGRATE_TYPES] = {
240 #ifdef CONFIG_MEMORY_ISOLATION
245 compound_page_dtor * const compound_page_dtors[] = {
248 #ifdef CONFIG_HUGETLB_PAGE
251 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
256 int min_free_kbytes = 1024;
257 int user_min_free_kbytes = -1;
258 int watermark_scale_factor = 10;
260 static unsigned long __meminitdata nr_kernel_pages;
261 static unsigned long __meminitdata nr_all_pages;
262 static unsigned long __meminitdata dma_reserve;
264 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
265 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
266 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
267 static unsigned long __initdata required_kernelcore;
268 static unsigned long __initdata required_movablecore;
269 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
270 static bool mirrored_kernelcore;
272 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
274 EXPORT_SYMBOL(movable_zone);
275 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
278 int nr_node_ids __read_mostly = MAX_NUMNODES;
279 int nr_online_nodes __read_mostly = 1;
280 EXPORT_SYMBOL(nr_node_ids);
281 EXPORT_SYMBOL(nr_online_nodes);
284 int page_group_by_mobility_disabled __read_mostly;
286 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
287 static inline void reset_deferred_meminit(pg_data_t *pgdat)
289 pgdat->first_deferred_pfn = ULONG_MAX;
292 /* Returns true if the struct page for the pfn is uninitialised */
293 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
295 int nid = early_pfn_to_nid(pfn);
297 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
307 static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
311 unsigned long max_initialise;
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
333 static inline void reset_deferred_meminit(pg_data_t *pgdat)
337 static inline bool early_page_uninitialised(unsigned long pfn)
342 static inline bool update_defer_init(pg_data_t *pgdat,
343 unsigned long pfn, unsigned long zone_end,
344 unsigned long *nr_initialised)
350 /* Return a pointer to the bitmap storing bits affecting a block of pages */
351 static inline unsigned long *get_pageblock_bitmap(struct page *page,
354 #ifdef CONFIG_SPARSEMEM
355 return __pfn_to_section(pfn)->pageblock_flags;
357 return page_zone(page)->pageblock_flags;
358 #endif /* CONFIG_SPARSEMEM */
361 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
363 #ifdef CONFIG_SPARSEMEM
364 pfn &= (PAGES_PER_SECTION-1);
365 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
367 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
368 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
369 #endif /* CONFIG_SPARSEMEM */
373 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
374 * @page: The page within the block of interest
375 * @pfn: The target page frame number
376 * @end_bitidx: The last bit of interest to retrieve
377 * @mask: mask of bits that the caller is interested in
379 * Return: pageblock_bits flags
381 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
383 unsigned long end_bitidx,
386 unsigned long *bitmap;
387 unsigned long bitidx, word_bitidx;
390 bitmap = get_pageblock_bitmap(page, pfn);
391 bitidx = pfn_to_bitidx(page, pfn);
392 word_bitidx = bitidx / BITS_PER_LONG;
393 bitidx &= (BITS_PER_LONG-1);
395 word = bitmap[word_bitidx];
396 bitidx += end_bitidx;
397 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
400 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
401 unsigned long end_bitidx,
404 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
407 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
409 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
413 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
414 * @page: The page within the block of interest
415 * @flags: The flags to set
416 * @pfn: The target page frame number
417 * @end_bitidx: The last bit of interest
418 * @mask: mask of bits that the caller is interested in
420 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
422 unsigned long end_bitidx,
425 unsigned long *bitmap;
426 unsigned long bitidx, word_bitidx;
427 unsigned long old_word, word;
429 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
431 bitmap = get_pageblock_bitmap(page, pfn);
432 bitidx = pfn_to_bitidx(page, pfn);
433 word_bitidx = bitidx / BITS_PER_LONG;
434 bitidx &= (BITS_PER_LONG-1);
436 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
438 bitidx += end_bitidx;
439 mask <<= (BITS_PER_LONG - bitidx - 1);
440 flags <<= (BITS_PER_LONG - bitidx - 1);
442 word = READ_ONCE(bitmap[word_bitidx]);
444 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
445 if (word == old_word)
451 void set_pageblock_migratetype(struct page *page, int migratetype)
453 if (unlikely(page_group_by_mobility_disabled &&
454 migratetype < MIGRATE_PCPTYPES))
455 migratetype = MIGRATE_UNMOVABLE;
457 set_pageblock_flags_group(page, (unsigned long)migratetype,
458 PB_migrate, PB_migrate_end);
461 #ifdef CONFIG_DEBUG_VM
462 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
466 unsigned long pfn = page_to_pfn(page);
467 unsigned long sp, start_pfn;
470 seq = zone_span_seqbegin(zone);
471 start_pfn = zone->zone_start_pfn;
472 sp = zone->spanned_pages;
473 if (!zone_spans_pfn(zone, pfn))
475 } while (zone_span_seqretry(zone, seq));
478 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
479 pfn, zone_to_nid(zone), zone->name,
480 start_pfn, start_pfn + sp);
485 static int page_is_consistent(struct zone *zone, struct page *page)
487 if (!pfn_valid_within(page_to_pfn(page)))
489 if (zone != page_zone(page))
495 * Temporary debugging check for pages not lying within a given zone.
497 static int bad_range(struct zone *zone, struct page *page)
499 if (page_outside_zone_boundaries(zone, page))
501 if (!page_is_consistent(zone, page))
507 static inline int bad_range(struct zone *zone, struct page *page)
513 static void bad_page(struct page *page, const char *reason,
514 unsigned long bad_flags)
516 static unsigned long resume;
517 static unsigned long nr_shown;
518 static unsigned long nr_unshown;
521 * Allow a burst of 60 reports, then keep quiet for that minute;
522 * or allow a steady drip of one report per second.
524 if (nr_shown == 60) {
525 if (time_before(jiffies, resume)) {
531 "BUG: Bad page state: %lu messages suppressed\n",
538 resume = jiffies + 60 * HZ;
540 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
541 current->comm, page_to_pfn(page));
542 __dump_page(page, reason);
543 bad_flags &= page->flags;
545 pr_alert("bad because of flags: %#lx(%pGp)\n",
546 bad_flags, &bad_flags);
547 dump_page_owner(page);
552 /* Leave bad fields for debug, except PageBuddy could make trouble */
553 page_mapcount_reset(page); /* remove PageBuddy */
554 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
558 * Higher-order pages are called "compound pages". They are structured thusly:
560 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
562 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
563 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
565 * The first tail page's ->compound_dtor holds the offset in array of compound
566 * page destructors. See compound_page_dtors.
568 * The first tail page's ->compound_order holds the order of allocation.
569 * This usage means that zero-order pages may not be compound.
572 void free_compound_page(struct page *page)
574 __free_pages_ok(page, compound_order(page));
577 void prep_compound_page(struct page *page, unsigned int order)
580 int nr_pages = 1 << order;
582 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
583 set_compound_order(page, order);
585 for (i = 1; i < nr_pages; i++) {
586 struct page *p = page + i;
587 set_page_count(p, 0);
588 p->mapping = TAIL_MAPPING;
589 set_compound_head(p, page);
591 atomic_set(compound_mapcount_ptr(page), -1);
594 #ifdef CONFIG_DEBUG_PAGEALLOC
595 unsigned int _debug_guardpage_minorder;
596 bool _debug_pagealloc_enabled __read_mostly
597 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
598 EXPORT_SYMBOL(_debug_pagealloc_enabled);
599 bool _debug_guardpage_enabled __read_mostly;
601 static int __init early_debug_pagealloc(char *buf)
605 return kstrtobool(buf, &_debug_pagealloc_enabled);
607 early_param("debug_pagealloc", early_debug_pagealloc);
609 static bool need_debug_guardpage(void)
611 /* If we don't use debug_pagealloc, we don't need guard page */
612 if (!debug_pagealloc_enabled())
615 if (!debug_guardpage_minorder())
621 static void init_debug_guardpage(void)
623 if (!debug_pagealloc_enabled())
626 if (!debug_guardpage_minorder())
629 _debug_guardpage_enabled = true;
632 struct page_ext_operations debug_guardpage_ops = {
633 .need = need_debug_guardpage,
634 .init = init_debug_guardpage,
637 static int __init debug_guardpage_minorder_setup(char *buf)
641 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
642 pr_err("Bad debug_guardpage_minorder value\n");
645 _debug_guardpage_minorder = res;
646 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
649 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
651 static inline bool set_page_guard(struct zone *zone, struct page *page,
652 unsigned int order, int migratetype)
654 struct page_ext *page_ext;
656 if (!debug_guardpage_enabled())
659 if (order >= debug_guardpage_minorder())
662 page_ext = lookup_page_ext(page);
663 if (unlikely(!page_ext))
666 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
668 INIT_LIST_HEAD(&page->lru);
669 set_page_private(page, order);
670 /* Guard pages are not available for any usage */
671 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
676 static inline void clear_page_guard(struct zone *zone, struct page *page,
677 unsigned int order, int migratetype)
679 struct page_ext *page_ext;
681 if (!debug_guardpage_enabled())
684 page_ext = lookup_page_ext(page);
685 if (unlikely(!page_ext))
688 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
690 set_page_private(page, 0);
691 if (!is_migrate_isolate(migratetype))
692 __mod_zone_freepage_state(zone, (1 << order), migratetype);
695 struct page_ext_operations debug_guardpage_ops;
696 static inline bool set_page_guard(struct zone *zone, struct page *page,
697 unsigned int order, int migratetype) { return false; }
698 static inline void clear_page_guard(struct zone *zone, struct page *page,
699 unsigned int order, int migratetype) {}
702 static inline void set_page_order(struct page *page, unsigned int order)
704 set_page_private(page, order);
705 __SetPageBuddy(page);
708 static inline void rmv_page_order(struct page *page)
710 __ClearPageBuddy(page);
711 set_page_private(page, 0);
715 * This function checks whether a page is free && is the buddy
716 * we can do coalesce a page and its buddy if
717 * (a) the buddy is not in a hole &&
718 * (b) the buddy is in the buddy system &&
719 * (c) a page and its buddy have the same order &&
720 * (d) a page and its buddy are in the same zone.
722 * For recording whether a page is in the buddy system, we set ->_mapcount
723 * PAGE_BUDDY_MAPCOUNT_VALUE.
724 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
725 * serialized by zone->lock.
727 * For recording page's order, we use page_private(page).
729 static inline int page_is_buddy(struct page *page, struct page *buddy,
732 if (!pfn_valid_within(page_to_pfn(buddy)))
735 if (page_is_guard(buddy) && page_order(buddy) == order) {
736 if (page_zone_id(page) != page_zone_id(buddy))
739 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
744 if (PageBuddy(buddy) && page_order(buddy) == order) {
746 * zone check is done late to avoid uselessly
747 * calculating zone/node ids for pages that could
750 if (page_zone_id(page) != page_zone_id(buddy))
753 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
761 * Freeing function for a buddy system allocator.
763 * The concept of a buddy system is to maintain direct-mapped table
764 * (containing bit values) for memory blocks of various "orders".
765 * The bottom level table contains the map for the smallest allocatable
766 * units of memory (here, pages), and each level above it describes
767 * pairs of units from the levels below, hence, "buddies".
768 * At a high level, all that happens here is marking the table entry
769 * at the bottom level available, and propagating the changes upward
770 * as necessary, plus some accounting needed to play nicely with other
771 * parts of the VM system.
772 * At each level, we keep a list of pages, which are heads of continuous
773 * free pages of length of (1 << order) and marked with _mapcount
774 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
776 * So when we are allocating or freeing one, we can derive the state of the
777 * other. That is, if we allocate a small block, and both were
778 * free, the remainder of the region must be split into blocks.
779 * If a block is freed, and its buddy is also free, then this
780 * triggers coalescing into a block of larger size.
785 static inline void __free_one_page(struct page *page,
787 struct zone *zone, unsigned int order,
790 unsigned long combined_pfn;
791 unsigned long uninitialized_var(buddy_pfn);
793 unsigned int max_order;
795 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
797 VM_BUG_ON(!zone_is_initialized(zone));
798 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
800 VM_BUG_ON(migratetype == -1);
801 if (likely(!is_migrate_isolate(migratetype)))
802 __mod_zone_freepage_state(zone, 1 << order, migratetype);
804 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
805 VM_BUG_ON_PAGE(bad_range(zone, page), page);
808 while (order < max_order - 1) {
809 buddy_pfn = __find_buddy_pfn(pfn, order);
810 buddy = page + (buddy_pfn - pfn);
811 if (!page_is_buddy(page, buddy, order))
814 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
815 * merge with it and move up one order.
817 if (page_is_guard(buddy)) {
818 clear_page_guard(zone, buddy, order, migratetype);
820 list_del(&buddy->lru);
821 zone->free_area[order].nr_free--;
822 rmv_page_order(buddy);
824 combined_pfn = buddy_pfn & pfn;
825 page = page + (combined_pfn - pfn);
829 if (max_order < MAX_ORDER) {
830 /* If we are here, it means order is >= pageblock_order.
831 * We want to prevent merge between freepages on isolate
832 * pageblock and normal pageblock. Without this, pageblock
833 * isolation could cause incorrect freepage or CMA accounting.
835 * We don't want to hit this code for the more frequent
838 if (unlikely(has_isolate_pageblock(zone))) {
841 buddy_pfn = __find_buddy_pfn(pfn, order);
842 buddy = page + (buddy_pfn - pfn);
843 buddy_mt = get_pageblock_migratetype(buddy);
845 if (migratetype != buddy_mt
846 && (is_migrate_isolate(migratetype) ||
847 is_migrate_isolate(buddy_mt)))
851 goto continue_merging;
855 set_page_order(page, order);
858 * If this is not the largest possible page, check if the buddy
859 * of the next-highest order is free. If it is, it's possible
860 * that pages are being freed that will coalesce soon. In case,
861 * that is happening, add the free page to the tail of the list
862 * so it's less likely to be used soon and more likely to be merged
863 * as a higher order page
865 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
866 struct page *higher_page, *higher_buddy;
867 combined_pfn = buddy_pfn & pfn;
868 higher_page = page + (combined_pfn - pfn);
869 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
870 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
871 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
872 list_add_tail(&page->lru,
873 &zone->free_area[order].free_list[migratetype]);
878 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
880 zone->free_area[order].nr_free++;
884 * A bad page could be due to a number of fields. Instead of multiple branches,
885 * try and check multiple fields with one check. The caller must do a detailed
886 * check if necessary.
888 static inline bool page_expected_state(struct page *page,
889 unsigned long check_flags)
891 if (unlikely(atomic_read(&page->_mapcount) != -1))
894 if (unlikely((unsigned long)page->mapping |
895 page_ref_count(page) |
897 (unsigned long)page->mem_cgroup |
899 (page->flags & check_flags)))
905 static void free_pages_check_bad(struct page *page)
907 const char *bad_reason;
908 unsigned long bad_flags;
913 if (unlikely(atomic_read(&page->_mapcount) != -1))
914 bad_reason = "nonzero mapcount";
915 if (unlikely(page->mapping != NULL))
916 bad_reason = "non-NULL mapping";
917 if (unlikely(page_ref_count(page) != 0))
918 bad_reason = "nonzero _refcount";
919 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
920 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
921 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
924 if (unlikely(page->mem_cgroup))
925 bad_reason = "page still charged to cgroup";
927 bad_page(page, bad_reason, bad_flags);
930 static inline int free_pages_check(struct page *page)
932 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
935 /* Something has gone sideways, find it */
936 free_pages_check_bad(page);
940 static int free_tail_pages_check(struct page *head_page, struct page *page)
945 * We rely page->lru.next never has bit 0 set, unless the page
946 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
948 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
950 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
954 switch (page - head_page) {
956 /* the first tail page: ->mapping is compound_mapcount() */
957 if (unlikely(compound_mapcount(page))) {
958 bad_page(page, "nonzero compound_mapcount", 0);
964 * the second tail page: ->mapping is
965 * page_deferred_list().next -- ignore value.
969 if (page->mapping != TAIL_MAPPING) {
970 bad_page(page, "corrupted mapping in tail page", 0);
975 if (unlikely(!PageTail(page))) {
976 bad_page(page, "PageTail not set", 0);
979 if (unlikely(compound_head(page) != head_page)) {
980 bad_page(page, "compound_head not consistent", 0);
985 page->mapping = NULL;
986 clear_compound_head(page);
990 static __always_inline bool free_pages_prepare(struct page *page,
991 unsigned int order, bool check_free)
995 VM_BUG_ON_PAGE(PageTail(page), page);
997 trace_mm_page_free(page, order);
998 kmemcheck_free_shadow(page, order);
1001 * Check tail pages before head page information is cleared to
1002 * avoid checking PageCompound for order-0 pages.
1004 if (unlikely(order)) {
1005 bool compound = PageCompound(page);
1008 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1011 ClearPageDoubleMap(page);
1012 for (i = 1; i < (1 << order); i++) {
1014 bad += free_tail_pages_check(page, page + i);
1015 if (unlikely(free_pages_check(page + i))) {
1019 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1022 if (PageMappingFlags(page))
1023 page->mapping = NULL;
1024 if (memcg_kmem_enabled() && PageKmemcg(page))
1025 memcg_kmem_uncharge(page, order);
1027 bad += free_pages_check(page);
1031 page_cpupid_reset_last(page);
1032 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1033 reset_page_owner(page, order);
1035 if (!PageHighMem(page)) {
1036 debug_check_no_locks_freed(page_address(page),
1037 PAGE_SIZE << order);
1038 debug_check_no_obj_freed(page_address(page),
1039 PAGE_SIZE << order);
1041 arch_free_page(page, order);
1042 kernel_poison_pages(page, 1 << order, 0);
1043 kernel_map_pages(page, 1 << order, 0);
1044 kasan_free_pages(page, order);
1049 #ifdef CONFIG_DEBUG_VM
1050 static inline bool free_pcp_prepare(struct page *page)
1052 return free_pages_prepare(page, 0, true);
1055 static inline bool bulkfree_pcp_prepare(struct page *page)
1060 static bool free_pcp_prepare(struct page *page)
1062 return free_pages_prepare(page, 0, false);
1065 static bool bulkfree_pcp_prepare(struct page *page)
1067 return free_pages_check(page);
1069 #endif /* CONFIG_DEBUG_VM */
1072 * Frees a number of pages from the PCP lists
1073 * Assumes all pages on list are in same zone, and of same order.
1074 * count is the number of pages to free.
1076 * If the zone was previously in an "all pages pinned" state then look to
1077 * see if this freeing clears that state.
1079 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1080 * pinned" detection logic.
1082 static void free_pcppages_bulk(struct zone *zone, int count,
1083 struct per_cpu_pages *pcp)
1085 int migratetype = 0;
1087 unsigned long nr_scanned;
1088 bool isolated_pageblocks;
1090 spin_lock(&zone->lock);
1091 isolated_pageblocks = has_isolate_pageblock(zone);
1092 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1094 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1098 struct list_head *list;
1101 * Remove pages from lists in a round-robin fashion. A
1102 * batch_free count is maintained that is incremented when an
1103 * empty list is encountered. This is so more pages are freed
1104 * off fuller lists instead of spinning excessively around empty
1109 if (++migratetype == MIGRATE_PCPTYPES)
1111 list = &pcp->lists[migratetype];
1112 } while (list_empty(list));
1114 /* This is the only non-empty list. Free them all. */
1115 if (batch_free == MIGRATE_PCPTYPES)
1119 int mt; /* migratetype of the to-be-freed page */
1121 page = list_last_entry(list, struct page, lru);
1122 /* must delete as __free_one_page list manipulates */
1123 list_del(&page->lru);
1125 mt = get_pcppage_migratetype(page);
1126 /* MIGRATE_ISOLATE page should not go to pcplists */
1127 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1128 /* Pageblock could have been isolated meanwhile */
1129 if (unlikely(isolated_pageblocks))
1130 mt = get_pageblock_migratetype(page);
1132 if (bulkfree_pcp_prepare(page))
1135 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1136 trace_mm_page_pcpu_drain(page, 0, mt);
1137 } while (--count && --batch_free && !list_empty(list));
1139 spin_unlock(&zone->lock);
1142 static void free_one_page(struct zone *zone,
1143 struct page *page, unsigned long pfn,
1147 unsigned long nr_scanned;
1148 spin_lock(&zone->lock);
1149 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1151 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1153 if (unlikely(has_isolate_pageblock(zone) ||
1154 is_migrate_isolate(migratetype))) {
1155 migratetype = get_pfnblock_migratetype(page, pfn);
1157 __free_one_page(page, pfn, zone, order, migratetype);
1158 spin_unlock(&zone->lock);
1161 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1162 unsigned long zone, int nid)
1164 set_page_links(page, zone, nid, pfn);
1165 init_page_count(page);
1166 page_mapcount_reset(page);
1167 page_cpupid_reset_last(page);
1169 INIT_LIST_HEAD(&page->lru);
1170 #ifdef WANT_PAGE_VIRTUAL
1171 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1172 if (!is_highmem_idx(zone))
1173 set_page_address(page, __va(pfn << PAGE_SHIFT));
1177 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1180 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1183 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1184 static void init_reserved_page(unsigned long pfn)
1189 if (!early_page_uninitialised(pfn))
1192 nid = early_pfn_to_nid(pfn);
1193 pgdat = NODE_DATA(nid);
1195 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1196 struct zone *zone = &pgdat->node_zones[zid];
1198 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1201 __init_single_pfn(pfn, zid, nid);
1204 static inline void init_reserved_page(unsigned long pfn)
1207 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1210 * Initialised pages do not have PageReserved set. This function is
1211 * called for each range allocated by the bootmem allocator and
1212 * marks the pages PageReserved. The remaining valid pages are later
1213 * sent to the buddy page allocator.
1215 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1217 unsigned long start_pfn = PFN_DOWN(start);
1218 unsigned long end_pfn = PFN_UP(end);
1220 for (; start_pfn < end_pfn; start_pfn++) {
1221 if (pfn_valid(start_pfn)) {
1222 struct page *page = pfn_to_page(start_pfn);
1224 init_reserved_page(start_pfn);
1226 /* Avoid false-positive PageTail() */
1227 INIT_LIST_HEAD(&page->lru);
1229 SetPageReserved(page);
1234 static void __free_pages_ok(struct page *page, unsigned int order)
1236 unsigned long flags;
1238 unsigned long pfn = page_to_pfn(page);
1240 if (!free_pages_prepare(page, order, true))
1243 migratetype = get_pfnblock_migratetype(page, pfn);
1244 local_irq_save(flags);
1245 __count_vm_events(PGFREE, 1 << order);
1246 free_one_page(page_zone(page), page, pfn, order, migratetype);
1247 local_irq_restore(flags);
1250 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1252 unsigned int nr_pages = 1 << order;
1253 struct page *p = page;
1257 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1259 __ClearPageReserved(p);
1260 set_page_count(p, 0);
1262 __ClearPageReserved(p);
1263 set_page_count(p, 0);
1265 page_zone(page)->managed_pages += nr_pages;
1266 set_page_refcounted(page);
1267 __free_pages(page, order);
1270 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1271 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1273 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1275 int __meminit early_pfn_to_nid(unsigned long pfn)
1277 static DEFINE_SPINLOCK(early_pfn_lock);
1280 spin_lock(&early_pfn_lock);
1281 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1283 nid = first_online_node;
1284 spin_unlock(&early_pfn_lock);
1290 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1291 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1292 struct mminit_pfnnid_cache *state)
1296 nid = __early_pfn_to_nid(pfn, state);
1297 if (nid >= 0 && nid != node)
1302 /* Only safe to use early in boot when initialisation is single-threaded */
1303 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1305 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1310 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1314 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1315 struct mminit_pfnnid_cache *state)
1322 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1325 if (early_page_uninitialised(pfn))
1327 return __free_pages_boot_core(page, order);
1331 * Check that the whole (or subset of) a pageblock given by the interval of
1332 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1333 * with the migration of free compaction scanner. The scanners then need to
1334 * use only pfn_valid_within() check for arches that allow holes within
1337 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1339 * It's possible on some configurations to have a setup like node0 node1 node0
1340 * i.e. it's possible that all pages within a zones range of pages do not
1341 * belong to a single zone. We assume that a border between node0 and node1
1342 * can occur within a single pageblock, but not a node0 node1 node0
1343 * interleaving within a single pageblock. It is therefore sufficient to check
1344 * the first and last page of a pageblock and avoid checking each individual
1345 * page in a pageblock.
1347 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1348 unsigned long end_pfn, struct zone *zone)
1350 struct page *start_page;
1351 struct page *end_page;
1353 /* end_pfn is one past the range we are checking */
1356 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1359 start_page = pfn_to_page(start_pfn);
1361 if (page_zone(start_page) != zone)
1364 end_page = pfn_to_page(end_pfn);
1366 /* This gives a shorter code than deriving page_zone(end_page) */
1367 if (page_zone_id(start_page) != page_zone_id(end_page))
1373 void set_zone_contiguous(struct zone *zone)
1375 unsigned long block_start_pfn = zone->zone_start_pfn;
1376 unsigned long block_end_pfn;
1378 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1379 for (; block_start_pfn < zone_end_pfn(zone);
1380 block_start_pfn = block_end_pfn,
1381 block_end_pfn += pageblock_nr_pages) {
1383 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1385 if (!__pageblock_pfn_to_page(block_start_pfn,
1386 block_end_pfn, zone))
1390 /* We confirm that there is no hole */
1391 zone->contiguous = true;
1394 void clear_zone_contiguous(struct zone *zone)
1396 zone->contiguous = false;
1399 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1400 static void __init deferred_free_range(struct page *page,
1401 unsigned long pfn, int nr_pages)
1408 /* Free a large naturally-aligned chunk if possible */
1409 if (nr_pages == pageblock_nr_pages &&
1410 (pfn & (pageblock_nr_pages - 1)) == 0) {
1411 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1412 __free_pages_boot_core(page, pageblock_order);
1416 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1417 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1418 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1419 __free_pages_boot_core(page, 0);
1423 /* Completion tracking for deferred_init_memmap() threads */
1424 static atomic_t pgdat_init_n_undone __initdata;
1425 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1427 static inline void __init pgdat_init_report_one_done(void)
1429 if (atomic_dec_and_test(&pgdat_init_n_undone))
1430 complete(&pgdat_init_all_done_comp);
1433 /* Initialise remaining memory on a node */
1434 static int __init deferred_init_memmap(void *data)
1436 pg_data_t *pgdat = data;
1437 int nid = pgdat->node_id;
1438 struct mminit_pfnnid_cache nid_init_state = { };
1439 unsigned long start = jiffies;
1440 unsigned long nr_pages = 0;
1441 unsigned long walk_start, walk_end;
1444 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1445 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1447 if (first_init_pfn == ULONG_MAX) {
1448 pgdat_init_report_one_done();
1452 /* Bind memory initialisation thread to a local node if possible */
1453 if (!cpumask_empty(cpumask))
1454 set_cpus_allowed_ptr(current, cpumask);
1456 /* Sanity check boundaries */
1457 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1458 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1459 pgdat->first_deferred_pfn = ULONG_MAX;
1461 /* Only the highest zone is deferred so find it */
1462 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1463 zone = pgdat->node_zones + zid;
1464 if (first_init_pfn < zone_end_pfn(zone))
1468 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1469 unsigned long pfn, end_pfn;
1470 struct page *page = NULL;
1471 struct page *free_base_page = NULL;
1472 unsigned long free_base_pfn = 0;
1475 end_pfn = min(walk_end, zone_end_pfn(zone));
1476 pfn = first_init_pfn;
1477 if (pfn < walk_start)
1479 if (pfn < zone->zone_start_pfn)
1480 pfn = zone->zone_start_pfn;
1482 for (; pfn < end_pfn; pfn++) {
1483 if (!pfn_valid_within(pfn))
1487 * Ensure pfn_valid is checked every
1488 * pageblock_nr_pages for memory holes
1490 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1491 if (!pfn_valid(pfn)) {
1497 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1502 /* Minimise pfn page lookups and scheduler checks */
1503 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1506 nr_pages += nr_to_free;
1507 deferred_free_range(free_base_page,
1508 free_base_pfn, nr_to_free);
1509 free_base_page = NULL;
1510 free_base_pfn = nr_to_free = 0;
1512 page = pfn_to_page(pfn);
1517 VM_BUG_ON(page_zone(page) != zone);
1521 __init_single_page(page, pfn, zid, nid);
1522 if (!free_base_page) {
1523 free_base_page = page;
1524 free_base_pfn = pfn;
1529 /* Where possible, batch up pages for a single free */
1532 /* Free the current block of pages to allocator */
1533 nr_pages += nr_to_free;
1534 deferred_free_range(free_base_page, free_base_pfn,
1536 free_base_page = NULL;
1537 free_base_pfn = nr_to_free = 0;
1539 /* Free the last block of pages to allocator */
1540 nr_pages += nr_to_free;
1541 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1543 first_init_pfn = max(end_pfn, first_init_pfn);
1546 /* Sanity check that the next zone really is unpopulated */
1547 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1549 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1550 jiffies_to_msecs(jiffies - start));
1552 pgdat_init_report_one_done();
1555 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1557 void __init page_alloc_init_late(void)
1561 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1564 /* There will be num_node_state(N_MEMORY) threads */
1565 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1566 for_each_node_state(nid, N_MEMORY) {
1567 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1570 /* Block until all are initialised */
1571 wait_for_completion(&pgdat_init_all_done_comp);
1573 /* Reinit limits that are based on free pages after the kernel is up */
1574 files_maxfiles_init();
1577 for_each_populated_zone(zone)
1578 set_zone_contiguous(zone);
1582 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1583 void __init init_cma_reserved_pageblock(struct page *page)
1585 unsigned i = pageblock_nr_pages;
1586 struct page *p = page;
1589 __ClearPageReserved(p);
1590 set_page_count(p, 0);
1593 set_pageblock_migratetype(page, MIGRATE_CMA);
1595 if (pageblock_order >= MAX_ORDER) {
1596 i = pageblock_nr_pages;
1599 set_page_refcounted(p);
1600 __free_pages(p, MAX_ORDER - 1);
1601 p += MAX_ORDER_NR_PAGES;
1602 } while (i -= MAX_ORDER_NR_PAGES);
1604 set_page_refcounted(page);
1605 __free_pages(page, pageblock_order);
1608 adjust_managed_page_count(page, pageblock_nr_pages);
1613 * The order of subdivision here is critical for the IO subsystem.
1614 * Please do not alter this order without good reasons and regression
1615 * testing. Specifically, as large blocks of memory are subdivided,
1616 * the order in which smaller blocks are delivered depends on the order
1617 * they're subdivided in this function. This is the primary factor
1618 * influencing the order in which pages are delivered to the IO
1619 * subsystem according to empirical testing, and this is also justified
1620 * by considering the behavior of a buddy system containing a single
1621 * large block of memory acted on by a series of small allocations.
1622 * This behavior is a critical factor in sglist merging's success.
1626 static inline void expand(struct zone *zone, struct page *page,
1627 int low, int high, struct free_area *area,
1630 unsigned long size = 1 << high;
1632 while (high > low) {
1636 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1639 * Mark as guard pages (or page), that will allow to
1640 * merge back to allocator when buddy will be freed.
1641 * Corresponding page table entries will not be touched,
1642 * pages will stay not present in virtual address space
1644 if (set_page_guard(zone, &page[size], high, migratetype))
1647 list_add(&page[size].lru, &area->free_list[migratetype]);
1649 set_page_order(&page[size], high);
1653 static void check_new_page_bad(struct page *page)
1655 const char *bad_reason = NULL;
1656 unsigned long bad_flags = 0;
1658 if (unlikely(atomic_read(&page->_mapcount) != -1))
1659 bad_reason = "nonzero mapcount";
1660 if (unlikely(page->mapping != NULL))
1661 bad_reason = "non-NULL mapping";
1662 if (unlikely(page_ref_count(page) != 0))
1663 bad_reason = "nonzero _count";
1664 if (unlikely(page->flags & __PG_HWPOISON)) {
1665 bad_reason = "HWPoisoned (hardware-corrupted)";
1666 bad_flags = __PG_HWPOISON;
1667 /* Don't complain about hwpoisoned pages */
1668 page_mapcount_reset(page); /* remove PageBuddy */
1671 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1672 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1673 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1676 if (unlikely(page->mem_cgroup))
1677 bad_reason = "page still charged to cgroup";
1679 bad_page(page, bad_reason, bad_flags);
1683 * This page is about to be returned from the page allocator
1685 static inline int check_new_page(struct page *page)
1687 if (likely(page_expected_state(page,
1688 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1691 check_new_page_bad(page);
1695 static inline bool free_pages_prezeroed(bool poisoned)
1697 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1698 page_poisoning_enabled() && poisoned;
1701 #ifdef CONFIG_DEBUG_VM
1702 static bool check_pcp_refill(struct page *page)
1707 static bool check_new_pcp(struct page *page)
1709 return check_new_page(page);
1712 static bool check_pcp_refill(struct page *page)
1714 return check_new_page(page);
1716 static bool check_new_pcp(struct page *page)
1720 #endif /* CONFIG_DEBUG_VM */
1722 static bool check_new_pages(struct page *page, unsigned int order)
1725 for (i = 0; i < (1 << order); i++) {
1726 struct page *p = page + i;
1728 if (unlikely(check_new_page(p)))
1735 inline void post_alloc_hook(struct page *page, unsigned int order,
1738 set_page_private(page, 0);
1739 set_page_refcounted(page);
1741 arch_alloc_page(page, order);
1742 kernel_map_pages(page, 1 << order, 1);
1743 kernel_poison_pages(page, 1 << order, 1);
1744 kasan_alloc_pages(page, order);
1745 set_page_owner(page, order, gfp_flags);
1748 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1749 unsigned int alloc_flags)
1752 bool poisoned = true;
1754 for (i = 0; i < (1 << order); i++) {
1755 struct page *p = page + i;
1757 poisoned &= page_is_poisoned(p);
1760 post_alloc_hook(page, order, gfp_flags);
1762 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1763 for (i = 0; i < (1 << order); i++)
1764 clear_highpage(page + i);
1766 if (order && (gfp_flags & __GFP_COMP))
1767 prep_compound_page(page, order);
1770 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1771 * allocate the page. The expectation is that the caller is taking
1772 * steps that will free more memory. The caller should avoid the page
1773 * being used for !PFMEMALLOC purposes.
1775 if (alloc_flags & ALLOC_NO_WATERMARKS)
1776 set_page_pfmemalloc(page);
1778 clear_page_pfmemalloc(page);
1782 * Go through the free lists for the given migratetype and remove
1783 * the smallest available page from the freelists
1786 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1789 unsigned int current_order;
1790 struct free_area *area;
1793 /* Find a page of the appropriate size in the preferred list */
1794 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1795 area = &(zone->free_area[current_order]);
1796 page = list_first_entry_or_null(&area->free_list[migratetype],
1800 list_del(&page->lru);
1801 rmv_page_order(page);
1803 expand(zone, page, order, current_order, area, migratetype);
1804 set_pcppage_migratetype(page, migratetype);
1813 * This array describes the order lists are fallen back to when
1814 * the free lists for the desirable migrate type are depleted
1816 static int fallbacks[MIGRATE_TYPES][4] = {
1817 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1818 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1819 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1821 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1823 #ifdef CONFIG_MEMORY_ISOLATION
1824 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1829 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1832 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1835 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1836 unsigned int order) { return NULL; }
1840 * Move the free pages in a range to the free lists of the requested type.
1841 * Note that start_page and end_pages are not aligned on a pageblock
1842 * boundary. If alignment is required, use move_freepages_block()
1844 int move_freepages(struct zone *zone,
1845 struct page *start_page, struct page *end_page,
1850 int pages_moved = 0;
1852 #ifndef CONFIG_HOLES_IN_ZONE
1854 * page_zone is not safe to call in this context when
1855 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1856 * anyway as we check zone boundaries in move_freepages_block().
1857 * Remove at a later date when no bug reports exist related to
1858 * grouping pages by mobility
1860 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1863 for (page = start_page; page <= end_page;) {
1864 if (!pfn_valid_within(page_to_pfn(page))) {
1869 /* Make sure we are not inadvertently changing nodes */
1870 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1872 if (!PageBuddy(page)) {
1877 order = page_order(page);
1878 list_move(&page->lru,
1879 &zone->free_area[order].free_list[migratetype]);
1881 pages_moved += 1 << order;
1887 int move_freepages_block(struct zone *zone, struct page *page,
1890 unsigned long start_pfn, end_pfn;
1891 struct page *start_page, *end_page;
1893 start_pfn = page_to_pfn(page);
1894 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1895 start_page = pfn_to_page(start_pfn);
1896 end_page = start_page + pageblock_nr_pages - 1;
1897 end_pfn = start_pfn + pageblock_nr_pages - 1;
1899 /* Do not cross zone boundaries */
1900 if (!zone_spans_pfn(zone, start_pfn))
1902 if (!zone_spans_pfn(zone, end_pfn))
1905 return move_freepages(zone, start_page, end_page, migratetype);
1908 static void change_pageblock_range(struct page *pageblock_page,
1909 int start_order, int migratetype)
1911 int nr_pageblocks = 1 << (start_order - pageblock_order);
1913 while (nr_pageblocks--) {
1914 set_pageblock_migratetype(pageblock_page, migratetype);
1915 pageblock_page += pageblock_nr_pages;
1920 * When we are falling back to another migratetype during allocation, try to
1921 * steal extra free pages from the same pageblocks to satisfy further
1922 * allocations, instead of polluting multiple pageblocks.
1924 * If we are stealing a relatively large buddy page, it is likely there will
1925 * be more free pages in the pageblock, so try to steal them all. For
1926 * reclaimable and unmovable allocations, we steal regardless of page size,
1927 * as fragmentation caused by those allocations polluting movable pageblocks
1928 * is worse than movable allocations stealing from unmovable and reclaimable
1931 static bool can_steal_fallback(unsigned int order, int start_mt)
1934 * Leaving this order check is intended, although there is
1935 * relaxed order check in next check. The reason is that
1936 * we can actually steal whole pageblock if this condition met,
1937 * but, below check doesn't guarantee it and that is just heuristic
1938 * so could be changed anytime.
1940 if (order >= pageblock_order)
1943 if (order >= pageblock_order / 2 ||
1944 start_mt == MIGRATE_RECLAIMABLE ||
1945 start_mt == MIGRATE_UNMOVABLE ||
1946 page_group_by_mobility_disabled)
1953 * This function implements actual steal behaviour. If order is large enough,
1954 * we can steal whole pageblock. If not, we first move freepages in this
1955 * pageblock and check whether half of pages are moved or not. If half of
1956 * pages are moved, we can change migratetype of pageblock and permanently
1957 * use it's pages as requested migratetype in the future.
1959 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1962 unsigned int current_order = page_order(page);
1965 /* Take ownership for orders >= pageblock_order */
1966 if (current_order >= pageblock_order) {
1967 change_pageblock_range(page, current_order, start_type);
1971 pages = move_freepages_block(zone, page, start_type);
1973 /* Claim the whole block if over half of it is free */
1974 if (pages >= (1 << (pageblock_order-1)) ||
1975 page_group_by_mobility_disabled)
1976 set_pageblock_migratetype(page, start_type);
1980 * Check whether there is a suitable fallback freepage with requested order.
1981 * If only_stealable is true, this function returns fallback_mt only if
1982 * we can steal other freepages all together. This would help to reduce
1983 * fragmentation due to mixed migratetype pages in one pageblock.
1985 int find_suitable_fallback(struct free_area *area, unsigned int order,
1986 int migratetype, bool only_stealable, bool *can_steal)
1991 if (area->nr_free == 0)
1996 fallback_mt = fallbacks[migratetype][i];
1997 if (fallback_mt == MIGRATE_TYPES)
2000 if (list_empty(&area->free_list[fallback_mt]))
2003 if (can_steal_fallback(order, migratetype))
2006 if (!only_stealable)
2017 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2018 * there are no empty page blocks that contain a page with a suitable order
2020 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2021 unsigned int alloc_order)
2024 unsigned long max_managed, flags;
2027 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2028 * Check is race-prone but harmless.
2030 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2031 if (zone->nr_reserved_highatomic >= max_managed)
2034 spin_lock_irqsave(&zone->lock, flags);
2036 /* Recheck the nr_reserved_highatomic limit under the lock */
2037 if (zone->nr_reserved_highatomic >= max_managed)
2041 mt = get_pageblock_migratetype(page);
2042 if (mt != MIGRATE_HIGHATOMIC &&
2043 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2044 zone->nr_reserved_highatomic += pageblock_nr_pages;
2045 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2046 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2050 spin_unlock_irqrestore(&zone->lock, flags);
2054 * Used when an allocation is about to fail under memory pressure. This
2055 * potentially hurts the reliability of high-order allocations when under
2056 * intense memory pressure but failed atomic allocations should be easier
2057 * to recover from than an OOM.
2059 * If @force is true, try to unreserve a pageblock even though highatomic
2060 * pageblock is exhausted.
2062 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2065 struct zonelist *zonelist = ac->zonelist;
2066 unsigned long flags;
2073 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2076 * Preserve at least one pageblock unless memory pressure
2079 if (!force && zone->nr_reserved_highatomic <=
2083 spin_lock_irqsave(&zone->lock, flags);
2084 for (order = 0; order < MAX_ORDER; order++) {
2085 struct free_area *area = &(zone->free_area[order]);
2087 page = list_first_entry_or_null(
2088 &area->free_list[MIGRATE_HIGHATOMIC],
2094 * In page freeing path, migratetype change is racy so
2095 * we can counter several free pages in a pageblock
2096 * in this loop althoug we changed the pageblock type
2097 * from highatomic to ac->migratetype. So we should
2098 * adjust the count once.
2100 if (get_pageblock_migratetype(page) ==
2101 MIGRATE_HIGHATOMIC) {
2103 * It should never happen but changes to
2104 * locking could inadvertently allow a per-cpu
2105 * drain to add pages to MIGRATE_HIGHATOMIC
2106 * while unreserving so be safe and watch for
2109 zone->nr_reserved_highatomic -= min(
2111 zone->nr_reserved_highatomic);
2115 * Convert to ac->migratetype and avoid the normal
2116 * pageblock stealing heuristics. Minimally, the caller
2117 * is doing the work and needs the pages. More
2118 * importantly, if the block was always converted to
2119 * MIGRATE_UNMOVABLE or another type then the number
2120 * of pageblocks that cannot be completely freed
2123 set_pageblock_migratetype(page, ac->migratetype);
2124 ret = move_freepages_block(zone, page, ac->migratetype);
2126 spin_unlock_irqrestore(&zone->lock, flags);
2130 spin_unlock_irqrestore(&zone->lock, flags);
2136 /* Remove an element from the buddy allocator from the fallback list */
2137 static inline struct page *
2138 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2140 struct free_area *area;
2141 unsigned int current_order;
2146 /* Find the largest possible block of pages in the other list */
2147 for (current_order = MAX_ORDER-1;
2148 current_order >= order && current_order <= MAX_ORDER-1;
2150 area = &(zone->free_area[current_order]);
2151 fallback_mt = find_suitable_fallback(area, current_order,
2152 start_migratetype, false, &can_steal);
2153 if (fallback_mt == -1)
2156 page = list_first_entry(&area->free_list[fallback_mt],
2159 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC)
2160 steal_suitable_fallback(zone, page, start_migratetype);
2162 /* Remove the page from the freelists */
2164 list_del(&page->lru);
2165 rmv_page_order(page);
2167 expand(zone, page, order, current_order, area,
2170 * The pcppage_migratetype may differ from pageblock's
2171 * migratetype depending on the decisions in
2172 * find_suitable_fallback(). This is OK as long as it does not
2173 * differ for MIGRATE_CMA pageblocks. Those can be used as
2174 * fallback only via special __rmqueue_cma_fallback() function
2176 set_pcppage_migratetype(page, start_migratetype);
2178 trace_mm_page_alloc_extfrag(page, order, current_order,
2179 start_migratetype, fallback_mt);
2188 * Do the hard work of removing an element from the buddy allocator.
2189 * Call me with the zone->lock already held.
2191 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2196 page = __rmqueue_smallest(zone, order, migratetype);
2197 if (unlikely(!page)) {
2198 if (migratetype == MIGRATE_MOVABLE)
2199 page = __rmqueue_cma_fallback(zone, order);
2202 page = __rmqueue_fallback(zone, order, migratetype);
2205 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2210 * Obtain a specified number of elements from the buddy allocator, all under
2211 * a single hold of the lock, for efficiency. Add them to the supplied list.
2212 * Returns the number of new pages which were placed at *list.
2214 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2215 unsigned long count, struct list_head *list,
2216 int migratetype, bool cold)
2220 spin_lock(&zone->lock);
2221 for (i = 0; i < count; ++i) {
2222 struct page *page = __rmqueue(zone, order, migratetype);
2223 if (unlikely(page == NULL))
2226 if (unlikely(check_pcp_refill(page)))
2230 * Split buddy pages returned by expand() are received here
2231 * in physical page order. The page is added to the callers and
2232 * list and the list head then moves forward. From the callers
2233 * perspective, the linked list is ordered by page number in
2234 * some conditions. This is useful for IO devices that can
2235 * merge IO requests if the physical pages are ordered
2239 list_add(&page->lru, list);
2241 list_add_tail(&page->lru, list);
2244 if (is_migrate_cma(get_pcppage_migratetype(page)))
2245 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2250 * i pages were removed from the buddy list even if some leak due
2251 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2252 * on i. Do not confuse with 'alloced' which is the number of
2253 * pages added to the pcp list.
2255 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2256 spin_unlock(&zone->lock);
2262 * Called from the vmstat counter updater to drain pagesets of this
2263 * currently executing processor on remote nodes after they have
2266 * Note that this function must be called with the thread pinned to
2267 * a single processor.
2269 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2271 unsigned long flags;
2272 int to_drain, batch;
2274 local_irq_save(flags);
2275 batch = READ_ONCE(pcp->batch);
2276 to_drain = min(pcp->count, batch);
2278 free_pcppages_bulk(zone, to_drain, pcp);
2279 pcp->count -= to_drain;
2281 local_irq_restore(flags);
2286 * Drain pcplists of the indicated processor and zone.
2288 * The processor must either be the current processor and the
2289 * thread pinned to the current processor or a processor that
2292 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2294 unsigned long flags;
2295 struct per_cpu_pageset *pset;
2296 struct per_cpu_pages *pcp;
2298 local_irq_save(flags);
2299 pset = per_cpu_ptr(zone->pageset, cpu);
2303 free_pcppages_bulk(zone, pcp->count, pcp);
2306 local_irq_restore(flags);
2310 * Drain pcplists of all zones on the indicated processor.
2312 * The processor must either be the current processor and the
2313 * thread pinned to the current processor or a processor that
2316 static void drain_pages(unsigned int cpu)
2320 for_each_populated_zone(zone) {
2321 drain_pages_zone(cpu, zone);
2326 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2328 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2329 * the single zone's pages.
2331 void drain_local_pages(struct zone *zone)
2333 int cpu = smp_processor_id();
2336 drain_pages_zone(cpu, zone);
2342 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2344 * When zone parameter is non-NULL, spill just the single zone's pages.
2346 * Note that this code is protected against sending an IPI to an offline
2347 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2348 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2349 * nothing keeps CPUs from showing up after we populated the cpumask and
2350 * before the call to on_each_cpu_mask().
2352 void drain_all_pages(struct zone *zone)
2357 * Allocate in the BSS so we wont require allocation in
2358 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2360 static cpumask_t cpus_with_pcps;
2363 * We don't care about racing with CPU hotplug event
2364 * as offline notification will cause the notified
2365 * cpu to drain that CPU pcps and on_each_cpu_mask
2366 * disables preemption as part of its processing
2368 for_each_online_cpu(cpu) {
2369 struct per_cpu_pageset *pcp;
2371 bool has_pcps = false;
2374 pcp = per_cpu_ptr(zone->pageset, cpu);
2378 for_each_populated_zone(z) {
2379 pcp = per_cpu_ptr(z->pageset, cpu);
2380 if (pcp->pcp.count) {
2388 cpumask_set_cpu(cpu, &cpus_with_pcps);
2390 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2392 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2396 #ifdef CONFIG_HIBERNATION
2398 void mark_free_pages(struct zone *zone)
2400 unsigned long pfn, max_zone_pfn;
2401 unsigned long flags;
2402 unsigned int order, t;
2405 if (zone_is_empty(zone))
2408 spin_lock_irqsave(&zone->lock, flags);
2410 max_zone_pfn = zone_end_pfn(zone);
2411 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2412 if (pfn_valid(pfn)) {
2413 page = pfn_to_page(pfn);
2415 if (page_zone(page) != zone)
2418 if (!swsusp_page_is_forbidden(page))
2419 swsusp_unset_page_free(page);
2422 for_each_migratetype_order(order, t) {
2423 list_for_each_entry(page,
2424 &zone->free_area[order].free_list[t], lru) {
2427 pfn = page_to_pfn(page);
2428 for (i = 0; i < (1UL << order); i++)
2429 swsusp_set_page_free(pfn_to_page(pfn + i));
2432 spin_unlock_irqrestore(&zone->lock, flags);
2434 #endif /* CONFIG_PM */
2437 * Free a 0-order page
2438 * cold == true ? free a cold page : free a hot page
2440 void free_hot_cold_page(struct page *page, bool cold)
2442 struct zone *zone = page_zone(page);
2443 struct per_cpu_pages *pcp;
2444 unsigned long flags;
2445 unsigned long pfn = page_to_pfn(page);
2448 if (!free_pcp_prepare(page))
2451 migratetype = get_pfnblock_migratetype(page, pfn);
2452 set_pcppage_migratetype(page, migratetype);
2453 local_irq_save(flags);
2454 __count_vm_event(PGFREE);
2457 * We only track unmovable, reclaimable and movable on pcp lists.
2458 * Free ISOLATE pages back to the allocator because they are being
2459 * offlined but treat RESERVE as movable pages so we can get those
2460 * areas back if necessary. Otherwise, we may have to free
2461 * excessively into the page allocator
2463 if (migratetype >= MIGRATE_PCPTYPES) {
2464 if (unlikely(is_migrate_isolate(migratetype))) {
2465 free_one_page(zone, page, pfn, 0, migratetype);
2468 migratetype = MIGRATE_MOVABLE;
2471 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2473 list_add(&page->lru, &pcp->lists[migratetype]);
2475 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2477 if (pcp->count >= pcp->high) {
2478 unsigned long batch = READ_ONCE(pcp->batch);
2479 free_pcppages_bulk(zone, batch, pcp);
2480 pcp->count -= batch;
2484 local_irq_restore(flags);
2488 * Free a list of 0-order pages
2490 void free_hot_cold_page_list(struct list_head *list, bool cold)
2492 struct page *page, *next;
2494 list_for_each_entry_safe(page, next, list, lru) {
2495 trace_mm_page_free_batched(page, cold);
2496 free_hot_cold_page(page, cold);
2501 * split_page takes a non-compound higher-order page, and splits it into
2502 * n (1<<order) sub-pages: page[0..n]
2503 * Each sub-page must be freed individually.
2505 * Note: this is probably too low level an operation for use in drivers.
2506 * Please consult with lkml before using this in your driver.
2508 void split_page(struct page *page, unsigned int order)
2512 VM_BUG_ON_PAGE(PageCompound(page), page);
2513 VM_BUG_ON_PAGE(!page_count(page), page);
2515 #ifdef CONFIG_KMEMCHECK
2517 * Split shadow pages too, because free(page[0]) would
2518 * otherwise free the whole shadow.
2520 if (kmemcheck_page_is_tracked(page))
2521 split_page(virt_to_page(page[0].shadow), order);
2524 for (i = 1; i < (1 << order); i++)
2525 set_page_refcounted(page + i);
2526 split_page_owner(page, order);
2528 EXPORT_SYMBOL_GPL(split_page);
2530 int __isolate_free_page(struct page *page, unsigned int order)
2532 unsigned long watermark;
2536 BUG_ON(!PageBuddy(page));
2538 zone = page_zone(page);
2539 mt = get_pageblock_migratetype(page);
2541 if (!is_migrate_isolate(mt)) {
2543 * Obey watermarks as if the page was being allocated. We can
2544 * emulate a high-order watermark check with a raised order-0
2545 * watermark, because we already know our high-order page
2548 watermark = min_wmark_pages(zone) + (1UL << order);
2549 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2552 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2555 /* Remove page from free list */
2556 list_del(&page->lru);
2557 zone->free_area[order].nr_free--;
2558 rmv_page_order(page);
2561 * Set the pageblock if the isolated page is at least half of a
2564 if (order >= pageblock_order - 1) {
2565 struct page *endpage = page + (1 << order) - 1;
2566 for (; page < endpage; page += pageblock_nr_pages) {
2567 int mt = get_pageblock_migratetype(page);
2568 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2569 && mt != MIGRATE_HIGHATOMIC)
2570 set_pageblock_migratetype(page,
2576 return 1UL << order;
2580 * Update NUMA hit/miss statistics
2582 * Must be called with interrupts disabled.
2584 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2587 enum zone_stat_item local_stat = NUMA_LOCAL;
2589 if (z->node != numa_node_id())
2590 local_stat = NUMA_OTHER;
2592 if (z->node == preferred_zone->node)
2593 __inc_zone_state(z, NUMA_HIT);
2595 __inc_zone_state(z, NUMA_MISS);
2596 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2598 __inc_zone_state(z, local_stat);
2603 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2606 struct page *buffered_rmqueue(struct zone *preferred_zone,
2607 struct zone *zone, unsigned int order,
2608 gfp_t gfp_flags, unsigned int alloc_flags,
2611 unsigned long flags;
2613 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2615 if (likely(order == 0)) {
2616 struct per_cpu_pages *pcp;
2617 struct list_head *list;
2619 local_irq_save(flags);
2621 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2622 list = &pcp->lists[migratetype];
2623 if (list_empty(list)) {
2624 pcp->count += rmqueue_bulk(zone, 0,
2627 if (unlikely(list_empty(list)))
2632 page = list_last_entry(list, struct page, lru);
2634 page = list_first_entry(list, struct page, lru);
2636 list_del(&page->lru);
2639 } while (check_new_pcp(page));
2642 * We most definitely don't want callers attempting to
2643 * allocate greater than order-1 page units with __GFP_NOFAIL.
2645 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2646 spin_lock_irqsave(&zone->lock, flags);
2650 if (alloc_flags & ALLOC_HARDER) {
2651 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2653 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2656 page = __rmqueue(zone, order, migratetype);
2657 } while (page && check_new_pages(page, order));
2658 spin_unlock(&zone->lock);
2661 __mod_zone_freepage_state(zone, -(1 << order),
2662 get_pcppage_migratetype(page));
2665 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2666 zone_statistics(preferred_zone, zone);
2667 local_irq_restore(flags);
2669 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2673 local_irq_restore(flags);
2677 #ifdef CONFIG_FAIL_PAGE_ALLOC
2680 struct fault_attr attr;
2682 bool ignore_gfp_highmem;
2683 bool ignore_gfp_reclaim;
2685 } fail_page_alloc = {
2686 .attr = FAULT_ATTR_INITIALIZER,
2687 .ignore_gfp_reclaim = true,
2688 .ignore_gfp_highmem = true,
2692 static int __init setup_fail_page_alloc(char *str)
2694 return setup_fault_attr(&fail_page_alloc.attr, str);
2696 __setup("fail_page_alloc=", setup_fail_page_alloc);
2698 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2700 if (order < fail_page_alloc.min_order)
2702 if (gfp_mask & __GFP_NOFAIL)
2704 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2706 if (fail_page_alloc.ignore_gfp_reclaim &&
2707 (gfp_mask & __GFP_DIRECT_RECLAIM))
2710 return should_fail(&fail_page_alloc.attr, 1 << order);
2713 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2715 static int __init fail_page_alloc_debugfs(void)
2717 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2720 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2721 &fail_page_alloc.attr);
2723 return PTR_ERR(dir);
2725 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2726 &fail_page_alloc.ignore_gfp_reclaim))
2728 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2729 &fail_page_alloc.ignore_gfp_highmem))
2731 if (!debugfs_create_u32("min-order", mode, dir,
2732 &fail_page_alloc.min_order))
2737 debugfs_remove_recursive(dir);
2742 late_initcall(fail_page_alloc_debugfs);
2744 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2746 #else /* CONFIG_FAIL_PAGE_ALLOC */
2748 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2753 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2756 * Return true if free base pages are above 'mark'. For high-order checks it
2757 * will return true of the order-0 watermark is reached and there is at least
2758 * one free page of a suitable size. Checking now avoids taking the zone lock
2759 * to check in the allocation paths if no pages are free.
2761 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2762 int classzone_idx, unsigned int alloc_flags,
2767 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2769 /* free_pages may go negative - that's OK */
2770 free_pages -= (1 << order) - 1;
2772 if (alloc_flags & ALLOC_HIGH)
2776 * If the caller does not have rights to ALLOC_HARDER then subtract
2777 * the high-atomic reserves. This will over-estimate the size of the
2778 * atomic reserve but it avoids a search.
2780 if (likely(!alloc_harder))
2781 free_pages -= z->nr_reserved_highatomic;
2786 /* If allocation can't use CMA areas don't use free CMA pages */
2787 if (!(alloc_flags & ALLOC_CMA))
2788 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2792 * Check watermarks for an order-0 allocation request. If these
2793 * are not met, then a high-order request also cannot go ahead
2794 * even if a suitable page happened to be free.
2796 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2799 /* If this is an order-0 request then the watermark is fine */
2803 /* For a high-order request, check at least one suitable page is free */
2804 for (o = order; o < MAX_ORDER; o++) {
2805 struct free_area *area = &z->free_area[o];
2814 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2815 if (!list_empty(&area->free_list[mt]))
2820 if ((alloc_flags & ALLOC_CMA) &&
2821 !list_empty(&area->free_list[MIGRATE_CMA])) {
2829 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2830 int classzone_idx, unsigned int alloc_flags)
2832 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2833 zone_page_state(z, NR_FREE_PAGES));
2836 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2837 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2839 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2843 /* If allocation can't use CMA areas don't use free CMA pages */
2844 if (!(alloc_flags & ALLOC_CMA))
2845 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2849 * Fast check for order-0 only. If this fails then the reserves
2850 * need to be calculated. There is a corner case where the check
2851 * passes but only the high-order atomic reserve are free. If
2852 * the caller is !atomic then it'll uselessly search the free
2853 * list. That corner case is then slower but it is harmless.
2855 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2858 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2862 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2863 unsigned long mark, int classzone_idx)
2865 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2867 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2868 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2870 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2875 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2877 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2880 #else /* CONFIG_NUMA */
2881 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2885 #endif /* CONFIG_NUMA */
2888 * get_page_from_freelist goes through the zonelist trying to allocate
2891 static struct page *
2892 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2893 const struct alloc_context *ac)
2895 struct zoneref *z = ac->preferred_zoneref;
2897 struct pglist_data *last_pgdat_dirty_limit = NULL;
2900 * Scan zonelist, looking for a zone with enough free.
2901 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2903 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2908 if (cpusets_enabled() &&
2909 (alloc_flags & ALLOC_CPUSET) &&
2910 !__cpuset_zone_allowed(zone, gfp_mask))
2913 * When allocating a page cache page for writing, we
2914 * want to get it from a node that is within its dirty
2915 * limit, such that no single node holds more than its
2916 * proportional share of globally allowed dirty pages.
2917 * The dirty limits take into account the node's
2918 * lowmem reserves and high watermark so that kswapd
2919 * should be able to balance it without having to
2920 * write pages from its LRU list.
2922 * XXX: For now, allow allocations to potentially
2923 * exceed the per-node dirty limit in the slowpath
2924 * (spread_dirty_pages unset) before going into reclaim,
2925 * which is important when on a NUMA setup the allowed
2926 * nodes are together not big enough to reach the
2927 * global limit. The proper fix for these situations
2928 * will require awareness of nodes in the
2929 * dirty-throttling and the flusher threads.
2931 if (ac->spread_dirty_pages) {
2932 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2935 if (!node_dirty_ok(zone->zone_pgdat)) {
2936 last_pgdat_dirty_limit = zone->zone_pgdat;
2941 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2942 if (!zone_watermark_fast(zone, order, mark,
2943 ac_classzone_idx(ac), alloc_flags)) {
2946 /* Checked here to keep the fast path fast */
2947 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2948 if (alloc_flags & ALLOC_NO_WATERMARKS)
2951 if (node_reclaim_mode == 0 ||
2952 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2955 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2957 case NODE_RECLAIM_NOSCAN:
2960 case NODE_RECLAIM_FULL:
2961 /* scanned but unreclaimable */
2964 /* did we reclaim enough */
2965 if (zone_watermark_ok(zone, order, mark,
2966 ac_classzone_idx(ac), alloc_flags))
2974 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2975 gfp_mask, alloc_flags, ac->migratetype);
2977 prep_new_page(page, order, gfp_mask, alloc_flags);
2980 * If this is a high-order atomic allocation then check
2981 * if the pageblock should be reserved for the future
2983 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2984 reserve_highatomic_pageblock(page, zone, order);
2994 * Large machines with many possible nodes should not always dump per-node
2995 * meminfo in irq context.
2997 static inline bool should_suppress_show_mem(void)
3002 ret = in_interrupt();
3007 static void warn_alloc_show_mem(gfp_t gfp_mask)
3009 unsigned int filter = SHOW_MEM_FILTER_NODES;
3010 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3012 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3016 * This documents exceptions given to allocations in certain
3017 * contexts that are allowed to allocate outside current's set
3020 if (!(gfp_mask & __GFP_NOMEMALLOC))
3021 if (test_thread_flag(TIF_MEMDIE) ||
3022 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3023 filter &= ~SHOW_MEM_FILTER_NODES;
3024 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3025 filter &= ~SHOW_MEM_FILTER_NODES;
3030 void warn_alloc(gfp_t gfp_mask, const char *fmt, ...)
3032 struct va_format vaf;
3034 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3035 DEFAULT_RATELIMIT_BURST);
3037 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3038 debug_guardpage_minorder() > 0)
3041 pr_warn("%s: ", current->comm);
3043 va_start(args, fmt);
3046 pr_cont("%pV", &vaf);
3049 pr_cont(", mode:%#x(%pGg)\n", gfp_mask, &gfp_mask);
3052 warn_alloc_show_mem(gfp_mask);
3055 static inline struct page *
3056 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3057 const struct alloc_context *ac, unsigned long *did_some_progress)
3059 struct oom_control oc = {
3060 .zonelist = ac->zonelist,
3061 .nodemask = ac->nodemask,
3063 .gfp_mask = gfp_mask,
3068 *did_some_progress = 0;
3071 * Acquire the oom lock. If that fails, somebody else is
3072 * making progress for us.
3074 if (!mutex_trylock(&oom_lock)) {
3075 *did_some_progress = 1;
3076 schedule_timeout_uninterruptible(1);
3081 * Go through the zonelist yet one more time, keep very high watermark
3082 * here, this is only to catch a parallel oom killing, we must fail if
3083 * we're still under heavy pressure.
3085 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3086 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3090 if (!(gfp_mask & __GFP_NOFAIL)) {
3091 /* Coredumps can quickly deplete all memory reserves */
3092 if (current->flags & PF_DUMPCORE)
3094 /* The OOM killer will not help higher order allocs */
3095 if (order > PAGE_ALLOC_COSTLY_ORDER)
3097 /* The OOM killer does not needlessly kill tasks for lowmem */
3098 if (ac->high_zoneidx < ZONE_NORMAL)
3100 if (pm_suspended_storage())
3103 * XXX: GFP_NOFS allocations should rather fail than rely on
3104 * other request to make a forward progress.
3105 * We are in an unfortunate situation where out_of_memory cannot
3106 * do much for this context but let's try it to at least get
3107 * access to memory reserved if the current task is killed (see
3108 * out_of_memory). Once filesystems are ready to handle allocation
3109 * failures more gracefully we should just bail out here.
3112 /* The OOM killer may not free memory on a specific node */
3113 if (gfp_mask & __GFP_THISNODE)
3116 /* Exhausted what can be done so it's blamo time */
3117 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3118 *did_some_progress = 1;
3120 if (gfp_mask & __GFP_NOFAIL) {
3121 page = get_page_from_freelist(gfp_mask, order,
3122 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3124 * fallback to ignore cpuset restriction if our nodes
3128 page = get_page_from_freelist(gfp_mask, order,
3129 ALLOC_NO_WATERMARKS, ac);
3133 mutex_unlock(&oom_lock);
3138 * Maximum number of compaction retries wit a progress before OOM
3139 * killer is consider as the only way to move forward.
3141 #define MAX_COMPACT_RETRIES 16
3143 #ifdef CONFIG_COMPACTION
3144 /* Try memory compaction for high-order allocations before reclaim */
3145 static struct page *
3146 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3147 unsigned int alloc_flags, const struct alloc_context *ac,
3148 enum compact_priority prio, enum compact_result *compact_result)
3155 current->flags |= PF_MEMALLOC;
3156 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3158 current->flags &= ~PF_MEMALLOC;
3160 if (*compact_result <= COMPACT_INACTIVE)
3164 * At least in one zone compaction wasn't deferred or skipped, so let's
3165 * count a compaction stall
3167 count_vm_event(COMPACTSTALL);
3169 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3172 struct zone *zone = page_zone(page);
3174 zone->compact_blockskip_flush = false;
3175 compaction_defer_reset(zone, order, true);
3176 count_vm_event(COMPACTSUCCESS);
3181 * It's bad if compaction run occurs and fails. The most likely reason
3182 * is that pages exist, but not enough to satisfy watermarks.
3184 count_vm_event(COMPACTFAIL);
3192 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3193 enum compact_result compact_result,
3194 enum compact_priority *compact_priority,
3195 int *compaction_retries)
3197 int max_retries = MAX_COMPACT_RETRIES;
3203 if (compaction_made_progress(compact_result))
3204 (*compaction_retries)++;
3207 * compaction considers all the zone as desperately out of memory
3208 * so it doesn't really make much sense to retry except when the
3209 * failure could be caused by insufficient priority
3211 if (compaction_failed(compact_result))
3212 goto check_priority;
3215 * make sure the compaction wasn't deferred or didn't bail out early
3216 * due to locks contention before we declare that we should give up.
3217 * But do not retry if the given zonelist is not suitable for
3220 if (compaction_withdrawn(compact_result))
3221 return compaction_zonelist_suitable(ac, order, alloc_flags);
3224 * !costly requests are much more important than __GFP_REPEAT
3225 * costly ones because they are de facto nofail and invoke OOM
3226 * killer to move on while costly can fail and users are ready
3227 * to cope with that. 1/4 retries is rather arbitrary but we
3228 * would need much more detailed feedback from compaction to
3229 * make a better decision.
3231 if (order > PAGE_ALLOC_COSTLY_ORDER)
3233 if (*compaction_retries <= max_retries)
3237 * Make sure there are attempts at the highest priority if we exhausted
3238 * all retries or failed at the lower priorities.
3241 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3242 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3243 if (*compact_priority > min_priority) {
3244 (*compact_priority)--;
3245 *compaction_retries = 0;
3251 static inline struct page *
3252 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3253 unsigned int alloc_flags, const struct alloc_context *ac,
3254 enum compact_priority prio, enum compact_result *compact_result)
3256 *compact_result = COMPACT_SKIPPED;
3261 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3262 enum compact_result compact_result,
3263 enum compact_priority *compact_priority,
3264 int *compaction_retries)
3269 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3273 * There are setups with compaction disabled which would prefer to loop
3274 * inside the allocator rather than hit the oom killer prematurely.
3275 * Let's give them a good hope and keep retrying while the order-0
3276 * watermarks are OK.
3278 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3280 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3281 ac_classzone_idx(ac), alloc_flags))
3286 #endif /* CONFIG_COMPACTION */
3288 /* Perform direct synchronous page reclaim */
3290 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3291 const struct alloc_context *ac)
3293 struct reclaim_state reclaim_state;
3298 /* We now go into synchronous reclaim */
3299 cpuset_memory_pressure_bump();
3300 current->flags |= PF_MEMALLOC;
3301 lockdep_set_current_reclaim_state(gfp_mask);
3302 reclaim_state.reclaimed_slab = 0;
3303 current->reclaim_state = &reclaim_state;
3305 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3308 current->reclaim_state = NULL;
3309 lockdep_clear_current_reclaim_state();
3310 current->flags &= ~PF_MEMALLOC;
3317 /* The really slow allocator path where we enter direct reclaim */
3318 static inline struct page *
3319 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3320 unsigned int alloc_flags, const struct alloc_context *ac,
3321 unsigned long *did_some_progress)
3323 struct page *page = NULL;
3324 bool drained = false;
3326 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3327 if (unlikely(!(*did_some_progress)))
3331 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3334 * If an allocation failed after direct reclaim, it could be because
3335 * pages are pinned on the per-cpu lists or in high alloc reserves.
3336 * Shrink them them and try again
3338 if (!page && !drained) {
3339 unreserve_highatomic_pageblock(ac, false);
3340 drain_all_pages(NULL);
3348 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3352 pg_data_t *last_pgdat = NULL;
3354 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3355 ac->high_zoneidx, ac->nodemask) {
3356 if (last_pgdat != zone->zone_pgdat)
3357 wakeup_kswapd(zone, order, ac->high_zoneidx);
3358 last_pgdat = zone->zone_pgdat;
3362 static inline unsigned int
3363 gfp_to_alloc_flags(gfp_t gfp_mask)
3365 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3367 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3368 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3371 * The caller may dip into page reserves a bit more if the caller
3372 * cannot run direct reclaim, or if the caller has realtime scheduling
3373 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3374 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3376 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3378 if (gfp_mask & __GFP_ATOMIC) {
3380 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3381 * if it can't schedule.
3383 if (!(gfp_mask & __GFP_NOMEMALLOC))
3384 alloc_flags |= ALLOC_HARDER;
3386 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3387 * comment for __cpuset_node_allowed().
3389 alloc_flags &= ~ALLOC_CPUSET;
3390 } else if (unlikely(rt_task(current)) && !in_interrupt())
3391 alloc_flags |= ALLOC_HARDER;
3394 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3395 alloc_flags |= ALLOC_CMA;
3400 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3402 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3405 if (gfp_mask & __GFP_MEMALLOC)
3407 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3409 if (!in_interrupt() &&
3410 ((current->flags & PF_MEMALLOC) ||
3411 unlikely(test_thread_flag(TIF_MEMDIE))))
3418 * Maximum number of reclaim retries without any progress before OOM killer
3419 * is consider as the only way to move forward.
3421 #define MAX_RECLAIM_RETRIES 16
3424 * Checks whether it makes sense to retry the reclaim to make a forward progress
3425 * for the given allocation request.
3426 * The reclaim feedback represented by did_some_progress (any progress during
3427 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3428 * any progress in a row) is considered as well as the reclaimable pages on the
3429 * applicable zone list (with a backoff mechanism which is a function of
3430 * no_progress_loops).
3432 * Returns true if a retry is viable or false to enter the oom path.
3435 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3436 struct alloc_context *ac, int alloc_flags,
3437 bool did_some_progress, int *no_progress_loops)
3443 * Costly allocations might have made a progress but this doesn't mean
3444 * their order will become available due to high fragmentation so
3445 * always increment the no progress counter for them
3447 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3448 *no_progress_loops = 0;
3450 (*no_progress_loops)++;
3453 * Make sure we converge to OOM if we cannot make any progress
3454 * several times in the row.
3456 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3457 /* Before OOM, exhaust highatomic_reserve */
3458 return unreserve_highatomic_pageblock(ac, true);
3462 * Keep reclaiming pages while there is a chance this will lead
3463 * somewhere. If none of the target zones can satisfy our allocation
3464 * request even if all reclaimable pages are considered then we are
3465 * screwed and have to go OOM.
3467 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3469 unsigned long available;
3470 unsigned long reclaimable;
3472 available = reclaimable = zone_reclaimable_pages(zone);
3473 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3474 MAX_RECLAIM_RETRIES);
3475 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3478 * Would the allocation succeed if we reclaimed the whole
3481 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3482 ac_classzone_idx(ac), alloc_flags, available)) {
3484 * If we didn't make any progress and have a lot of
3485 * dirty + writeback pages then we should wait for
3486 * an IO to complete to slow down the reclaim and
3487 * prevent from pre mature OOM
3489 if (!did_some_progress) {
3490 unsigned long write_pending;
3492 write_pending = zone_page_state_snapshot(zone,
3493 NR_ZONE_WRITE_PENDING);
3495 if (2 * write_pending > reclaimable) {
3496 congestion_wait(BLK_RW_ASYNC, HZ/10);
3502 * Memory allocation/reclaim might be called from a WQ
3503 * context and the current implementation of the WQ
3504 * concurrency control doesn't recognize that
3505 * a particular WQ is congested if the worker thread is
3506 * looping without ever sleeping. Therefore we have to
3507 * do a short sleep here rather than calling
3510 if (current->flags & PF_WQ_WORKER)
3511 schedule_timeout_uninterruptible(1);
3522 static inline struct page *
3523 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3524 struct alloc_context *ac)
3526 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3527 struct page *page = NULL;
3528 unsigned int alloc_flags;
3529 unsigned long did_some_progress;
3530 enum compact_priority compact_priority;
3531 enum compact_result compact_result;
3532 int compaction_retries;
3533 int no_progress_loops;
3534 unsigned long alloc_start = jiffies;
3535 unsigned int stall_timeout = 10 * HZ;
3536 unsigned int cpuset_mems_cookie;
3539 * In the slowpath, we sanity check order to avoid ever trying to
3540 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3541 * be using allocators in order of preference for an area that is
3544 if (order >= MAX_ORDER) {
3545 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3550 * We also sanity check to catch abuse of atomic reserves being used by
3551 * callers that are not in atomic context.
3553 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3554 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3555 gfp_mask &= ~__GFP_ATOMIC;
3558 compaction_retries = 0;
3559 no_progress_loops = 0;
3560 compact_priority = DEF_COMPACT_PRIORITY;
3561 cpuset_mems_cookie = read_mems_allowed_begin();
3563 * We need to recalculate the starting point for the zonelist iterator
3564 * because we might have used different nodemask in the fast path, or
3565 * there was a cpuset modification and we are retrying - otherwise we
3566 * could end up iterating over non-eligible zones endlessly.
3568 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3569 ac->high_zoneidx, ac->nodemask);
3570 if (!ac->preferred_zoneref->zone)
3575 * The fast path uses conservative alloc_flags to succeed only until
3576 * kswapd needs to be woken up, and to avoid the cost of setting up
3577 * alloc_flags precisely. So we do that now.
3579 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3581 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3582 wake_all_kswapds(order, ac);
3585 * The adjusted alloc_flags might result in immediate success, so try
3588 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3593 * For costly allocations, try direct compaction first, as it's likely
3594 * that we have enough base pages and don't need to reclaim. Don't try
3595 * that for allocations that are allowed to ignore watermarks, as the
3596 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3598 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3599 !gfp_pfmemalloc_allowed(gfp_mask)) {
3600 page = __alloc_pages_direct_compact(gfp_mask, order,
3602 INIT_COMPACT_PRIORITY,
3608 * Checks for costly allocations with __GFP_NORETRY, which
3609 * includes THP page fault allocations
3611 if (gfp_mask & __GFP_NORETRY) {
3613 * If compaction is deferred for high-order allocations,
3614 * it is because sync compaction recently failed. If
3615 * this is the case and the caller requested a THP
3616 * allocation, we do not want to heavily disrupt the
3617 * system, so we fail the allocation instead of entering
3620 if (compact_result == COMPACT_DEFERRED)
3624 * Looks like reclaim/compaction is worth trying, but
3625 * sync compaction could be very expensive, so keep
3626 * using async compaction.
3628 compact_priority = INIT_COMPACT_PRIORITY;
3633 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3634 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3635 wake_all_kswapds(order, ac);
3637 if (gfp_pfmemalloc_allowed(gfp_mask))
3638 alloc_flags = ALLOC_NO_WATERMARKS;
3641 * Reset the zonelist iterators if memory policies can be ignored.
3642 * These allocations are high priority and system rather than user
3645 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3646 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3647 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3648 ac->high_zoneidx, ac->nodemask);
3651 /* Attempt with potentially adjusted zonelist and alloc_flags */
3652 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3656 /* Caller is not willing to reclaim, we can't balance anything */
3657 if (!can_direct_reclaim) {
3659 * All existing users of the __GFP_NOFAIL are blockable, so warn
3660 * of any new users that actually allow this type of allocation
3663 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3667 /* Avoid recursion of direct reclaim */
3668 if (current->flags & PF_MEMALLOC) {
3670 * __GFP_NOFAIL request from this context is rather bizarre
3671 * because we cannot reclaim anything and only can loop waiting
3672 * for somebody to do a work for us.
3674 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3681 /* Avoid allocations with no watermarks from looping endlessly */
3682 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3686 /* Try direct reclaim and then allocating */
3687 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3688 &did_some_progress);
3692 /* Try direct compaction and then allocating */
3693 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3694 compact_priority, &compact_result);
3698 /* Do not loop if specifically requested */
3699 if (gfp_mask & __GFP_NORETRY)
3703 * Do not retry costly high order allocations unless they are
3706 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3709 /* Make sure we know about allocations which stall for too long */
3710 if (time_after(jiffies, alloc_start + stall_timeout)) {
3711 warn_alloc(gfp_mask,
3712 "page allocation stalls for %ums, order:%u",
3713 jiffies_to_msecs(jiffies-alloc_start), order);
3714 stall_timeout += 10 * HZ;
3717 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3718 did_some_progress > 0, &no_progress_loops))
3722 * It doesn't make any sense to retry for the compaction if the order-0
3723 * reclaim is not able to make any progress because the current
3724 * implementation of the compaction depends on the sufficient amount
3725 * of free memory (see __compaction_suitable)
3727 if (did_some_progress > 0 &&
3728 should_compact_retry(ac, order, alloc_flags,
3729 compact_result, &compact_priority,
3730 &compaction_retries))
3734 * It's possible we raced with cpuset update so the OOM would be
3735 * premature (see below the nopage: label for full explanation).
3737 if (read_mems_allowed_retry(cpuset_mems_cookie))
3740 /* Reclaim has failed us, start killing things */
3741 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3745 /* Retry as long as the OOM killer is making progress */
3746 if (did_some_progress) {
3747 no_progress_loops = 0;
3753 * When updating a task's mems_allowed or mempolicy nodemask, it is
3754 * possible to race with parallel threads in such a way that our
3755 * allocation can fail while the mask is being updated. If we are about
3756 * to fail, check if the cpuset changed during allocation and if so,
3759 if (read_mems_allowed_retry(cpuset_mems_cookie))
3762 warn_alloc(gfp_mask,
3763 "page allocation failure: order:%u", order);
3769 * This is the 'heart' of the zoned buddy allocator.
3772 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3773 struct zonelist *zonelist, nodemask_t *nodemask)
3776 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3777 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3778 struct alloc_context ac = {
3779 .high_zoneidx = gfp_zone(gfp_mask),
3780 .zonelist = zonelist,
3781 .nodemask = nodemask,
3782 .migratetype = gfpflags_to_migratetype(gfp_mask),
3785 if (cpusets_enabled()) {
3786 alloc_mask |= __GFP_HARDWALL;
3787 alloc_flags |= ALLOC_CPUSET;
3789 ac.nodemask = &cpuset_current_mems_allowed;
3792 gfp_mask &= gfp_allowed_mask;
3794 lockdep_trace_alloc(gfp_mask);
3796 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3798 if (should_fail_alloc_page(gfp_mask, order))
3802 * Check the zones suitable for the gfp_mask contain at least one
3803 * valid zone. It's possible to have an empty zonelist as a result
3804 * of __GFP_THISNODE and a memoryless node
3806 if (unlikely(!zonelist->_zonerefs->zone))
3809 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3810 alloc_flags |= ALLOC_CMA;
3812 /* Dirty zone balancing only done in the fast path */
3813 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3816 * The preferred zone is used for statistics but crucially it is
3817 * also used as the starting point for the zonelist iterator. It
3818 * may get reset for allocations that ignore memory policies.
3820 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3821 ac.high_zoneidx, ac.nodemask);
3822 if (!ac.preferred_zoneref->zone) {
3825 * This might be due to race with cpuset_current_mems_allowed
3826 * update, so make sure we retry with original nodemask in the
3832 /* First allocation attempt */
3833 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3839 * Runtime PM, block IO and its error handling path can deadlock
3840 * because I/O on the device might not complete.
3842 alloc_mask = memalloc_noio_flags(gfp_mask);
3843 ac.spread_dirty_pages = false;
3846 * Restore the original nodemask if it was potentially replaced with
3847 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3849 if (unlikely(ac.nodemask != nodemask))
3850 ac.nodemask = nodemask;
3852 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3855 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3856 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3857 __free_pages(page, order);
3861 if (kmemcheck_enabled && page)
3862 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3864 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3868 EXPORT_SYMBOL(__alloc_pages_nodemask);
3871 * Common helper functions.
3873 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3878 * __get_free_pages() returns a 32-bit address, which cannot represent
3881 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3883 page = alloc_pages(gfp_mask, order);
3886 return (unsigned long) page_address(page);
3888 EXPORT_SYMBOL(__get_free_pages);
3890 unsigned long get_zeroed_page(gfp_t gfp_mask)
3892 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3894 EXPORT_SYMBOL(get_zeroed_page);
3896 void __free_pages(struct page *page, unsigned int order)
3898 if (put_page_testzero(page)) {
3900 free_hot_cold_page(page, false);
3902 __free_pages_ok(page, order);
3906 EXPORT_SYMBOL(__free_pages);
3908 void free_pages(unsigned long addr, unsigned int order)
3911 VM_BUG_ON(!virt_addr_valid((void *)addr));
3912 __free_pages(virt_to_page((void *)addr), order);
3916 EXPORT_SYMBOL(free_pages);
3920 * An arbitrary-length arbitrary-offset area of memory which resides
3921 * within a 0 or higher order page. Multiple fragments within that page
3922 * are individually refcounted, in the page's reference counter.
3924 * The page_frag functions below provide a simple allocation framework for
3925 * page fragments. This is used by the network stack and network device
3926 * drivers to provide a backing region of memory for use as either an
3927 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3929 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
3932 struct page *page = NULL;
3933 gfp_t gfp = gfp_mask;
3935 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3936 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3938 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3939 PAGE_FRAG_CACHE_MAX_ORDER);
3940 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3942 if (unlikely(!page))
3943 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3945 nc->va = page ? page_address(page) : NULL;
3950 void __page_frag_cache_drain(struct page *page, unsigned int count)
3952 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
3954 if (page_ref_sub_and_test(page, count)) {
3955 unsigned int order = compound_order(page);
3958 free_hot_cold_page(page, false);
3960 __free_pages_ok(page, order);
3963 EXPORT_SYMBOL(__page_frag_cache_drain);
3965 void *page_frag_alloc(struct page_frag_cache *nc,
3966 unsigned int fragsz, gfp_t gfp_mask)
3968 unsigned int size = PAGE_SIZE;
3972 if (unlikely(!nc->va)) {
3974 page = __page_frag_cache_refill(nc, gfp_mask);
3978 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3979 /* if size can vary use size else just use PAGE_SIZE */
3982 /* Even if we own the page, we do not use atomic_set().
3983 * This would break get_page_unless_zero() users.
3985 page_ref_add(page, size - 1);
3987 /* reset page count bias and offset to start of new frag */
3988 nc->pfmemalloc = page_is_pfmemalloc(page);
3989 nc->pagecnt_bias = size;
3993 offset = nc->offset - fragsz;
3994 if (unlikely(offset < 0)) {
3995 page = virt_to_page(nc->va);
3997 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4000 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4001 /* if size can vary use size else just use PAGE_SIZE */
4004 /* OK, page count is 0, we can safely set it */
4005 set_page_count(page, size);
4007 /* reset page count bias and offset to start of new frag */
4008 nc->pagecnt_bias = size;
4009 offset = size - fragsz;
4013 nc->offset = offset;
4015 return nc->va + offset;
4017 EXPORT_SYMBOL(page_frag_alloc);
4020 * Frees a page fragment allocated out of either a compound or order 0 page.
4022 void page_frag_free(void *addr)
4024 struct page *page = virt_to_head_page(addr);
4026 if (unlikely(put_page_testzero(page)))
4027 __free_pages_ok(page, compound_order(page));
4029 EXPORT_SYMBOL(page_frag_free);
4031 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4035 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4036 unsigned long used = addr + PAGE_ALIGN(size);
4038 split_page(virt_to_page((void *)addr), order);
4039 while (used < alloc_end) {
4044 return (void *)addr;
4048 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4049 * @size: the number of bytes to allocate
4050 * @gfp_mask: GFP flags for the allocation
4052 * This function is similar to alloc_pages(), except that it allocates the
4053 * minimum number of pages to satisfy the request. alloc_pages() can only
4054 * allocate memory in power-of-two pages.
4056 * This function is also limited by MAX_ORDER.
4058 * Memory allocated by this function must be released by free_pages_exact().
4060 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4062 unsigned int order = get_order(size);
4065 addr = __get_free_pages(gfp_mask, order);
4066 return make_alloc_exact(addr, order, size);
4068 EXPORT_SYMBOL(alloc_pages_exact);
4071 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4073 * @nid: the preferred node ID where memory should be allocated
4074 * @size: the number of bytes to allocate
4075 * @gfp_mask: GFP flags for the allocation
4077 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4080 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4082 unsigned int order = get_order(size);
4083 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4086 return make_alloc_exact((unsigned long)page_address(p), order, size);
4090 * free_pages_exact - release memory allocated via alloc_pages_exact()
4091 * @virt: the value returned by alloc_pages_exact.
4092 * @size: size of allocation, same value as passed to alloc_pages_exact().
4094 * Release the memory allocated by a previous call to alloc_pages_exact.
4096 void free_pages_exact(void *virt, size_t size)
4098 unsigned long addr = (unsigned long)virt;
4099 unsigned long end = addr + PAGE_ALIGN(size);
4101 while (addr < end) {
4106 EXPORT_SYMBOL(free_pages_exact);
4109 * nr_free_zone_pages - count number of pages beyond high watermark
4110 * @offset: The zone index of the highest zone
4112 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4113 * high watermark within all zones at or below a given zone index. For each
4114 * zone, the number of pages is calculated as:
4115 * managed_pages - high_pages
4117 static unsigned long nr_free_zone_pages(int offset)
4122 /* Just pick one node, since fallback list is circular */
4123 unsigned long sum = 0;
4125 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4127 for_each_zone_zonelist(zone, z, zonelist, offset) {
4128 unsigned long size = zone->managed_pages;
4129 unsigned long high = high_wmark_pages(zone);
4138 * nr_free_buffer_pages - count number of pages beyond high watermark
4140 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4141 * watermark within ZONE_DMA and ZONE_NORMAL.
4143 unsigned long nr_free_buffer_pages(void)
4145 return nr_free_zone_pages(gfp_zone(GFP_USER));
4147 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4150 * nr_free_pagecache_pages - count number of pages beyond high watermark
4152 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4153 * high watermark within all zones.
4155 unsigned long nr_free_pagecache_pages(void)
4157 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4160 static inline void show_node(struct zone *zone)
4162 if (IS_ENABLED(CONFIG_NUMA))
4163 printk("Node %d ", zone_to_nid(zone));
4166 long si_mem_available(void)
4169 unsigned long pagecache;
4170 unsigned long wmark_low = 0;
4171 unsigned long pages[NR_LRU_LISTS];
4175 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4176 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4179 wmark_low += zone->watermark[WMARK_LOW];
4182 * Estimate the amount of memory available for userspace allocations,
4183 * without causing swapping.
4185 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4188 * Not all the page cache can be freed, otherwise the system will
4189 * start swapping. Assume at least half of the page cache, or the
4190 * low watermark worth of cache, needs to stay.
4192 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4193 pagecache -= min(pagecache / 2, wmark_low);
4194 available += pagecache;
4197 * Part of the reclaimable slab consists of items that are in use,
4198 * and cannot be freed. Cap this estimate at the low watermark.
4200 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4201 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4207 EXPORT_SYMBOL_GPL(si_mem_available);
4209 void si_meminfo(struct sysinfo *val)
4211 val->totalram = totalram_pages;
4212 val->sharedram = global_node_page_state(NR_SHMEM);
4213 val->freeram = global_page_state(NR_FREE_PAGES);
4214 val->bufferram = nr_blockdev_pages();
4215 val->totalhigh = totalhigh_pages;
4216 val->freehigh = nr_free_highpages();
4217 val->mem_unit = PAGE_SIZE;
4220 EXPORT_SYMBOL(si_meminfo);
4223 void si_meminfo_node(struct sysinfo *val, int nid)
4225 int zone_type; /* needs to be signed */
4226 unsigned long managed_pages = 0;
4227 unsigned long managed_highpages = 0;
4228 unsigned long free_highpages = 0;
4229 pg_data_t *pgdat = NODE_DATA(nid);
4231 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4232 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4233 val->totalram = managed_pages;
4234 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4235 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4236 #ifdef CONFIG_HIGHMEM
4237 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4238 struct zone *zone = &pgdat->node_zones[zone_type];
4240 if (is_highmem(zone)) {
4241 managed_highpages += zone->managed_pages;
4242 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4245 val->totalhigh = managed_highpages;
4246 val->freehigh = free_highpages;
4248 val->totalhigh = managed_highpages;
4249 val->freehigh = free_highpages;
4251 val->mem_unit = PAGE_SIZE;
4256 * Determine whether the node should be displayed or not, depending on whether
4257 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4259 bool skip_free_areas_node(unsigned int flags, int nid)
4262 unsigned int cpuset_mems_cookie;
4264 if (!(flags & SHOW_MEM_FILTER_NODES))
4268 cpuset_mems_cookie = read_mems_allowed_begin();
4269 ret = !node_isset(nid, cpuset_current_mems_allowed);
4270 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4275 #define K(x) ((x) << (PAGE_SHIFT-10))
4277 static void show_migration_types(unsigned char type)
4279 static const char types[MIGRATE_TYPES] = {
4280 [MIGRATE_UNMOVABLE] = 'U',
4281 [MIGRATE_MOVABLE] = 'M',
4282 [MIGRATE_RECLAIMABLE] = 'E',
4283 [MIGRATE_HIGHATOMIC] = 'H',
4285 [MIGRATE_CMA] = 'C',
4287 #ifdef CONFIG_MEMORY_ISOLATION
4288 [MIGRATE_ISOLATE] = 'I',
4291 char tmp[MIGRATE_TYPES + 1];
4295 for (i = 0; i < MIGRATE_TYPES; i++) {
4296 if (type & (1 << i))
4301 printk(KERN_CONT "(%s) ", tmp);
4305 * Show free area list (used inside shift_scroll-lock stuff)
4306 * We also calculate the percentage fragmentation. We do this by counting the
4307 * memory on each free list with the exception of the first item on the list.
4310 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4313 void show_free_areas(unsigned int filter)
4315 unsigned long free_pcp = 0;
4320 for_each_populated_zone(zone) {
4321 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4324 for_each_online_cpu(cpu)
4325 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4328 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4329 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4330 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4331 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4332 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4333 " free:%lu free_pcp:%lu free_cma:%lu\n",
4334 global_node_page_state(NR_ACTIVE_ANON),
4335 global_node_page_state(NR_INACTIVE_ANON),
4336 global_node_page_state(NR_ISOLATED_ANON),
4337 global_node_page_state(NR_ACTIVE_FILE),
4338 global_node_page_state(NR_INACTIVE_FILE),
4339 global_node_page_state(NR_ISOLATED_FILE),
4340 global_node_page_state(NR_UNEVICTABLE),
4341 global_node_page_state(NR_FILE_DIRTY),
4342 global_node_page_state(NR_WRITEBACK),
4343 global_node_page_state(NR_UNSTABLE_NFS),
4344 global_page_state(NR_SLAB_RECLAIMABLE),
4345 global_page_state(NR_SLAB_UNRECLAIMABLE),
4346 global_node_page_state(NR_FILE_MAPPED),
4347 global_node_page_state(NR_SHMEM),
4348 global_page_state(NR_PAGETABLE),
4349 global_page_state(NR_BOUNCE),
4350 global_page_state(NR_FREE_PAGES),
4352 global_page_state(NR_FREE_CMA_PAGES));
4354 for_each_online_pgdat(pgdat) {
4356 " active_anon:%lukB"
4357 " inactive_anon:%lukB"
4358 " active_file:%lukB"
4359 " inactive_file:%lukB"
4360 " unevictable:%lukB"
4361 " isolated(anon):%lukB"
4362 " isolated(file):%lukB"
4367 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4369 " shmem_pmdmapped: %lukB"
4372 " writeback_tmp:%lukB"
4374 " pages_scanned:%lu"
4375 " all_unreclaimable? %s"
4378 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4379 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4380 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4381 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4382 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4383 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4384 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4385 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4386 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4387 K(node_page_state(pgdat, NR_WRITEBACK)),
4388 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4389 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4390 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4392 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4394 K(node_page_state(pgdat, NR_SHMEM)),
4395 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4396 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4397 node_page_state(pgdat, NR_PAGES_SCANNED),
4398 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4401 for_each_populated_zone(zone) {
4404 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4408 for_each_online_cpu(cpu)
4409 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4418 " active_anon:%lukB"
4419 " inactive_anon:%lukB"
4420 " active_file:%lukB"
4421 " inactive_file:%lukB"
4422 " unevictable:%lukB"
4423 " writepending:%lukB"
4427 " slab_reclaimable:%lukB"
4428 " slab_unreclaimable:%lukB"
4429 " kernel_stack:%lukB"
4437 K(zone_page_state(zone, NR_FREE_PAGES)),
4438 K(min_wmark_pages(zone)),
4439 K(low_wmark_pages(zone)),
4440 K(high_wmark_pages(zone)),
4441 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4442 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4443 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4444 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4445 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4446 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4447 K(zone->present_pages),
4448 K(zone->managed_pages),
4449 K(zone_page_state(zone, NR_MLOCK)),
4450 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4451 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4452 zone_page_state(zone, NR_KERNEL_STACK_KB),
4453 K(zone_page_state(zone, NR_PAGETABLE)),
4454 K(zone_page_state(zone, NR_BOUNCE)),
4456 K(this_cpu_read(zone->pageset->pcp.count)),
4457 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4458 printk("lowmem_reserve[]:");
4459 for (i = 0; i < MAX_NR_ZONES; i++)
4460 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4461 printk(KERN_CONT "\n");
4464 for_each_populated_zone(zone) {
4466 unsigned long nr[MAX_ORDER], flags, total = 0;
4467 unsigned char types[MAX_ORDER];
4469 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4472 printk(KERN_CONT "%s: ", zone->name);
4474 spin_lock_irqsave(&zone->lock, flags);
4475 for (order = 0; order < MAX_ORDER; order++) {
4476 struct free_area *area = &zone->free_area[order];
4479 nr[order] = area->nr_free;
4480 total += nr[order] << order;
4483 for (type = 0; type < MIGRATE_TYPES; type++) {
4484 if (!list_empty(&area->free_list[type]))
4485 types[order] |= 1 << type;
4488 spin_unlock_irqrestore(&zone->lock, flags);
4489 for (order = 0; order < MAX_ORDER; order++) {
4490 printk(KERN_CONT "%lu*%lukB ",
4491 nr[order], K(1UL) << order);
4493 show_migration_types(types[order]);
4495 printk(KERN_CONT "= %lukB\n", K(total));
4498 hugetlb_show_meminfo();
4500 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4502 show_swap_cache_info();
4505 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4507 zoneref->zone = zone;
4508 zoneref->zone_idx = zone_idx(zone);
4512 * Builds allocation fallback zone lists.
4514 * Add all populated zones of a node to the zonelist.
4516 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4520 enum zone_type zone_type = MAX_NR_ZONES;
4524 zone = pgdat->node_zones + zone_type;
4525 if (managed_zone(zone)) {
4526 zoneref_set_zone(zone,
4527 &zonelist->_zonerefs[nr_zones++]);
4528 check_highest_zone(zone_type);
4530 } while (zone_type);
4538 * 0 = automatic detection of better ordering.
4539 * 1 = order by ([node] distance, -zonetype)
4540 * 2 = order by (-zonetype, [node] distance)
4542 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4543 * the same zonelist. So only NUMA can configure this param.
4545 #define ZONELIST_ORDER_DEFAULT 0
4546 #define ZONELIST_ORDER_NODE 1
4547 #define ZONELIST_ORDER_ZONE 2
4549 /* zonelist order in the kernel.
4550 * set_zonelist_order() will set this to NODE or ZONE.
4552 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4553 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4557 /* The value user specified ....changed by config */
4558 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4559 /* string for sysctl */
4560 #define NUMA_ZONELIST_ORDER_LEN 16
4561 char numa_zonelist_order[16] = "default";
4564 * interface for configure zonelist ordering.
4565 * command line option "numa_zonelist_order"
4566 * = "[dD]efault - default, automatic configuration.
4567 * = "[nN]ode - order by node locality, then by zone within node
4568 * = "[zZ]one - order by zone, then by locality within zone
4571 static int __parse_numa_zonelist_order(char *s)
4573 if (*s == 'd' || *s == 'D') {
4574 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4575 } else if (*s == 'n' || *s == 'N') {
4576 user_zonelist_order = ZONELIST_ORDER_NODE;
4577 } else if (*s == 'z' || *s == 'Z') {
4578 user_zonelist_order = ZONELIST_ORDER_ZONE;
4580 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4586 static __init int setup_numa_zonelist_order(char *s)
4593 ret = __parse_numa_zonelist_order(s);
4595 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4599 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4602 * sysctl handler for numa_zonelist_order
4604 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4605 void __user *buffer, size_t *length,
4608 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4610 static DEFINE_MUTEX(zl_order_mutex);
4612 mutex_lock(&zl_order_mutex);
4614 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4618 strcpy(saved_string, (char *)table->data);
4620 ret = proc_dostring(table, write, buffer, length, ppos);
4624 int oldval = user_zonelist_order;
4626 ret = __parse_numa_zonelist_order((char *)table->data);
4629 * bogus value. restore saved string
4631 strncpy((char *)table->data, saved_string,
4632 NUMA_ZONELIST_ORDER_LEN);
4633 user_zonelist_order = oldval;
4634 } else if (oldval != user_zonelist_order) {
4635 mutex_lock(&zonelists_mutex);
4636 build_all_zonelists(NULL, NULL);
4637 mutex_unlock(&zonelists_mutex);
4641 mutex_unlock(&zl_order_mutex);
4646 #define MAX_NODE_LOAD (nr_online_nodes)
4647 static int node_load[MAX_NUMNODES];
4650 * find_next_best_node - find the next node that should appear in a given node's fallback list
4651 * @node: node whose fallback list we're appending
4652 * @used_node_mask: nodemask_t of already used nodes
4654 * We use a number of factors to determine which is the next node that should
4655 * appear on a given node's fallback list. The node should not have appeared
4656 * already in @node's fallback list, and it should be the next closest node
4657 * according to the distance array (which contains arbitrary distance values
4658 * from each node to each node in the system), and should also prefer nodes
4659 * with no CPUs, since presumably they'll have very little allocation pressure
4660 * on them otherwise.
4661 * It returns -1 if no node is found.
4663 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4666 int min_val = INT_MAX;
4667 int best_node = NUMA_NO_NODE;
4668 const struct cpumask *tmp = cpumask_of_node(0);
4670 /* Use the local node if we haven't already */
4671 if (!node_isset(node, *used_node_mask)) {
4672 node_set(node, *used_node_mask);
4676 for_each_node_state(n, N_MEMORY) {
4678 /* Don't want a node to appear more than once */
4679 if (node_isset(n, *used_node_mask))
4682 /* Use the distance array to find the distance */
4683 val = node_distance(node, n);
4685 /* Penalize nodes under us ("prefer the next node") */
4688 /* Give preference to headless and unused nodes */
4689 tmp = cpumask_of_node(n);
4690 if (!cpumask_empty(tmp))
4691 val += PENALTY_FOR_NODE_WITH_CPUS;
4693 /* Slight preference for less loaded node */
4694 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4695 val += node_load[n];
4697 if (val < min_val) {
4704 node_set(best_node, *used_node_mask);
4711 * Build zonelists ordered by node and zones within node.
4712 * This results in maximum locality--normal zone overflows into local
4713 * DMA zone, if any--but risks exhausting DMA zone.
4715 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4718 struct zonelist *zonelist;
4720 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4721 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4723 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4724 zonelist->_zonerefs[j].zone = NULL;
4725 zonelist->_zonerefs[j].zone_idx = 0;
4729 * Build gfp_thisnode zonelists
4731 static void build_thisnode_zonelists(pg_data_t *pgdat)
4734 struct zonelist *zonelist;
4736 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4737 j = build_zonelists_node(pgdat, zonelist, 0);
4738 zonelist->_zonerefs[j].zone = NULL;
4739 zonelist->_zonerefs[j].zone_idx = 0;
4743 * Build zonelists ordered by zone and nodes within zones.
4744 * This results in conserving DMA zone[s] until all Normal memory is
4745 * exhausted, but results in overflowing to remote node while memory
4746 * may still exist in local DMA zone.
4748 static int node_order[MAX_NUMNODES];
4750 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4753 int zone_type; /* needs to be signed */
4755 struct zonelist *zonelist;
4757 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4759 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4760 for (j = 0; j < nr_nodes; j++) {
4761 node = node_order[j];
4762 z = &NODE_DATA(node)->node_zones[zone_type];
4763 if (managed_zone(z)) {
4765 &zonelist->_zonerefs[pos++]);
4766 check_highest_zone(zone_type);
4770 zonelist->_zonerefs[pos].zone = NULL;
4771 zonelist->_zonerefs[pos].zone_idx = 0;
4774 #if defined(CONFIG_64BIT)
4776 * Devices that require DMA32/DMA are relatively rare and do not justify a
4777 * penalty to every machine in case the specialised case applies. Default
4778 * to Node-ordering on 64-bit NUMA machines
4780 static int default_zonelist_order(void)
4782 return ZONELIST_ORDER_NODE;
4786 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4787 * by the kernel. If processes running on node 0 deplete the low memory zone
4788 * then reclaim will occur more frequency increasing stalls and potentially
4789 * be easier to OOM if a large percentage of the zone is under writeback or
4790 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4791 * Hence, default to zone ordering on 32-bit.
4793 static int default_zonelist_order(void)
4795 return ZONELIST_ORDER_ZONE;
4797 #endif /* CONFIG_64BIT */
4799 static void set_zonelist_order(void)
4801 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4802 current_zonelist_order = default_zonelist_order();
4804 current_zonelist_order = user_zonelist_order;
4807 static void build_zonelists(pg_data_t *pgdat)
4810 nodemask_t used_mask;
4811 int local_node, prev_node;
4812 struct zonelist *zonelist;
4813 unsigned int order = current_zonelist_order;
4815 /* initialize zonelists */
4816 for (i = 0; i < MAX_ZONELISTS; i++) {
4817 zonelist = pgdat->node_zonelists + i;
4818 zonelist->_zonerefs[0].zone = NULL;
4819 zonelist->_zonerefs[0].zone_idx = 0;
4822 /* NUMA-aware ordering of nodes */
4823 local_node = pgdat->node_id;
4824 load = nr_online_nodes;
4825 prev_node = local_node;
4826 nodes_clear(used_mask);
4828 memset(node_order, 0, sizeof(node_order));
4831 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4833 * We don't want to pressure a particular node.
4834 * So adding penalty to the first node in same
4835 * distance group to make it round-robin.
4837 if (node_distance(local_node, node) !=
4838 node_distance(local_node, prev_node))
4839 node_load[node] = load;
4843 if (order == ZONELIST_ORDER_NODE)
4844 build_zonelists_in_node_order(pgdat, node);
4846 node_order[i++] = node; /* remember order */
4849 if (order == ZONELIST_ORDER_ZONE) {
4850 /* calculate node order -- i.e., DMA last! */
4851 build_zonelists_in_zone_order(pgdat, i);
4854 build_thisnode_zonelists(pgdat);
4857 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4859 * Return node id of node used for "local" allocations.
4860 * I.e., first node id of first zone in arg node's generic zonelist.
4861 * Used for initializing percpu 'numa_mem', which is used primarily
4862 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4864 int local_memory_node(int node)
4868 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4869 gfp_zone(GFP_KERNEL),
4871 return z->zone->node;
4875 static void setup_min_unmapped_ratio(void);
4876 static void setup_min_slab_ratio(void);
4877 #else /* CONFIG_NUMA */
4879 static void set_zonelist_order(void)
4881 current_zonelist_order = ZONELIST_ORDER_ZONE;
4884 static void build_zonelists(pg_data_t *pgdat)
4886 int node, local_node;
4888 struct zonelist *zonelist;
4890 local_node = pgdat->node_id;
4892 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4893 j = build_zonelists_node(pgdat, zonelist, 0);
4896 * Now we build the zonelist so that it contains the zones
4897 * of all the other nodes.
4898 * We don't want to pressure a particular node, so when
4899 * building the zones for node N, we make sure that the
4900 * zones coming right after the local ones are those from
4901 * node N+1 (modulo N)
4903 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4904 if (!node_online(node))
4906 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4908 for (node = 0; node < local_node; node++) {
4909 if (!node_online(node))
4911 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4914 zonelist->_zonerefs[j].zone = NULL;
4915 zonelist->_zonerefs[j].zone_idx = 0;
4918 #endif /* CONFIG_NUMA */
4921 * Boot pageset table. One per cpu which is going to be used for all
4922 * zones and all nodes. The parameters will be set in such a way
4923 * that an item put on a list will immediately be handed over to
4924 * the buddy list. This is safe since pageset manipulation is done
4925 * with interrupts disabled.
4927 * The boot_pagesets must be kept even after bootup is complete for
4928 * unused processors and/or zones. They do play a role for bootstrapping
4929 * hotplugged processors.
4931 * zoneinfo_show() and maybe other functions do
4932 * not check if the processor is online before following the pageset pointer.
4933 * Other parts of the kernel may not check if the zone is available.
4935 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4936 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4937 static void setup_zone_pageset(struct zone *zone);
4940 * Global mutex to protect against size modification of zonelists
4941 * as well as to serialize pageset setup for the new populated zone.
4943 DEFINE_MUTEX(zonelists_mutex);
4945 /* return values int ....just for stop_machine() */
4946 static int __build_all_zonelists(void *data)
4950 pg_data_t *self = data;
4953 memset(node_load, 0, sizeof(node_load));
4956 if (self && !node_online(self->node_id)) {
4957 build_zonelists(self);
4960 for_each_online_node(nid) {
4961 pg_data_t *pgdat = NODE_DATA(nid);
4963 build_zonelists(pgdat);
4967 * Initialize the boot_pagesets that are going to be used
4968 * for bootstrapping processors. The real pagesets for
4969 * each zone will be allocated later when the per cpu
4970 * allocator is available.
4972 * boot_pagesets are used also for bootstrapping offline
4973 * cpus if the system is already booted because the pagesets
4974 * are needed to initialize allocators on a specific cpu too.
4975 * F.e. the percpu allocator needs the page allocator which
4976 * needs the percpu allocator in order to allocate its pagesets
4977 * (a chicken-egg dilemma).
4979 for_each_possible_cpu(cpu) {
4980 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4982 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4984 * We now know the "local memory node" for each node--
4985 * i.e., the node of the first zone in the generic zonelist.
4986 * Set up numa_mem percpu variable for on-line cpus. During
4987 * boot, only the boot cpu should be on-line; we'll init the
4988 * secondary cpus' numa_mem as they come on-line. During
4989 * node/memory hotplug, we'll fixup all on-line cpus.
4991 if (cpu_online(cpu))
4992 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4999 static noinline void __init
5000 build_all_zonelists_init(void)
5002 __build_all_zonelists(NULL);
5003 mminit_verify_zonelist();
5004 cpuset_init_current_mems_allowed();
5008 * Called with zonelists_mutex held always
5009 * unless system_state == SYSTEM_BOOTING.
5011 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5012 * [we're only called with non-NULL zone through __meminit paths] and
5013 * (2) call of __init annotated helper build_all_zonelists_init
5014 * [protected by SYSTEM_BOOTING].
5016 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5018 set_zonelist_order();
5020 if (system_state == SYSTEM_BOOTING) {
5021 build_all_zonelists_init();
5023 #ifdef CONFIG_MEMORY_HOTPLUG
5025 setup_zone_pageset(zone);
5027 /* we have to stop all cpus to guarantee there is no user
5029 stop_machine(__build_all_zonelists, pgdat, NULL);
5030 /* cpuset refresh routine should be here */
5032 vm_total_pages = nr_free_pagecache_pages();
5034 * Disable grouping by mobility if the number of pages in the
5035 * system is too low to allow the mechanism to work. It would be
5036 * more accurate, but expensive to check per-zone. This check is
5037 * made on memory-hotadd so a system can start with mobility
5038 * disabled and enable it later
5040 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5041 page_group_by_mobility_disabled = 1;
5043 page_group_by_mobility_disabled = 0;
5045 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5047 zonelist_order_name[current_zonelist_order],
5048 page_group_by_mobility_disabled ? "off" : "on",
5051 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5056 * Initially all pages are reserved - free ones are freed
5057 * up by free_all_bootmem() once the early boot process is
5058 * done. Non-atomic initialization, single-pass.
5060 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5061 unsigned long start_pfn, enum memmap_context context)
5063 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5064 unsigned long end_pfn = start_pfn + size;
5065 pg_data_t *pgdat = NODE_DATA(nid);
5067 unsigned long nr_initialised = 0;
5068 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5069 struct memblock_region *r = NULL, *tmp;
5072 if (highest_memmap_pfn < end_pfn - 1)
5073 highest_memmap_pfn = end_pfn - 1;
5076 * Honor reservation requested by the driver for this ZONE_DEVICE
5079 if (altmap && start_pfn == altmap->base_pfn)
5080 start_pfn += altmap->reserve;
5082 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5084 * There can be holes in boot-time mem_map[]s handed to this
5085 * function. They do not exist on hotplugged memory.
5087 if (context != MEMMAP_EARLY)
5090 if (!early_pfn_valid(pfn))
5092 if (!early_pfn_in_nid(pfn, nid))
5094 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5097 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5099 * Check given memblock attribute by firmware which can affect
5100 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5101 * mirrored, it's an overlapped memmap init. skip it.
5103 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5104 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5105 for_each_memblock(memory, tmp)
5106 if (pfn < memblock_region_memory_end_pfn(tmp))
5110 if (pfn >= memblock_region_memory_base_pfn(r) &&
5111 memblock_is_mirror(r)) {
5112 /* already initialized as NORMAL */
5113 pfn = memblock_region_memory_end_pfn(r);
5121 * Mark the block movable so that blocks are reserved for
5122 * movable at startup. This will force kernel allocations
5123 * to reserve their blocks rather than leaking throughout
5124 * the address space during boot when many long-lived
5125 * kernel allocations are made.
5127 * bitmap is created for zone's valid pfn range. but memmap
5128 * can be created for invalid pages (for alignment)
5129 * check here not to call set_pageblock_migratetype() against
5132 if (!(pfn & (pageblock_nr_pages - 1))) {
5133 struct page *page = pfn_to_page(pfn);
5135 __init_single_page(page, pfn, zone, nid);
5136 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5138 __init_single_pfn(pfn, zone, nid);
5143 static void __meminit zone_init_free_lists(struct zone *zone)
5145 unsigned int order, t;
5146 for_each_migratetype_order(order, t) {
5147 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5148 zone->free_area[order].nr_free = 0;
5152 #ifndef __HAVE_ARCH_MEMMAP_INIT
5153 #define memmap_init(size, nid, zone, start_pfn) \
5154 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5157 static int zone_batchsize(struct zone *zone)
5163 * The per-cpu-pages pools are set to around 1000th of the
5164 * size of the zone. But no more than 1/2 of a meg.
5166 * OK, so we don't know how big the cache is. So guess.
5168 batch = zone->managed_pages / 1024;
5169 if (batch * PAGE_SIZE > 512 * 1024)
5170 batch = (512 * 1024) / PAGE_SIZE;
5171 batch /= 4; /* We effectively *= 4 below */
5176 * Clamp the batch to a 2^n - 1 value. Having a power
5177 * of 2 value was found to be more likely to have
5178 * suboptimal cache aliasing properties in some cases.
5180 * For example if 2 tasks are alternately allocating
5181 * batches of pages, one task can end up with a lot
5182 * of pages of one half of the possible page colors
5183 * and the other with pages of the other colors.
5185 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5190 /* The deferral and batching of frees should be suppressed under NOMMU
5193 * The problem is that NOMMU needs to be able to allocate large chunks
5194 * of contiguous memory as there's no hardware page translation to
5195 * assemble apparent contiguous memory from discontiguous pages.
5197 * Queueing large contiguous runs of pages for batching, however,
5198 * causes the pages to actually be freed in smaller chunks. As there
5199 * can be a significant delay between the individual batches being
5200 * recycled, this leads to the once large chunks of space being
5201 * fragmented and becoming unavailable for high-order allocations.
5208 * pcp->high and pcp->batch values are related and dependent on one another:
5209 * ->batch must never be higher then ->high.
5210 * The following function updates them in a safe manner without read side
5213 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5214 * those fields changing asynchronously (acording the the above rule).
5216 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5217 * outside of boot time (or some other assurance that no concurrent updaters
5220 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5221 unsigned long batch)
5223 /* start with a fail safe value for batch */
5227 /* Update high, then batch, in order */
5234 /* a companion to pageset_set_high() */
5235 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5237 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5240 static void pageset_init(struct per_cpu_pageset *p)
5242 struct per_cpu_pages *pcp;
5245 memset(p, 0, sizeof(*p));
5249 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5250 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5253 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5256 pageset_set_batch(p, batch);
5260 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5261 * to the value high for the pageset p.
5263 static void pageset_set_high(struct per_cpu_pageset *p,
5266 unsigned long batch = max(1UL, high / 4);
5267 if ((high / 4) > (PAGE_SHIFT * 8))
5268 batch = PAGE_SHIFT * 8;
5270 pageset_update(&p->pcp, high, batch);
5273 static void pageset_set_high_and_batch(struct zone *zone,
5274 struct per_cpu_pageset *pcp)
5276 if (percpu_pagelist_fraction)
5277 pageset_set_high(pcp,
5278 (zone->managed_pages /
5279 percpu_pagelist_fraction));
5281 pageset_set_batch(pcp, zone_batchsize(zone));
5284 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5286 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5289 pageset_set_high_and_batch(zone, pcp);
5292 static void __meminit setup_zone_pageset(struct zone *zone)
5295 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5296 for_each_possible_cpu(cpu)
5297 zone_pageset_init(zone, cpu);
5301 * Allocate per cpu pagesets and initialize them.
5302 * Before this call only boot pagesets were available.
5304 void __init setup_per_cpu_pageset(void)
5306 struct pglist_data *pgdat;
5309 for_each_populated_zone(zone)
5310 setup_zone_pageset(zone);
5312 for_each_online_pgdat(pgdat)
5313 pgdat->per_cpu_nodestats =
5314 alloc_percpu(struct per_cpu_nodestat);
5317 static __meminit void zone_pcp_init(struct zone *zone)
5320 * per cpu subsystem is not up at this point. The following code
5321 * relies on the ability of the linker to provide the
5322 * offset of a (static) per cpu variable into the per cpu area.
5324 zone->pageset = &boot_pageset;
5326 if (populated_zone(zone))
5327 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5328 zone->name, zone->present_pages,
5329 zone_batchsize(zone));
5332 int __meminit init_currently_empty_zone(struct zone *zone,
5333 unsigned long zone_start_pfn,
5336 struct pglist_data *pgdat = zone->zone_pgdat;
5338 pgdat->nr_zones = zone_idx(zone) + 1;
5340 zone->zone_start_pfn = zone_start_pfn;
5342 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5343 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5345 (unsigned long)zone_idx(zone),
5346 zone_start_pfn, (zone_start_pfn + size));
5348 zone_init_free_lists(zone);
5349 zone->initialized = 1;
5354 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5355 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5358 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5360 int __meminit __early_pfn_to_nid(unsigned long pfn,
5361 struct mminit_pfnnid_cache *state)
5363 unsigned long start_pfn, end_pfn;
5366 if (state->last_start <= pfn && pfn < state->last_end)
5367 return state->last_nid;
5369 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5371 state->last_start = start_pfn;
5372 state->last_end = end_pfn;
5373 state->last_nid = nid;
5378 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5381 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5382 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5383 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5385 * If an architecture guarantees that all ranges registered contain no holes
5386 * and may be freed, this this function may be used instead of calling
5387 * memblock_free_early_nid() manually.
5389 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5391 unsigned long start_pfn, end_pfn;
5394 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5395 start_pfn = min(start_pfn, max_low_pfn);
5396 end_pfn = min(end_pfn, max_low_pfn);
5398 if (start_pfn < end_pfn)
5399 memblock_free_early_nid(PFN_PHYS(start_pfn),
5400 (end_pfn - start_pfn) << PAGE_SHIFT,
5406 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5407 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5409 * If an architecture guarantees that all ranges registered contain no holes and may
5410 * be freed, this function may be used instead of calling memory_present() manually.
5412 void __init sparse_memory_present_with_active_regions(int nid)
5414 unsigned long start_pfn, end_pfn;
5417 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5418 memory_present(this_nid, start_pfn, end_pfn);
5422 * get_pfn_range_for_nid - Return the start and end page frames for a node
5423 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5424 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5425 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5427 * It returns the start and end page frame of a node based on information
5428 * provided by memblock_set_node(). If called for a node
5429 * with no available memory, a warning is printed and the start and end
5432 void __meminit get_pfn_range_for_nid(unsigned int nid,
5433 unsigned long *start_pfn, unsigned long *end_pfn)
5435 unsigned long this_start_pfn, this_end_pfn;
5441 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5442 *start_pfn = min(*start_pfn, this_start_pfn);
5443 *end_pfn = max(*end_pfn, this_end_pfn);
5446 if (*start_pfn == -1UL)
5451 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5452 * assumption is made that zones within a node are ordered in monotonic
5453 * increasing memory addresses so that the "highest" populated zone is used
5455 static void __init find_usable_zone_for_movable(void)
5458 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5459 if (zone_index == ZONE_MOVABLE)
5462 if (arch_zone_highest_possible_pfn[zone_index] >
5463 arch_zone_lowest_possible_pfn[zone_index])
5467 VM_BUG_ON(zone_index == -1);
5468 movable_zone = zone_index;
5472 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5473 * because it is sized independent of architecture. Unlike the other zones,
5474 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5475 * in each node depending on the size of each node and how evenly kernelcore
5476 * is distributed. This helper function adjusts the zone ranges
5477 * provided by the architecture for a given node by using the end of the
5478 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5479 * zones within a node are in order of monotonic increases memory addresses
5481 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5482 unsigned long zone_type,
5483 unsigned long node_start_pfn,
5484 unsigned long node_end_pfn,
5485 unsigned long *zone_start_pfn,
5486 unsigned long *zone_end_pfn)
5488 /* Only adjust if ZONE_MOVABLE is on this node */
5489 if (zone_movable_pfn[nid]) {
5490 /* Size ZONE_MOVABLE */
5491 if (zone_type == ZONE_MOVABLE) {
5492 *zone_start_pfn = zone_movable_pfn[nid];
5493 *zone_end_pfn = min(node_end_pfn,
5494 arch_zone_highest_possible_pfn[movable_zone]);
5496 /* Adjust for ZONE_MOVABLE starting within this range */
5497 } else if (!mirrored_kernelcore &&
5498 *zone_start_pfn < zone_movable_pfn[nid] &&
5499 *zone_end_pfn > zone_movable_pfn[nid]) {
5500 *zone_end_pfn = zone_movable_pfn[nid];
5502 /* Check if this whole range is within ZONE_MOVABLE */
5503 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5504 *zone_start_pfn = *zone_end_pfn;
5509 * Return the number of pages a zone spans in a node, including holes
5510 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5512 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5513 unsigned long zone_type,
5514 unsigned long node_start_pfn,
5515 unsigned long node_end_pfn,
5516 unsigned long *zone_start_pfn,
5517 unsigned long *zone_end_pfn,
5518 unsigned long *ignored)
5520 /* When hotadd a new node from cpu_up(), the node should be empty */
5521 if (!node_start_pfn && !node_end_pfn)
5524 /* Get the start and end of the zone */
5525 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5526 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5527 adjust_zone_range_for_zone_movable(nid, zone_type,
5528 node_start_pfn, node_end_pfn,
5529 zone_start_pfn, zone_end_pfn);
5531 /* Check that this node has pages within the zone's required range */
5532 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5535 /* Move the zone boundaries inside the node if necessary */
5536 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5537 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5539 /* Return the spanned pages */
5540 return *zone_end_pfn - *zone_start_pfn;
5544 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5545 * then all holes in the requested range will be accounted for.
5547 unsigned long __meminit __absent_pages_in_range(int nid,
5548 unsigned long range_start_pfn,
5549 unsigned long range_end_pfn)
5551 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5552 unsigned long start_pfn, end_pfn;
5555 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5556 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5557 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5558 nr_absent -= end_pfn - start_pfn;
5564 * absent_pages_in_range - Return number of page frames in holes within a range
5565 * @start_pfn: The start PFN to start searching for holes
5566 * @end_pfn: The end PFN to stop searching for holes
5568 * It returns the number of pages frames in memory holes within a range.
5570 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5571 unsigned long end_pfn)
5573 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5576 /* Return the number of page frames in holes in a zone on a node */
5577 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5578 unsigned long zone_type,
5579 unsigned long node_start_pfn,
5580 unsigned long node_end_pfn,
5581 unsigned long *ignored)
5583 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5584 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5585 unsigned long zone_start_pfn, zone_end_pfn;
5586 unsigned long nr_absent;
5588 /* When hotadd a new node from cpu_up(), the node should be empty */
5589 if (!node_start_pfn && !node_end_pfn)
5592 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5593 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5595 adjust_zone_range_for_zone_movable(nid, zone_type,
5596 node_start_pfn, node_end_pfn,
5597 &zone_start_pfn, &zone_end_pfn);
5598 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5601 * ZONE_MOVABLE handling.
5602 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5605 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5606 unsigned long start_pfn, end_pfn;
5607 struct memblock_region *r;
5609 for_each_memblock(memory, r) {
5610 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5611 zone_start_pfn, zone_end_pfn);
5612 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5613 zone_start_pfn, zone_end_pfn);
5615 if (zone_type == ZONE_MOVABLE &&
5616 memblock_is_mirror(r))
5617 nr_absent += end_pfn - start_pfn;
5619 if (zone_type == ZONE_NORMAL &&
5620 !memblock_is_mirror(r))
5621 nr_absent += end_pfn - start_pfn;
5628 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5629 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5630 unsigned long zone_type,
5631 unsigned long node_start_pfn,
5632 unsigned long node_end_pfn,
5633 unsigned long *zone_start_pfn,
5634 unsigned long *zone_end_pfn,
5635 unsigned long *zones_size)
5639 *zone_start_pfn = node_start_pfn;
5640 for (zone = 0; zone < zone_type; zone++)
5641 *zone_start_pfn += zones_size[zone];
5643 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5645 return zones_size[zone_type];
5648 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5649 unsigned long zone_type,
5650 unsigned long node_start_pfn,
5651 unsigned long node_end_pfn,
5652 unsigned long *zholes_size)
5657 return zholes_size[zone_type];
5660 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5662 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5663 unsigned long node_start_pfn,
5664 unsigned long node_end_pfn,
5665 unsigned long *zones_size,
5666 unsigned long *zholes_size)
5668 unsigned long realtotalpages = 0, totalpages = 0;
5671 for (i = 0; i < MAX_NR_ZONES; i++) {
5672 struct zone *zone = pgdat->node_zones + i;
5673 unsigned long zone_start_pfn, zone_end_pfn;
5674 unsigned long size, real_size;
5676 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5682 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5683 node_start_pfn, node_end_pfn,
5686 zone->zone_start_pfn = zone_start_pfn;
5688 zone->zone_start_pfn = 0;
5689 zone->spanned_pages = size;
5690 zone->present_pages = real_size;
5693 realtotalpages += real_size;
5696 pgdat->node_spanned_pages = totalpages;
5697 pgdat->node_present_pages = realtotalpages;
5698 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5702 #ifndef CONFIG_SPARSEMEM
5704 * Calculate the size of the zone->blockflags rounded to an unsigned long
5705 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5706 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5707 * round what is now in bits to nearest long in bits, then return it in
5710 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5712 unsigned long usemapsize;
5714 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5715 usemapsize = roundup(zonesize, pageblock_nr_pages);
5716 usemapsize = usemapsize >> pageblock_order;
5717 usemapsize *= NR_PAGEBLOCK_BITS;
5718 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5720 return usemapsize / 8;
5723 static void __init setup_usemap(struct pglist_data *pgdat,
5725 unsigned long zone_start_pfn,
5726 unsigned long zonesize)
5728 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5729 zone->pageblock_flags = NULL;
5731 zone->pageblock_flags =
5732 memblock_virt_alloc_node_nopanic(usemapsize,
5736 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5737 unsigned long zone_start_pfn, unsigned long zonesize) {}
5738 #endif /* CONFIG_SPARSEMEM */
5740 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5742 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5743 void __paginginit set_pageblock_order(void)
5747 /* Check that pageblock_nr_pages has not already been setup */
5748 if (pageblock_order)
5751 if (HPAGE_SHIFT > PAGE_SHIFT)
5752 order = HUGETLB_PAGE_ORDER;
5754 order = MAX_ORDER - 1;
5757 * Assume the largest contiguous order of interest is a huge page.
5758 * This value may be variable depending on boot parameters on IA64 and
5761 pageblock_order = order;
5763 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5766 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5767 * is unused as pageblock_order is set at compile-time. See
5768 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5771 void __paginginit set_pageblock_order(void)
5775 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5777 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5778 unsigned long present_pages)
5780 unsigned long pages = spanned_pages;
5783 * Provide a more accurate estimation if there are holes within
5784 * the zone and SPARSEMEM is in use. If there are holes within the
5785 * zone, each populated memory region may cost us one or two extra
5786 * memmap pages due to alignment because memmap pages for each
5787 * populated regions may not naturally algined on page boundary.
5788 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5790 if (spanned_pages > present_pages + (present_pages >> 4) &&
5791 IS_ENABLED(CONFIG_SPARSEMEM))
5792 pages = present_pages;
5794 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5798 * Set up the zone data structures:
5799 * - mark all pages reserved
5800 * - mark all memory queues empty
5801 * - clear the memory bitmaps
5803 * NOTE: pgdat should get zeroed by caller.
5805 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5808 int nid = pgdat->node_id;
5811 pgdat_resize_init(pgdat);
5812 #ifdef CONFIG_NUMA_BALANCING
5813 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5814 pgdat->numabalancing_migrate_nr_pages = 0;
5815 pgdat->numabalancing_migrate_next_window = jiffies;
5817 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5818 spin_lock_init(&pgdat->split_queue_lock);
5819 INIT_LIST_HEAD(&pgdat->split_queue);
5820 pgdat->split_queue_len = 0;
5822 init_waitqueue_head(&pgdat->kswapd_wait);
5823 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5824 #ifdef CONFIG_COMPACTION
5825 init_waitqueue_head(&pgdat->kcompactd_wait);
5827 pgdat_page_ext_init(pgdat);
5828 spin_lock_init(&pgdat->lru_lock);
5829 lruvec_init(node_lruvec(pgdat));
5831 for (j = 0; j < MAX_NR_ZONES; j++) {
5832 struct zone *zone = pgdat->node_zones + j;
5833 unsigned long size, realsize, freesize, memmap_pages;
5834 unsigned long zone_start_pfn = zone->zone_start_pfn;
5836 size = zone->spanned_pages;
5837 realsize = freesize = zone->present_pages;
5840 * Adjust freesize so that it accounts for how much memory
5841 * is used by this zone for memmap. This affects the watermark
5842 * and per-cpu initialisations
5844 memmap_pages = calc_memmap_size(size, realsize);
5845 if (!is_highmem_idx(j)) {
5846 if (freesize >= memmap_pages) {
5847 freesize -= memmap_pages;
5850 " %s zone: %lu pages used for memmap\n",
5851 zone_names[j], memmap_pages);
5853 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5854 zone_names[j], memmap_pages, freesize);
5857 /* Account for reserved pages */
5858 if (j == 0 && freesize > dma_reserve) {
5859 freesize -= dma_reserve;
5860 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5861 zone_names[0], dma_reserve);
5864 if (!is_highmem_idx(j))
5865 nr_kernel_pages += freesize;
5866 /* Charge for highmem memmap if there are enough kernel pages */
5867 else if (nr_kernel_pages > memmap_pages * 2)
5868 nr_kernel_pages -= memmap_pages;
5869 nr_all_pages += freesize;
5872 * Set an approximate value for lowmem here, it will be adjusted
5873 * when the bootmem allocator frees pages into the buddy system.
5874 * And all highmem pages will be managed by the buddy system.
5876 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5880 zone->name = zone_names[j];
5881 zone->zone_pgdat = pgdat;
5882 spin_lock_init(&zone->lock);
5883 zone_seqlock_init(zone);
5884 zone_pcp_init(zone);
5889 set_pageblock_order();
5890 setup_usemap(pgdat, zone, zone_start_pfn, size);
5891 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5893 memmap_init(size, nid, j, zone_start_pfn);
5897 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5899 unsigned long __maybe_unused start = 0;
5900 unsigned long __maybe_unused offset = 0;
5902 /* Skip empty nodes */
5903 if (!pgdat->node_spanned_pages)
5906 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5907 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5908 offset = pgdat->node_start_pfn - start;
5909 /* ia64 gets its own node_mem_map, before this, without bootmem */
5910 if (!pgdat->node_mem_map) {
5911 unsigned long size, end;
5915 * The zone's endpoints aren't required to be MAX_ORDER
5916 * aligned but the node_mem_map endpoints must be in order
5917 * for the buddy allocator to function correctly.
5919 end = pgdat_end_pfn(pgdat);
5920 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5921 size = (end - start) * sizeof(struct page);
5922 map = alloc_remap(pgdat->node_id, size);
5924 map = memblock_virt_alloc_node_nopanic(size,
5926 pgdat->node_mem_map = map + offset;
5928 #ifndef CONFIG_NEED_MULTIPLE_NODES
5930 * With no DISCONTIG, the global mem_map is just set as node 0's
5932 if (pgdat == NODE_DATA(0)) {
5933 mem_map = NODE_DATA(0)->node_mem_map;
5934 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5935 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5937 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5940 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5943 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5944 unsigned long node_start_pfn, unsigned long *zholes_size)
5946 pg_data_t *pgdat = NODE_DATA(nid);
5947 unsigned long start_pfn = 0;
5948 unsigned long end_pfn = 0;
5950 /* pg_data_t should be reset to zero when it's allocated */
5951 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5953 reset_deferred_meminit(pgdat);
5954 pgdat->node_id = nid;
5955 pgdat->node_start_pfn = node_start_pfn;
5956 pgdat->per_cpu_nodestats = NULL;
5957 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5958 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5959 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5960 (u64)start_pfn << PAGE_SHIFT,
5961 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5963 start_pfn = node_start_pfn;
5965 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5966 zones_size, zholes_size);
5968 alloc_node_mem_map(pgdat);
5969 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5970 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5971 nid, (unsigned long)pgdat,
5972 (unsigned long)pgdat->node_mem_map);
5975 free_area_init_core(pgdat);
5978 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5980 #if MAX_NUMNODES > 1
5982 * Figure out the number of possible node ids.
5984 void __init setup_nr_node_ids(void)
5986 unsigned int highest;
5988 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5989 nr_node_ids = highest + 1;
5994 * node_map_pfn_alignment - determine the maximum internode alignment
5996 * This function should be called after node map is populated and sorted.
5997 * It calculates the maximum power of two alignment which can distinguish
6000 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6001 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6002 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6003 * shifted, 1GiB is enough and this function will indicate so.
6005 * This is used to test whether pfn -> nid mapping of the chosen memory
6006 * model has fine enough granularity to avoid incorrect mapping for the
6007 * populated node map.
6009 * Returns the determined alignment in pfn's. 0 if there is no alignment
6010 * requirement (single node).
6012 unsigned long __init node_map_pfn_alignment(void)
6014 unsigned long accl_mask = 0, last_end = 0;
6015 unsigned long start, end, mask;
6019 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6020 if (!start || last_nid < 0 || last_nid == nid) {
6027 * Start with a mask granular enough to pin-point to the
6028 * start pfn and tick off bits one-by-one until it becomes
6029 * too coarse to separate the current node from the last.
6031 mask = ~((1 << __ffs(start)) - 1);
6032 while (mask && last_end <= (start & (mask << 1)))
6035 /* accumulate all internode masks */
6039 /* convert mask to number of pages */
6040 return ~accl_mask + 1;
6043 /* Find the lowest pfn for a node */
6044 static unsigned long __init find_min_pfn_for_node(int nid)
6046 unsigned long min_pfn = ULONG_MAX;
6047 unsigned long start_pfn;
6050 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6051 min_pfn = min(min_pfn, start_pfn);
6053 if (min_pfn == ULONG_MAX) {
6054 pr_warn("Could not find start_pfn for node %d\n", nid);
6062 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6064 * It returns the minimum PFN based on information provided via
6065 * memblock_set_node().
6067 unsigned long __init find_min_pfn_with_active_regions(void)
6069 return find_min_pfn_for_node(MAX_NUMNODES);
6073 * early_calculate_totalpages()
6074 * Sum pages in active regions for movable zone.
6075 * Populate N_MEMORY for calculating usable_nodes.
6077 static unsigned long __init early_calculate_totalpages(void)
6079 unsigned long totalpages = 0;
6080 unsigned long start_pfn, end_pfn;
6083 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6084 unsigned long pages = end_pfn - start_pfn;
6086 totalpages += pages;
6088 node_set_state(nid, N_MEMORY);
6094 * Find the PFN the Movable zone begins in each node. Kernel memory
6095 * is spread evenly between nodes as long as the nodes have enough
6096 * memory. When they don't, some nodes will have more kernelcore than
6099 static void __init find_zone_movable_pfns_for_nodes(void)
6102 unsigned long usable_startpfn;
6103 unsigned long kernelcore_node, kernelcore_remaining;
6104 /* save the state before borrow the nodemask */
6105 nodemask_t saved_node_state = node_states[N_MEMORY];
6106 unsigned long totalpages = early_calculate_totalpages();
6107 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6108 struct memblock_region *r;
6110 /* Need to find movable_zone earlier when movable_node is specified. */
6111 find_usable_zone_for_movable();
6114 * If movable_node is specified, ignore kernelcore and movablecore
6117 if (movable_node_is_enabled()) {
6118 for_each_memblock(memory, r) {
6119 if (!memblock_is_hotpluggable(r))
6124 usable_startpfn = PFN_DOWN(r->base);
6125 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6126 min(usable_startpfn, zone_movable_pfn[nid]) :
6134 * If kernelcore=mirror is specified, ignore movablecore option
6136 if (mirrored_kernelcore) {
6137 bool mem_below_4gb_not_mirrored = false;
6139 for_each_memblock(memory, r) {
6140 if (memblock_is_mirror(r))
6145 usable_startpfn = memblock_region_memory_base_pfn(r);
6147 if (usable_startpfn < 0x100000) {
6148 mem_below_4gb_not_mirrored = true;
6152 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6153 min(usable_startpfn, zone_movable_pfn[nid]) :
6157 if (mem_below_4gb_not_mirrored)
6158 pr_warn("This configuration results in unmirrored kernel memory.");
6164 * If movablecore=nn[KMG] was specified, calculate what size of
6165 * kernelcore that corresponds so that memory usable for
6166 * any allocation type is evenly spread. If both kernelcore
6167 * and movablecore are specified, then the value of kernelcore
6168 * will be used for required_kernelcore if it's greater than
6169 * what movablecore would have allowed.
6171 if (required_movablecore) {
6172 unsigned long corepages;
6175 * Round-up so that ZONE_MOVABLE is at least as large as what
6176 * was requested by the user
6178 required_movablecore =
6179 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6180 required_movablecore = min(totalpages, required_movablecore);
6181 corepages = totalpages - required_movablecore;
6183 required_kernelcore = max(required_kernelcore, corepages);
6187 * If kernelcore was not specified or kernelcore size is larger
6188 * than totalpages, there is no ZONE_MOVABLE.
6190 if (!required_kernelcore || required_kernelcore >= totalpages)
6193 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6194 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6197 /* Spread kernelcore memory as evenly as possible throughout nodes */
6198 kernelcore_node = required_kernelcore / usable_nodes;
6199 for_each_node_state(nid, N_MEMORY) {
6200 unsigned long start_pfn, end_pfn;
6203 * Recalculate kernelcore_node if the division per node
6204 * now exceeds what is necessary to satisfy the requested
6205 * amount of memory for the kernel
6207 if (required_kernelcore < kernelcore_node)
6208 kernelcore_node = required_kernelcore / usable_nodes;
6211 * As the map is walked, we track how much memory is usable
6212 * by the kernel using kernelcore_remaining. When it is
6213 * 0, the rest of the node is usable by ZONE_MOVABLE
6215 kernelcore_remaining = kernelcore_node;
6217 /* Go through each range of PFNs within this node */
6218 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6219 unsigned long size_pages;
6221 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6222 if (start_pfn >= end_pfn)
6225 /* Account for what is only usable for kernelcore */
6226 if (start_pfn < usable_startpfn) {
6227 unsigned long kernel_pages;
6228 kernel_pages = min(end_pfn, usable_startpfn)
6231 kernelcore_remaining -= min(kernel_pages,
6232 kernelcore_remaining);
6233 required_kernelcore -= min(kernel_pages,
6234 required_kernelcore);
6236 /* Continue if range is now fully accounted */
6237 if (end_pfn <= usable_startpfn) {
6240 * Push zone_movable_pfn to the end so
6241 * that if we have to rebalance
6242 * kernelcore across nodes, we will
6243 * not double account here
6245 zone_movable_pfn[nid] = end_pfn;
6248 start_pfn = usable_startpfn;
6252 * The usable PFN range for ZONE_MOVABLE is from
6253 * start_pfn->end_pfn. Calculate size_pages as the
6254 * number of pages used as kernelcore
6256 size_pages = end_pfn - start_pfn;
6257 if (size_pages > kernelcore_remaining)
6258 size_pages = kernelcore_remaining;
6259 zone_movable_pfn[nid] = start_pfn + size_pages;
6262 * Some kernelcore has been met, update counts and
6263 * break if the kernelcore for this node has been
6266 required_kernelcore -= min(required_kernelcore,
6268 kernelcore_remaining -= size_pages;
6269 if (!kernelcore_remaining)
6275 * If there is still required_kernelcore, we do another pass with one
6276 * less node in the count. This will push zone_movable_pfn[nid] further
6277 * along on the nodes that still have memory until kernelcore is
6281 if (usable_nodes && required_kernelcore > usable_nodes)
6285 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6286 for (nid = 0; nid < MAX_NUMNODES; nid++)
6287 zone_movable_pfn[nid] =
6288 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6291 /* restore the node_state */
6292 node_states[N_MEMORY] = saved_node_state;
6295 /* Any regular or high memory on that node ? */
6296 static void check_for_memory(pg_data_t *pgdat, int nid)
6298 enum zone_type zone_type;
6300 if (N_MEMORY == N_NORMAL_MEMORY)
6303 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6304 struct zone *zone = &pgdat->node_zones[zone_type];
6305 if (populated_zone(zone)) {
6306 node_set_state(nid, N_HIGH_MEMORY);
6307 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6308 zone_type <= ZONE_NORMAL)
6309 node_set_state(nid, N_NORMAL_MEMORY);
6316 * free_area_init_nodes - Initialise all pg_data_t and zone data
6317 * @max_zone_pfn: an array of max PFNs for each zone
6319 * This will call free_area_init_node() for each active node in the system.
6320 * Using the page ranges provided by memblock_set_node(), the size of each
6321 * zone in each node and their holes is calculated. If the maximum PFN
6322 * between two adjacent zones match, it is assumed that the zone is empty.
6323 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6324 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6325 * starts where the previous one ended. For example, ZONE_DMA32 starts
6326 * at arch_max_dma_pfn.
6328 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6330 unsigned long start_pfn, end_pfn;
6333 /* Record where the zone boundaries are */
6334 memset(arch_zone_lowest_possible_pfn, 0,
6335 sizeof(arch_zone_lowest_possible_pfn));
6336 memset(arch_zone_highest_possible_pfn, 0,
6337 sizeof(arch_zone_highest_possible_pfn));
6339 start_pfn = find_min_pfn_with_active_regions();
6341 for (i = 0; i < MAX_NR_ZONES; i++) {
6342 if (i == ZONE_MOVABLE)
6345 end_pfn = max(max_zone_pfn[i], start_pfn);
6346 arch_zone_lowest_possible_pfn[i] = start_pfn;
6347 arch_zone_highest_possible_pfn[i] = end_pfn;
6349 start_pfn = end_pfn;
6351 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6352 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6354 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6355 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6356 find_zone_movable_pfns_for_nodes();
6358 /* Print out the zone ranges */
6359 pr_info("Zone ranges:\n");
6360 for (i = 0; i < MAX_NR_ZONES; i++) {
6361 if (i == ZONE_MOVABLE)
6363 pr_info(" %-8s ", zone_names[i]);
6364 if (arch_zone_lowest_possible_pfn[i] ==
6365 arch_zone_highest_possible_pfn[i])
6368 pr_cont("[mem %#018Lx-%#018Lx]\n",
6369 (u64)arch_zone_lowest_possible_pfn[i]
6371 ((u64)arch_zone_highest_possible_pfn[i]
6372 << PAGE_SHIFT) - 1);
6375 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6376 pr_info("Movable zone start for each node\n");
6377 for (i = 0; i < MAX_NUMNODES; i++) {
6378 if (zone_movable_pfn[i])
6379 pr_info(" Node %d: %#018Lx\n", i,
6380 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6383 /* Print out the early node map */
6384 pr_info("Early memory node ranges\n");
6385 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6386 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6387 (u64)start_pfn << PAGE_SHIFT,
6388 ((u64)end_pfn << PAGE_SHIFT) - 1);
6390 /* Initialise every node */
6391 mminit_verify_pageflags_layout();
6392 setup_nr_node_ids();
6393 for_each_online_node(nid) {
6394 pg_data_t *pgdat = NODE_DATA(nid);
6395 free_area_init_node(nid, NULL,
6396 find_min_pfn_for_node(nid), NULL);
6398 /* Any memory on that node */
6399 if (pgdat->node_present_pages)
6400 node_set_state(nid, N_MEMORY);
6401 check_for_memory(pgdat, nid);
6405 static int __init cmdline_parse_core(char *p, unsigned long *core)
6407 unsigned long long coremem;
6411 coremem = memparse(p, &p);
6412 *core = coremem >> PAGE_SHIFT;
6414 /* Paranoid check that UL is enough for the coremem value */
6415 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6421 * kernelcore=size sets the amount of memory for use for allocations that
6422 * cannot be reclaimed or migrated.
6424 static int __init cmdline_parse_kernelcore(char *p)
6426 /* parse kernelcore=mirror */
6427 if (parse_option_str(p, "mirror")) {
6428 mirrored_kernelcore = true;
6432 return cmdline_parse_core(p, &required_kernelcore);
6436 * movablecore=size sets the amount of memory for use for allocations that
6437 * can be reclaimed or migrated.
6439 static int __init cmdline_parse_movablecore(char *p)
6441 return cmdline_parse_core(p, &required_movablecore);
6444 early_param("kernelcore", cmdline_parse_kernelcore);
6445 early_param("movablecore", cmdline_parse_movablecore);
6447 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6449 void adjust_managed_page_count(struct page *page, long count)
6451 spin_lock(&managed_page_count_lock);
6452 page_zone(page)->managed_pages += count;
6453 totalram_pages += count;
6454 #ifdef CONFIG_HIGHMEM
6455 if (PageHighMem(page))
6456 totalhigh_pages += count;
6458 spin_unlock(&managed_page_count_lock);
6460 EXPORT_SYMBOL(adjust_managed_page_count);
6462 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6465 unsigned long pages = 0;
6467 start = (void *)PAGE_ALIGN((unsigned long)start);
6468 end = (void *)((unsigned long)end & PAGE_MASK);
6469 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6470 if ((unsigned int)poison <= 0xFF)
6471 memset(pos, poison, PAGE_SIZE);
6472 free_reserved_page(virt_to_page(pos));
6476 pr_info("Freeing %s memory: %ldK\n",
6477 s, pages << (PAGE_SHIFT - 10));
6481 EXPORT_SYMBOL(free_reserved_area);
6483 #ifdef CONFIG_HIGHMEM
6484 void free_highmem_page(struct page *page)
6486 __free_reserved_page(page);
6488 page_zone(page)->managed_pages++;
6494 void __init mem_init_print_info(const char *str)
6496 unsigned long physpages, codesize, datasize, rosize, bss_size;
6497 unsigned long init_code_size, init_data_size;
6499 physpages = get_num_physpages();
6500 codesize = _etext - _stext;
6501 datasize = _edata - _sdata;
6502 rosize = __end_rodata - __start_rodata;
6503 bss_size = __bss_stop - __bss_start;
6504 init_data_size = __init_end - __init_begin;
6505 init_code_size = _einittext - _sinittext;
6508 * Detect special cases and adjust section sizes accordingly:
6509 * 1) .init.* may be embedded into .data sections
6510 * 2) .init.text.* may be out of [__init_begin, __init_end],
6511 * please refer to arch/tile/kernel/vmlinux.lds.S.
6512 * 3) .rodata.* may be embedded into .text or .data sections.
6514 #define adj_init_size(start, end, size, pos, adj) \
6516 if (start <= pos && pos < end && size > adj) \
6520 adj_init_size(__init_begin, __init_end, init_data_size,
6521 _sinittext, init_code_size);
6522 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6523 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6524 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6525 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6527 #undef adj_init_size
6529 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6530 #ifdef CONFIG_HIGHMEM
6534 nr_free_pages() << (PAGE_SHIFT - 10),
6535 physpages << (PAGE_SHIFT - 10),
6536 codesize >> 10, datasize >> 10, rosize >> 10,
6537 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6538 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6539 totalcma_pages << (PAGE_SHIFT - 10),
6540 #ifdef CONFIG_HIGHMEM
6541 totalhigh_pages << (PAGE_SHIFT - 10),
6543 str ? ", " : "", str ? str : "");
6547 * set_dma_reserve - set the specified number of pages reserved in the first zone
6548 * @new_dma_reserve: The number of pages to mark reserved
6550 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6551 * In the DMA zone, a significant percentage may be consumed by kernel image
6552 * and other unfreeable allocations which can skew the watermarks badly. This
6553 * function may optionally be used to account for unfreeable pages in the
6554 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6555 * smaller per-cpu batchsize.
6557 void __init set_dma_reserve(unsigned long new_dma_reserve)
6559 dma_reserve = new_dma_reserve;
6562 void __init free_area_init(unsigned long *zones_size)
6564 free_area_init_node(0, zones_size,
6565 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6568 static int page_alloc_cpu_dead(unsigned int cpu)
6571 lru_add_drain_cpu(cpu);
6575 * Spill the event counters of the dead processor
6576 * into the current processors event counters.
6577 * This artificially elevates the count of the current
6580 vm_events_fold_cpu(cpu);
6583 * Zero the differential counters of the dead processor
6584 * so that the vm statistics are consistent.
6586 * This is only okay since the processor is dead and cannot
6587 * race with what we are doing.
6589 cpu_vm_stats_fold(cpu);
6593 void __init page_alloc_init(void)
6597 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6598 "mm/page_alloc:dead", NULL,
6599 page_alloc_cpu_dead);
6604 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6605 * or min_free_kbytes changes.
6607 static void calculate_totalreserve_pages(void)
6609 struct pglist_data *pgdat;
6610 unsigned long reserve_pages = 0;
6611 enum zone_type i, j;
6613 for_each_online_pgdat(pgdat) {
6615 pgdat->totalreserve_pages = 0;
6617 for (i = 0; i < MAX_NR_ZONES; i++) {
6618 struct zone *zone = pgdat->node_zones + i;
6621 /* Find valid and maximum lowmem_reserve in the zone */
6622 for (j = i; j < MAX_NR_ZONES; j++) {
6623 if (zone->lowmem_reserve[j] > max)
6624 max = zone->lowmem_reserve[j];
6627 /* we treat the high watermark as reserved pages. */
6628 max += high_wmark_pages(zone);
6630 if (max > zone->managed_pages)
6631 max = zone->managed_pages;
6633 pgdat->totalreserve_pages += max;
6635 reserve_pages += max;
6638 totalreserve_pages = reserve_pages;
6642 * setup_per_zone_lowmem_reserve - called whenever
6643 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6644 * has a correct pages reserved value, so an adequate number of
6645 * pages are left in the zone after a successful __alloc_pages().
6647 static void setup_per_zone_lowmem_reserve(void)
6649 struct pglist_data *pgdat;
6650 enum zone_type j, idx;
6652 for_each_online_pgdat(pgdat) {
6653 for (j = 0; j < MAX_NR_ZONES; j++) {
6654 struct zone *zone = pgdat->node_zones + j;
6655 unsigned long managed_pages = zone->managed_pages;
6657 zone->lowmem_reserve[j] = 0;
6661 struct zone *lower_zone;
6665 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6666 sysctl_lowmem_reserve_ratio[idx] = 1;
6668 lower_zone = pgdat->node_zones + idx;
6669 lower_zone->lowmem_reserve[j] = managed_pages /
6670 sysctl_lowmem_reserve_ratio[idx];
6671 managed_pages += lower_zone->managed_pages;
6676 /* update totalreserve_pages */
6677 calculate_totalreserve_pages();
6680 static void __setup_per_zone_wmarks(void)
6682 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6683 unsigned long lowmem_pages = 0;
6685 unsigned long flags;
6687 /* Calculate total number of !ZONE_HIGHMEM pages */
6688 for_each_zone(zone) {
6689 if (!is_highmem(zone))
6690 lowmem_pages += zone->managed_pages;
6693 for_each_zone(zone) {
6696 spin_lock_irqsave(&zone->lock, flags);
6697 tmp = (u64)pages_min * zone->managed_pages;
6698 do_div(tmp, lowmem_pages);
6699 if (is_highmem(zone)) {
6701 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6702 * need highmem pages, so cap pages_min to a small
6705 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6706 * deltas control asynch page reclaim, and so should
6707 * not be capped for highmem.
6709 unsigned long min_pages;
6711 min_pages = zone->managed_pages / 1024;
6712 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6713 zone->watermark[WMARK_MIN] = min_pages;
6716 * If it's a lowmem zone, reserve a number of pages
6717 * proportionate to the zone's size.
6719 zone->watermark[WMARK_MIN] = tmp;
6723 * Set the kswapd watermarks distance according to the
6724 * scale factor in proportion to available memory, but
6725 * ensure a minimum size on small systems.
6727 tmp = max_t(u64, tmp >> 2,
6728 mult_frac(zone->managed_pages,
6729 watermark_scale_factor, 10000));
6731 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6732 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6734 spin_unlock_irqrestore(&zone->lock, flags);
6737 /* update totalreserve_pages */
6738 calculate_totalreserve_pages();
6742 * setup_per_zone_wmarks - called when min_free_kbytes changes
6743 * or when memory is hot-{added|removed}
6745 * Ensures that the watermark[min,low,high] values for each zone are set
6746 * correctly with respect to min_free_kbytes.
6748 void setup_per_zone_wmarks(void)
6750 mutex_lock(&zonelists_mutex);
6751 __setup_per_zone_wmarks();
6752 mutex_unlock(&zonelists_mutex);
6756 * Initialise min_free_kbytes.
6758 * For small machines we want it small (128k min). For large machines
6759 * we want it large (64MB max). But it is not linear, because network
6760 * bandwidth does not increase linearly with machine size. We use
6762 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6763 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6779 int __meminit init_per_zone_wmark_min(void)
6781 unsigned long lowmem_kbytes;
6782 int new_min_free_kbytes;
6784 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6785 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6787 if (new_min_free_kbytes > user_min_free_kbytes) {
6788 min_free_kbytes = new_min_free_kbytes;
6789 if (min_free_kbytes < 128)
6790 min_free_kbytes = 128;
6791 if (min_free_kbytes > 65536)
6792 min_free_kbytes = 65536;
6794 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6795 new_min_free_kbytes, user_min_free_kbytes);
6797 setup_per_zone_wmarks();
6798 refresh_zone_stat_thresholds();
6799 setup_per_zone_lowmem_reserve();
6802 setup_min_unmapped_ratio();
6803 setup_min_slab_ratio();
6808 core_initcall(init_per_zone_wmark_min)
6811 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6812 * that we can call two helper functions whenever min_free_kbytes
6815 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6816 void __user *buffer, size_t *length, loff_t *ppos)
6820 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6825 user_min_free_kbytes = min_free_kbytes;
6826 setup_per_zone_wmarks();
6831 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6832 void __user *buffer, size_t *length, loff_t *ppos)
6836 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6841 setup_per_zone_wmarks();
6847 static void setup_min_unmapped_ratio(void)
6852 for_each_online_pgdat(pgdat)
6853 pgdat->min_unmapped_pages = 0;
6856 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6857 sysctl_min_unmapped_ratio) / 100;
6861 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6862 void __user *buffer, size_t *length, loff_t *ppos)
6866 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6870 setup_min_unmapped_ratio();
6875 static void setup_min_slab_ratio(void)
6880 for_each_online_pgdat(pgdat)
6881 pgdat->min_slab_pages = 0;
6884 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6885 sysctl_min_slab_ratio) / 100;
6888 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6889 void __user *buffer, size_t *length, loff_t *ppos)
6893 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6897 setup_min_slab_ratio();
6904 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6905 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6906 * whenever sysctl_lowmem_reserve_ratio changes.
6908 * The reserve ratio obviously has absolutely no relation with the
6909 * minimum watermarks. The lowmem reserve ratio can only make sense
6910 * if in function of the boot time zone sizes.
6912 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6913 void __user *buffer, size_t *length, loff_t *ppos)
6915 proc_dointvec_minmax(table, write, buffer, length, ppos);
6916 setup_per_zone_lowmem_reserve();
6921 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6922 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6923 * pagelist can have before it gets flushed back to buddy allocator.
6925 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6926 void __user *buffer, size_t *length, loff_t *ppos)
6929 int old_percpu_pagelist_fraction;
6932 mutex_lock(&pcp_batch_high_lock);
6933 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6935 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6936 if (!write || ret < 0)
6939 /* Sanity checking to avoid pcp imbalance */
6940 if (percpu_pagelist_fraction &&
6941 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6942 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6948 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6951 for_each_populated_zone(zone) {
6954 for_each_possible_cpu(cpu)
6955 pageset_set_high_and_batch(zone,
6956 per_cpu_ptr(zone->pageset, cpu));
6959 mutex_unlock(&pcp_batch_high_lock);
6964 int hashdist = HASHDIST_DEFAULT;
6966 static int __init set_hashdist(char *str)
6970 hashdist = simple_strtoul(str, &str, 0);
6973 __setup("hashdist=", set_hashdist);
6976 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6978 * Returns the number of pages that arch has reserved but
6979 * is not known to alloc_large_system_hash().
6981 static unsigned long __init arch_reserved_kernel_pages(void)
6988 * allocate a large system hash table from bootmem
6989 * - it is assumed that the hash table must contain an exact power-of-2
6990 * quantity of entries
6991 * - limit is the number of hash buckets, not the total allocation size
6993 void *__init alloc_large_system_hash(const char *tablename,
6994 unsigned long bucketsize,
6995 unsigned long numentries,
6998 unsigned int *_hash_shift,
6999 unsigned int *_hash_mask,
7000 unsigned long low_limit,
7001 unsigned long high_limit)
7003 unsigned long long max = high_limit;
7004 unsigned long log2qty, size;
7007 /* allow the kernel cmdline to have a say */
7009 /* round applicable memory size up to nearest megabyte */
7010 numentries = nr_kernel_pages;
7011 numentries -= arch_reserved_kernel_pages();
7013 /* It isn't necessary when PAGE_SIZE >= 1MB */
7014 if (PAGE_SHIFT < 20)
7015 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7017 /* limit to 1 bucket per 2^scale bytes of low memory */
7018 if (scale > PAGE_SHIFT)
7019 numentries >>= (scale - PAGE_SHIFT);
7021 numentries <<= (PAGE_SHIFT - scale);
7023 /* Make sure we've got at least a 0-order allocation.. */
7024 if (unlikely(flags & HASH_SMALL)) {
7025 /* Makes no sense without HASH_EARLY */
7026 WARN_ON(!(flags & HASH_EARLY));
7027 if (!(numentries >> *_hash_shift)) {
7028 numentries = 1UL << *_hash_shift;
7029 BUG_ON(!numentries);
7031 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7032 numentries = PAGE_SIZE / bucketsize;
7034 numentries = roundup_pow_of_two(numentries);
7036 /* limit allocation size to 1/16 total memory by default */
7038 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7039 do_div(max, bucketsize);
7041 max = min(max, 0x80000000ULL);
7043 if (numentries < low_limit)
7044 numentries = low_limit;
7045 if (numentries > max)
7048 log2qty = ilog2(numentries);
7051 size = bucketsize << log2qty;
7052 if (flags & HASH_EARLY)
7053 table = memblock_virt_alloc_nopanic(size, 0);
7055 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7058 * If bucketsize is not a power-of-two, we may free
7059 * some pages at the end of hash table which
7060 * alloc_pages_exact() automatically does
7062 if (get_order(size) < MAX_ORDER) {
7063 table = alloc_pages_exact(size, GFP_ATOMIC);
7064 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7067 } while (!table && size > PAGE_SIZE && --log2qty);
7070 panic("Failed to allocate %s hash table\n", tablename);
7072 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7073 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7076 *_hash_shift = log2qty;
7078 *_hash_mask = (1 << log2qty) - 1;
7084 * This function checks whether pageblock includes unmovable pages or not.
7085 * If @count is not zero, it is okay to include less @count unmovable pages
7087 * PageLRU check without isolation or lru_lock could race so that
7088 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7089 * expect this function should be exact.
7091 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7092 bool skip_hwpoisoned_pages)
7094 unsigned long pfn, iter, found;
7098 * For avoiding noise data, lru_add_drain_all() should be called
7099 * If ZONE_MOVABLE, the zone never contains unmovable pages
7101 if (zone_idx(zone) == ZONE_MOVABLE)
7103 mt = get_pageblock_migratetype(page);
7104 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7107 pfn = page_to_pfn(page);
7108 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7109 unsigned long check = pfn + iter;
7111 if (!pfn_valid_within(check))
7114 page = pfn_to_page(check);
7117 * Hugepages are not in LRU lists, but they're movable.
7118 * We need not scan over tail pages bacause we don't
7119 * handle each tail page individually in migration.
7121 if (PageHuge(page)) {
7122 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7127 * We can't use page_count without pin a page
7128 * because another CPU can free compound page.
7129 * This check already skips compound tails of THP
7130 * because their page->_refcount is zero at all time.
7132 if (!page_ref_count(page)) {
7133 if (PageBuddy(page))
7134 iter += (1 << page_order(page)) - 1;
7139 * The HWPoisoned page may be not in buddy system, and
7140 * page_count() is not 0.
7142 if (skip_hwpoisoned_pages && PageHWPoison(page))
7148 * If there are RECLAIMABLE pages, we need to check
7149 * it. But now, memory offline itself doesn't call
7150 * shrink_node_slabs() and it still to be fixed.
7153 * If the page is not RAM, page_count()should be 0.
7154 * we don't need more check. This is an _used_ not-movable page.
7156 * The problematic thing here is PG_reserved pages. PG_reserved
7157 * is set to both of a memory hole page and a _used_ kernel
7166 bool is_pageblock_removable_nolock(struct page *page)
7172 * We have to be careful here because we are iterating over memory
7173 * sections which are not zone aware so we might end up outside of
7174 * the zone but still within the section.
7175 * We have to take care about the node as well. If the node is offline
7176 * its NODE_DATA will be NULL - see page_zone.
7178 if (!node_online(page_to_nid(page)))
7181 zone = page_zone(page);
7182 pfn = page_to_pfn(page);
7183 if (!zone_spans_pfn(zone, pfn))
7186 return !has_unmovable_pages(zone, page, 0, true);
7189 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7191 static unsigned long pfn_max_align_down(unsigned long pfn)
7193 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7194 pageblock_nr_pages) - 1);
7197 static unsigned long pfn_max_align_up(unsigned long pfn)
7199 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7200 pageblock_nr_pages));
7203 /* [start, end) must belong to a single zone. */
7204 static int __alloc_contig_migrate_range(struct compact_control *cc,
7205 unsigned long start, unsigned long end)
7207 /* This function is based on compact_zone() from compaction.c. */
7208 unsigned long nr_reclaimed;
7209 unsigned long pfn = start;
7210 unsigned int tries = 0;
7215 while (pfn < end || !list_empty(&cc->migratepages)) {
7216 if (fatal_signal_pending(current)) {
7221 if (list_empty(&cc->migratepages)) {
7222 cc->nr_migratepages = 0;
7223 pfn = isolate_migratepages_range(cc, pfn, end);
7229 } else if (++tries == 5) {
7230 ret = ret < 0 ? ret : -EBUSY;
7234 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7236 cc->nr_migratepages -= nr_reclaimed;
7238 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7239 NULL, 0, cc->mode, MR_CMA);
7242 putback_movable_pages(&cc->migratepages);
7249 * alloc_contig_range() -- tries to allocate given range of pages
7250 * @start: start PFN to allocate
7251 * @end: one-past-the-last PFN to allocate
7252 * @migratetype: migratetype of the underlaying pageblocks (either
7253 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7254 * in range must have the same migratetype and it must
7255 * be either of the two.
7257 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7258 * aligned, however it's the caller's responsibility to guarantee that
7259 * we are the only thread that changes migrate type of pageblocks the
7262 * The PFN range must belong to a single zone.
7264 * Returns zero on success or negative error code. On success all
7265 * pages which PFN is in [start, end) are allocated for the caller and
7266 * need to be freed with free_contig_range().
7268 int alloc_contig_range(unsigned long start, unsigned long end,
7269 unsigned migratetype)
7271 unsigned long outer_start, outer_end;
7275 struct compact_control cc = {
7276 .nr_migratepages = 0,
7278 .zone = page_zone(pfn_to_page(start)),
7279 .mode = MIGRATE_SYNC,
7280 .ignore_skip_hint = true,
7281 .gfp_mask = GFP_KERNEL,
7283 INIT_LIST_HEAD(&cc.migratepages);
7286 * What we do here is we mark all pageblocks in range as
7287 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7288 * have different sizes, and due to the way page allocator
7289 * work, we align the range to biggest of the two pages so
7290 * that page allocator won't try to merge buddies from
7291 * different pageblocks and change MIGRATE_ISOLATE to some
7292 * other migration type.
7294 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7295 * migrate the pages from an unaligned range (ie. pages that
7296 * we are interested in). This will put all the pages in
7297 * range back to page allocator as MIGRATE_ISOLATE.
7299 * When this is done, we take the pages in range from page
7300 * allocator removing them from the buddy system. This way
7301 * page allocator will never consider using them.
7303 * This lets us mark the pageblocks back as
7304 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7305 * aligned range but not in the unaligned, original range are
7306 * put back to page allocator so that buddy can use them.
7309 ret = start_isolate_page_range(pfn_max_align_down(start),
7310 pfn_max_align_up(end), migratetype,
7316 * In case of -EBUSY, we'd like to know which page causes problem.
7317 * So, just fall through. We will check it in test_pages_isolated().
7319 ret = __alloc_contig_migrate_range(&cc, start, end);
7320 if (ret && ret != -EBUSY)
7324 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7325 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7326 * more, all pages in [start, end) are free in page allocator.
7327 * What we are going to do is to allocate all pages from
7328 * [start, end) (that is remove them from page allocator).
7330 * The only problem is that pages at the beginning and at the
7331 * end of interesting range may be not aligned with pages that
7332 * page allocator holds, ie. they can be part of higher order
7333 * pages. Because of this, we reserve the bigger range and
7334 * once this is done free the pages we are not interested in.
7336 * We don't have to hold zone->lock here because the pages are
7337 * isolated thus they won't get removed from buddy.
7340 lru_add_drain_all();
7341 drain_all_pages(cc.zone);
7344 outer_start = start;
7345 while (!PageBuddy(pfn_to_page(outer_start))) {
7346 if (++order >= MAX_ORDER) {
7347 outer_start = start;
7350 outer_start &= ~0UL << order;
7353 if (outer_start != start) {
7354 order = page_order(pfn_to_page(outer_start));
7357 * outer_start page could be small order buddy page and
7358 * it doesn't include start page. Adjust outer_start
7359 * in this case to report failed page properly
7360 * on tracepoint in test_pages_isolated()
7362 if (outer_start + (1UL << order) <= start)
7363 outer_start = start;
7366 /* Make sure the range is really isolated. */
7367 if (test_pages_isolated(outer_start, end, false)) {
7368 pr_info("%s: [%lx, %lx) PFNs busy\n",
7369 __func__, outer_start, end);
7374 /* Grab isolated pages from freelists. */
7375 outer_end = isolate_freepages_range(&cc, outer_start, end);
7381 /* Free head and tail (if any) */
7382 if (start != outer_start)
7383 free_contig_range(outer_start, start - outer_start);
7384 if (end != outer_end)
7385 free_contig_range(end, outer_end - end);
7388 undo_isolate_page_range(pfn_max_align_down(start),
7389 pfn_max_align_up(end), migratetype);
7393 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7395 unsigned int count = 0;
7397 for (; nr_pages--; pfn++) {
7398 struct page *page = pfn_to_page(pfn);
7400 count += page_count(page) != 1;
7403 WARN(count != 0, "%d pages are still in use!\n", count);
7407 #ifdef CONFIG_MEMORY_HOTPLUG
7409 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7410 * page high values need to be recalulated.
7412 void __meminit zone_pcp_update(struct zone *zone)
7415 mutex_lock(&pcp_batch_high_lock);
7416 for_each_possible_cpu(cpu)
7417 pageset_set_high_and_batch(zone,
7418 per_cpu_ptr(zone->pageset, cpu));
7419 mutex_unlock(&pcp_batch_high_lock);
7423 void zone_pcp_reset(struct zone *zone)
7425 unsigned long flags;
7427 struct per_cpu_pageset *pset;
7429 /* avoid races with drain_pages() */
7430 local_irq_save(flags);
7431 if (zone->pageset != &boot_pageset) {
7432 for_each_online_cpu(cpu) {
7433 pset = per_cpu_ptr(zone->pageset, cpu);
7434 drain_zonestat(zone, pset);
7436 free_percpu(zone->pageset);
7437 zone->pageset = &boot_pageset;
7439 local_irq_restore(flags);
7442 #ifdef CONFIG_MEMORY_HOTREMOVE
7444 * All pages in the range must be in a single zone and isolated
7445 * before calling this.
7448 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7452 unsigned int order, i;
7454 unsigned long flags;
7455 /* find the first valid pfn */
7456 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7461 zone = page_zone(pfn_to_page(pfn));
7462 spin_lock_irqsave(&zone->lock, flags);
7464 while (pfn < end_pfn) {
7465 if (!pfn_valid(pfn)) {
7469 page = pfn_to_page(pfn);
7471 * The HWPoisoned page may be not in buddy system, and
7472 * page_count() is not 0.
7474 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7476 SetPageReserved(page);
7480 BUG_ON(page_count(page));
7481 BUG_ON(!PageBuddy(page));
7482 order = page_order(page);
7483 #ifdef CONFIG_DEBUG_VM
7484 pr_info("remove from free list %lx %d %lx\n",
7485 pfn, 1 << order, end_pfn);
7487 list_del(&page->lru);
7488 rmv_page_order(page);
7489 zone->free_area[order].nr_free--;
7490 for (i = 0; i < (1 << order); i++)
7491 SetPageReserved((page+i));
7492 pfn += (1 << order);
7494 spin_unlock_irqrestore(&zone->lock, flags);
7498 bool is_free_buddy_page(struct page *page)
7500 struct zone *zone = page_zone(page);
7501 unsigned long pfn = page_to_pfn(page);
7502 unsigned long flags;
7505 spin_lock_irqsave(&zone->lock, flags);
7506 for (order = 0; order < MAX_ORDER; order++) {
7507 struct page *page_head = page - (pfn & ((1 << order) - 1));
7509 if (PageBuddy(page_head) && page_order(page_head) >= order)
7512 spin_unlock_irqrestore(&zone->lock, flags);
7514 return order < MAX_ORDER;