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/locallock.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
69 #include <asm/sections.h>
70 #include <asm/tlbflush.h>
71 #include <asm/div64.h>
74 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
75 static DEFINE_MUTEX(pcp_batch_high_lock);
76 #define MIN_PERCPU_PAGELIST_FRACTION (8)
78 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
79 DEFINE_PER_CPU(int, numa_node);
80 EXPORT_PER_CPU_SYMBOL(numa_node);
83 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
85 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
86 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
87 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
88 * defined in <linux/topology.h>.
90 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
91 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
92 int _node_numa_mem_[MAX_NUMNODES];
95 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
96 volatile unsigned long latent_entropy __latent_entropy;
97 EXPORT_SYMBOL(latent_entropy);
101 * Array of node states.
103 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
104 [N_POSSIBLE] = NODE_MASK_ALL,
105 [N_ONLINE] = { { [0] = 1UL } },
107 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
108 #ifdef CONFIG_HIGHMEM
109 [N_HIGH_MEMORY] = { { [0] = 1UL } },
111 #ifdef CONFIG_MOVABLE_NODE
112 [N_MEMORY] = { { [0] = 1UL } },
114 [N_CPU] = { { [0] = 1UL } },
117 EXPORT_SYMBOL(node_states);
119 /* Protect totalram_pages and zone->managed_pages */
120 static DEFINE_SPINLOCK(managed_page_count_lock);
122 unsigned long totalram_pages __read_mostly;
123 unsigned long totalreserve_pages __read_mostly;
124 unsigned long totalcma_pages __read_mostly;
126 int percpu_pagelist_fraction;
127 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
130 * A cached value of the page's pageblock's migratetype, used when the page is
131 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
132 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
133 * Also the migratetype set in the page does not necessarily match the pcplist
134 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
135 * other index - this ensures that it will be put on the correct CMA freelist.
137 static inline int get_pcppage_migratetype(struct page *page)
142 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
144 page->index = migratetype;
147 #ifdef CONFIG_PM_SLEEP
149 * The following functions are used by the suspend/hibernate code to temporarily
150 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
151 * while devices are suspended. To avoid races with the suspend/hibernate code,
152 * they should always be called with pm_mutex held (gfp_allowed_mask also should
153 * only be modified with pm_mutex held, unless the suspend/hibernate code is
154 * guaranteed not to run in parallel with that modification).
157 static gfp_t saved_gfp_mask;
159 void pm_restore_gfp_mask(void)
161 WARN_ON(!mutex_is_locked(&pm_mutex));
162 if (saved_gfp_mask) {
163 gfp_allowed_mask = saved_gfp_mask;
168 void pm_restrict_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&pm_mutex));
171 WARN_ON(saved_gfp_mask);
172 saved_gfp_mask = gfp_allowed_mask;
173 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
176 bool pm_suspended_storage(void)
178 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
182 #endif /* CONFIG_PM_SLEEP */
184 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
185 unsigned int pageblock_order __read_mostly;
188 static void __free_pages_ok(struct page *page, unsigned int order);
191 * results with 256, 32 in the lowmem_reserve sysctl:
192 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
193 * 1G machine -> (16M dma, 784M normal, 224M high)
194 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
195 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
196 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
198 * TBD: should special case ZONE_DMA32 machines here - in those we normally
199 * don't need any ZONE_NORMAL reservation
201 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
202 #ifdef CONFIG_ZONE_DMA
205 #ifdef CONFIG_ZONE_DMA32
208 #ifdef CONFIG_HIGHMEM
214 EXPORT_SYMBOL(totalram_pages);
216 static char * const zone_names[MAX_NR_ZONES] = {
217 #ifdef CONFIG_ZONE_DMA
220 #ifdef CONFIG_ZONE_DMA32
224 #ifdef CONFIG_HIGHMEM
228 #ifdef CONFIG_ZONE_DEVICE
233 char * const migratetype_names[MIGRATE_TYPES] = {
241 #ifdef CONFIG_MEMORY_ISOLATION
246 compound_page_dtor * const compound_page_dtors[] = {
249 #ifdef CONFIG_HUGETLB_PAGE
252 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
257 int min_free_kbytes = 1024;
258 int user_min_free_kbytes = -1;
259 int watermark_scale_factor = 10;
261 static unsigned long __meminitdata nr_kernel_pages;
262 static unsigned long __meminitdata nr_all_pages;
263 static unsigned long __meminitdata dma_reserve;
265 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
266 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
267 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
268 static unsigned long __initdata required_kernelcore;
269 static unsigned long __initdata required_movablecore;
270 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
271 static bool mirrored_kernelcore;
273 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
275 EXPORT_SYMBOL(movable_zone);
276 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
279 int nr_node_ids __read_mostly = MAX_NUMNODES;
280 int nr_online_nodes __read_mostly = 1;
281 EXPORT_SYMBOL(nr_node_ids);
282 EXPORT_SYMBOL(nr_online_nodes);
285 static DEFINE_LOCAL_IRQ_LOCK(pa_lock);
287 #ifdef CONFIG_PREEMPT_RT_BASE
288 # define cpu_lock_irqsave(cpu, flags) \
289 local_lock_irqsave_on(pa_lock, flags, cpu)
290 # define cpu_unlock_irqrestore(cpu, flags) \
291 local_unlock_irqrestore_on(pa_lock, flags, cpu)
293 # define cpu_lock_irqsave(cpu, flags) local_irq_save(flags)
294 # define cpu_unlock_irqrestore(cpu, flags) local_irq_restore(flags)
297 int page_group_by_mobility_disabled __read_mostly;
299 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
300 static inline void reset_deferred_meminit(pg_data_t *pgdat)
302 pgdat->first_deferred_pfn = ULONG_MAX;
305 /* Returns true if the struct page for the pfn is uninitialised */
306 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
308 int nid = early_pfn_to_nid(pfn);
310 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
317 * Returns false when the remaining initialisation should be deferred until
318 * later in the boot cycle when it can be parallelised.
320 static inline bool update_defer_init(pg_data_t *pgdat,
321 unsigned long pfn, unsigned long zone_end,
322 unsigned long *nr_initialised)
324 unsigned long max_initialise;
326 /* Always populate low zones for address-contrained allocations */
327 if (zone_end < pgdat_end_pfn(pgdat))
330 * Initialise at least 2G of a node but also take into account that
331 * two large system hashes that can take up 1GB for 0.25TB/node.
333 max_initialise = max(2UL << (30 - PAGE_SHIFT),
334 (pgdat->node_spanned_pages >> 8));
337 if ((*nr_initialised > max_initialise) &&
338 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
339 pgdat->first_deferred_pfn = pfn;
346 static inline void reset_deferred_meminit(pg_data_t *pgdat)
350 static inline bool early_page_uninitialised(unsigned long pfn)
355 static inline bool update_defer_init(pg_data_t *pgdat,
356 unsigned long pfn, unsigned long zone_end,
357 unsigned long *nr_initialised)
363 /* Return a pointer to the bitmap storing bits affecting a block of pages */
364 static inline unsigned long *get_pageblock_bitmap(struct page *page,
367 #ifdef CONFIG_SPARSEMEM
368 return __pfn_to_section(pfn)->pageblock_flags;
370 return page_zone(page)->pageblock_flags;
371 #endif /* CONFIG_SPARSEMEM */
374 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
376 #ifdef CONFIG_SPARSEMEM
377 pfn &= (PAGES_PER_SECTION-1);
378 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
380 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
381 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
382 #endif /* CONFIG_SPARSEMEM */
386 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
387 * @page: The page within the block of interest
388 * @pfn: The target page frame number
389 * @end_bitidx: The last bit of interest to retrieve
390 * @mask: mask of bits that the caller is interested in
392 * Return: pageblock_bits flags
394 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
396 unsigned long end_bitidx,
399 unsigned long *bitmap;
400 unsigned long bitidx, word_bitidx;
403 bitmap = get_pageblock_bitmap(page, pfn);
404 bitidx = pfn_to_bitidx(page, pfn);
405 word_bitidx = bitidx / BITS_PER_LONG;
406 bitidx &= (BITS_PER_LONG-1);
408 word = bitmap[word_bitidx];
409 bitidx += end_bitidx;
410 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
413 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
414 unsigned long end_bitidx,
417 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
420 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
422 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
426 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
427 * @page: The page within the block of interest
428 * @flags: The flags to set
429 * @pfn: The target page frame number
430 * @end_bitidx: The last bit of interest
431 * @mask: mask of bits that the caller is interested in
433 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
435 unsigned long end_bitidx,
438 unsigned long *bitmap;
439 unsigned long bitidx, word_bitidx;
440 unsigned long old_word, word;
442 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
444 bitmap = get_pageblock_bitmap(page, pfn);
445 bitidx = pfn_to_bitidx(page, pfn);
446 word_bitidx = bitidx / BITS_PER_LONG;
447 bitidx &= (BITS_PER_LONG-1);
449 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
451 bitidx += end_bitidx;
452 mask <<= (BITS_PER_LONG - bitidx - 1);
453 flags <<= (BITS_PER_LONG - bitidx - 1);
455 word = READ_ONCE(bitmap[word_bitidx]);
457 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
458 if (word == old_word)
464 void set_pageblock_migratetype(struct page *page, int migratetype)
466 if (unlikely(page_group_by_mobility_disabled &&
467 migratetype < MIGRATE_PCPTYPES))
468 migratetype = MIGRATE_UNMOVABLE;
470 set_pageblock_flags_group(page, (unsigned long)migratetype,
471 PB_migrate, PB_migrate_end);
474 #ifdef CONFIG_DEBUG_VM
475 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
479 unsigned long pfn = page_to_pfn(page);
480 unsigned long sp, start_pfn;
483 seq = zone_span_seqbegin(zone);
484 start_pfn = zone->zone_start_pfn;
485 sp = zone->spanned_pages;
486 if (!zone_spans_pfn(zone, pfn))
488 } while (zone_span_seqretry(zone, seq));
491 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
492 pfn, zone_to_nid(zone), zone->name,
493 start_pfn, start_pfn + sp);
498 static int page_is_consistent(struct zone *zone, struct page *page)
500 if (!pfn_valid_within(page_to_pfn(page)))
502 if (zone != page_zone(page))
508 * Temporary debugging check for pages not lying within a given zone.
510 static int bad_range(struct zone *zone, struct page *page)
512 if (page_outside_zone_boundaries(zone, page))
514 if (!page_is_consistent(zone, page))
520 static inline int bad_range(struct zone *zone, struct page *page)
526 static void bad_page(struct page *page, const char *reason,
527 unsigned long bad_flags)
529 static unsigned long resume;
530 static unsigned long nr_shown;
531 static unsigned long nr_unshown;
534 * Allow a burst of 60 reports, then keep quiet for that minute;
535 * or allow a steady drip of one report per second.
537 if (nr_shown == 60) {
538 if (time_before(jiffies, resume)) {
544 "BUG: Bad page state: %lu messages suppressed\n",
551 resume = jiffies + 60 * HZ;
553 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
554 current->comm, page_to_pfn(page));
555 __dump_page(page, reason);
556 bad_flags &= page->flags;
558 pr_alert("bad because of flags: %#lx(%pGp)\n",
559 bad_flags, &bad_flags);
560 dump_page_owner(page);
565 /* Leave bad fields for debug, except PageBuddy could make trouble */
566 page_mapcount_reset(page); /* remove PageBuddy */
567 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
571 * Higher-order pages are called "compound pages". They are structured thusly:
573 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
575 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
576 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
578 * The first tail page's ->compound_dtor holds the offset in array of compound
579 * page destructors. See compound_page_dtors.
581 * The first tail page's ->compound_order holds the order of allocation.
582 * This usage means that zero-order pages may not be compound.
585 void free_compound_page(struct page *page)
587 __free_pages_ok(page, compound_order(page));
590 void prep_compound_page(struct page *page, unsigned int order)
593 int nr_pages = 1 << order;
595 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
596 set_compound_order(page, order);
598 for (i = 1; i < nr_pages; i++) {
599 struct page *p = page + i;
600 set_page_count(p, 0);
601 p->mapping = TAIL_MAPPING;
602 set_compound_head(p, page);
604 atomic_set(compound_mapcount_ptr(page), -1);
607 #ifdef CONFIG_DEBUG_PAGEALLOC
608 unsigned int _debug_guardpage_minorder;
609 bool _debug_pagealloc_enabled __read_mostly
610 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
611 EXPORT_SYMBOL(_debug_pagealloc_enabled);
612 bool _debug_guardpage_enabled __read_mostly;
614 static int __init early_debug_pagealloc(char *buf)
618 return kstrtobool(buf, &_debug_pagealloc_enabled);
620 early_param("debug_pagealloc", early_debug_pagealloc);
622 static bool need_debug_guardpage(void)
624 /* If we don't use debug_pagealloc, we don't need guard page */
625 if (!debug_pagealloc_enabled())
628 if (!debug_guardpage_minorder())
634 static void init_debug_guardpage(void)
636 if (!debug_pagealloc_enabled())
639 if (!debug_guardpage_minorder())
642 _debug_guardpage_enabled = true;
645 struct page_ext_operations debug_guardpage_ops = {
646 .need = need_debug_guardpage,
647 .init = init_debug_guardpage,
650 static int __init debug_guardpage_minorder_setup(char *buf)
654 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
655 pr_err("Bad debug_guardpage_minorder value\n");
658 _debug_guardpage_minorder = res;
659 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
662 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
664 static inline bool set_page_guard(struct zone *zone, struct page *page,
665 unsigned int order, int migratetype)
667 struct page_ext *page_ext;
669 if (!debug_guardpage_enabled())
672 if (order >= debug_guardpage_minorder())
675 page_ext = lookup_page_ext(page);
676 if (unlikely(!page_ext))
679 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
681 INIT_LIST_HEAD(&page->lru);
682 set_page_private(page, order);
683 /* Guard pages are not available for any usage */
684 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
689 static inline void clear_page_guard(struct zone *zone, struct page *page,
690 unsigned int order, int migratetype)
692 struct page_ext *page_ext;
694 if (!debug_guardpage_enabled())
697 page_ext = lookup_page_ext(page);
698 if (unlikely(!page_ext))
701 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
703 set_page_private(page, 0);
704 if (!is_migrate_isolate(migratetype))
705 __mod_zone_freepage_state(zone, (1 << order), migratetype);
708 struct page_ext_operations debug_guardpage_ops;
709 static inline bool set_page_guard(struct zone *zone, struct page *page,
710 unsigned int order, int migratetype) { return false; }
711 static inline void clear_page_guard(struct zone *zone, struct page *page,
712 unsigned int order, int migratetype) {}
715 static inline void set_page_order(struct page *page, unsigned int order)
717 set_page_private(page, order);
718 __SetPageBuddy(page);
721 static inline void rmv_page_order(struct page *page)
723 __ClearPageBuddy(page);
724 set_page_private(page, 0);
728 * This function checks whether a page is free && is the buddy
729 * we can do coalesce a page and its buddy if
730 * (a) the buddy is not in a hole &&
731 * (b) the buddy is in the buddy system &&
732 * (c) a page and its buddy have the same order &&
733 * (d) a page and its buddy are in the same zone.
735 * For recording whether a page is in the buddy system, we set ->_mapcount
736 * PAGE_BUDDY_MAPCOUNT_VALUE.
737 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
738 * serialized by zone->lock.
740 * For recording page's order, we use page_private(page).
742 static inline int page_is_buddy(struct page *page, struct page *buddy,
745 if (!pfn_valid_within(page_to_pfn(buddy)))
748 if (page_is_guard(buddy) && page_order(buddy) == order) {
749 if (page_zone_id(page) != page_zone_id(buddy))
752 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
757 if (PageBuddy(buddy) && page_order(buddy) == order) {
759 * zone check is done late to avoid uselessly
760 * calculating zone/node ids for pages that could
763 if (page_zone_id(page) != page_zone_id(buddy))
766 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
774 * Freeing function for a buddy system allocator.
776 * The concept of a buddy system is to maintain direct-mapped table
777 * (containing bit values) for memory blocks of various "orders".
778 * The bottom level table contains the map for the smallest allocatable
779 * units of memory (here, pages), and each level above it describes
780 * pairs of units from the levels below, hence, "buddies".
781 * At a high level, all that happens here is marking the table entry
782 * at the bottom level available, and propagating the changes upward
783 * as necessary, plus some accounting needed to play nicely with other
784 * parts of the VM system.
785 * At each level, we keep a list of pages, which are heads of continuous
786 * free pages of length of (1 << order) and marked with _mapcount
787 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
789 * So when we are allocating or freeing one, we can derive the state of the
790 * other. That is, if we allocate a small block, and both were
791 * free, the remainder of the region must be split into blocks.
792 * If a block is freed, and its buddy is also free, then this
793 * triggers coalescing into a block of larger size.
798 static inline void __free_one_page(struct page *page,
800 struct zone *zone, unsigned int order,
803 unsigned long page_idx;
804 unsigned long combined_idx;
805 unsigned long uninitialized_var(buddy_idx);
807 unsigned int max_order;
809 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
811 VM_BUG_ON(!zone_is_initialized(zone));
812 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
814 VM_BUG_ON(migratetype == -1);
815 if (likely(!is_migrate_isolate(migratetype)))
816 __mod_zone_freepage_state(zone, 1 << order, migratetype);
818 page_idx = pfn & ((1 << MAX_ORDER) - 1);
820 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
821 VM_BUG_ON_PAGE(bad_range(zone, page), page);
824 while (order < max_order - 1) {
825 buddy_idx = __find_buddy_index(page_idx, order);
826 buddy = page + (buddy_idx - page_idx);
827 if (!page_is_buddy(page, buddy, order))
830 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
831 * merge with it and move up one order.
833 if (page_is_guard(buddy)) {
834 clear_page_guard(zone, buddy, order, migratetype);
836 list_del(&buddy->lru);
837 zone->free_area[order].nr_free--;
838 rmv_page_order(buddy);
840 combined_idx = buddy_idx & page_idx;
841 page = page + (combined_idx - page_idx);
842 page_idx = combined_idx;
845 if (max_order < MAX_ORDER) {
846 /* If we are here, it means order is >= pageblock_order.
847 * We want to prevent merge between freepages on isolate
848 * pageblock and normal pageblock. Without this, pageblock
849 * isolation could cause incorrect freepage or CMA accounting.
851 * We don't want to hit this code for the more frequent
854 if (unlikely(has_isolate_pageblock(zone))) {
857 buddy_idx = __find_buddy_index(page_idx, order);
858 buddy = page + (buddy_idx - page_idx);
859 buddy_mt = get_pageblock_migratetype(buddy);
861 if (migratetype != buddy_mt
862 && (is_migrate_isolate(migratetype) ||
863 is_migrate_isolate(buddy_mt)))
867 goto continue_merging;
871 set_page_order(page, order);
874 * If this is not the largest possible page, check if the buddy
875 * of the next-highest order is free. If it is, it's possible
876 * that pages are being freed that will coalesce soon. In case,
877 * that is happening, add the free page to the tail of the list
878 * so it's less likely to be used soon and more likely to be merged
879 * as a higher order page
881 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
882 struct page *higher_page, *higher_buddy;
883 combined_idx = buddy_idx & page_idx;
884 higher_page = page + (combined_idx - page_idx);
885 buddy_idx = __find_buddy_index(combined_idx, order + 1);
886 higher_buddy = higher_page + (buddy_idx - combined_idx);
887 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
888 list_add_tail(&page->lru,
889 &zone->free_area[order].free_list[migratetype]);
894 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
896 zone->free_area[order].nr_free++;
900 * A bad page could be due to a number of fields. Instead of multiple branches,
901 * try and check multiple fields with one check. The caller must do a detailed
902 * check if necessary.
904 static inline bool page_expected_state(struct page *page,
905 unsigned long check_flags)
907 if (unlikely(atomic_read(&page->_mapcount) != -1))
910 if (unlikely((unsigned long)page->mapping |
911 page_ref_count(page) |
913 (unsigned long)page->mem_cgroup |
915 (page->flags & check_flags)))
921 static void free_pages_check_bad(struct page *page)
923 const char *bad_reason;
924 unsigned long bad_flags;
929 if (unlikely(atomic_read(&page->_mapcount) != -1))
930 bad_reason = "nonzero mapcount";
931 if (unlikely(page->mapping != NULL))
932 bad_reason = "non-NULL mapping";
933 if (unlikely(page_ref_count(page) != 0))
934 bad_reason = "nonzero _refcount";
935 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
936 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
937 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
940 if (unlikely(page->mem_cgroup))
941 bad_reason = "page still charged to cgroup";
943 bad_page(page, bad_reason, bad_flags);
946 static inline int free_pages_check(struct page *page)
948 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
951 /* Something has gone sideways, find it */
952 free_pages_check_bad(page);
956 static int free_tail_pages_check(struct page *head_page, struct page *page)
961 * We rely page->lru.next never has bit 0 set, unless the page
962 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
964 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
966 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
970 switch (page - head_page) {
972 /* the first tail page: ->mapping is compound_mapcount() */
973 if (unlikely(compound_mapcount(page))) {
974 bad_page(page, "nonzero compound_mapcount", 0);
980 * the second tail page: ->mapping is
981 * page_deferred_list().next -- ignore value.
985 if (page->mapping != TAIL_MAPPING) {
986 bad_page(page, "corrupted mapping in tail page", 0);
991 if (unlikely(!PageTail(page))) {
992 bad_page(page, "PageTail not set", 0);
995 if (unlikely(compound_head(page) != head_page)) {
996 bad_page(page, "compound_head not consistent", 0);
1001 page->mapping = NULL;
1002 clear_compound_head(page);
1006 static __always_inline bool free_pages_prepare(struct page *page,
1007 unsigned int order, bool check_free)
1011 VM_BUG_ON_PAGE(PageTail(page), page);
1013 trace_mm_page_free(page, order);
1014 kmemcheck_free_shadow(page, order);
1017 * Check tail pages before head page information is cleared to
1018 * avoid checking PageCompound for order-0 pages.
1020 if (unlikely(order)) {
1021 bool compound = PageCompound(page);
1024 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1027 ClearPageDoubleMap(page);
1028 for (i = 1; i < (1 << order); i++) {
1030 bad += free_tail_pages_check(page, page + i);
1031 if (unlikely(free_pages_check(page + i))) {
1035 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1038 if (PageMappingFlags(page))
1039 page->mapping = NULL;
1040 if (memcg_kmem_enabled() && PageKmemcg(page))
1041 memcg_kmem_uncharge(page, order);
1043 bad += free_pages_check(page);
1047 page_cpupid_reset_last(page);
1048 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 reset_page_owner(page, order);
1051 if (!PageHighMem(page)) {
1052 debug_check_no_locks_freed(page_address(page),
1053 PAGE_SIZE << order);
1054 debug_check_no_obj_freed(page_address(page),
1055 PAGE_SIZE << order);
1057 arch_free_page(page, order);
1058 kernel_poison_pages(page, 1 << order, 0);
1059 kernel_map_pages(page, 1 << order, 0);
1060 kasan_free_pages(page, order);
1065 #ifdef CONFIG_DEBUG_VM
1066 static inline bool free_pcp_prepare(struct page *page)
1068 return free_pages_prepare(page, 0, true);
1071 static inline bool bulkfree_pcp_prepare(struct page *page)
1076 static bool free_pcp_prepare(struct page *page)
1078 return free_pages_prepare(page, 0, false);
1081 static bool bulkfree_pcp_prepare(struct page *page)
1083 return free_pages_check(page);
1085 #endif /* CONFIG_DEBUG_VM */
1088 * Frees a number of pages which have been collected from the pcp lists.
1089 * Assumes all pages on list are in same zone, and of same order.
1090 * count is the number of pages to free.
1092 * If the zone was previously in an "all pages pinned" state then look to
1093 * see if this freeing clears that state.
1095 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1096 * pinned" detection logic.
1098 static void free_pcppages_bulk(struct zone *zone, int count,
1099 struct list_head *list)
1101 unsigned long nr_scanned;
1102 bool isolated_pageblocks;
1103 unsigned long flags;
1105 spin_lock_irqsave(&zone->lock, flags);
1107 isolated_pageblocks = has_isolate_pageblock(zone);
1108 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1110 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1112 while (!list_empty(list)) {
1114 int mt; /* migratetype of the to-be-freed page */
1116 page = list_first_entry(list, struct page, lru);
1117 /* must delete as __free_one_page list manipulates */
1118 list_del(&page->lru);
1120 mt = get_pcppage_migratetype(page);
1121 /* MIGRATE_ISOLATE page should not go to pcplists */
1122 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1123 /* Pageblock could have been isolated meanwhile */
1124 if (unlikely(isolated_pageblocks))
1125 mt = get_pageblock_migratetype(page);
1127 if (bulkfree_pcp_prepare(page))
1130 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1131 trace_mm_page_pcpu_drain(page, 0, mt);
1134 WARN_ON(count != 0);
1135 spin_unlock_irqrestore(&zone->lock, flags);
1139 * Moves a number of pages from the PCP lists to free list which
1140 * is freed outside of the locked region.
1142 * Assumes all pages on list are in same zone, and of same order.
1143 * count is the number of pages to free.
1145 static void isolate_pcp_pages(int count, struct per_cpu_pages *src,
1146 struct list_head *dst)
1148 int migratetype = 0;
1153 struct list_head *list;
1156 * Remove pages from lists in a round-robin fashion. A
1157 * batch_free count is maintained that is incremented when an
1158 * empty list is encountered. This is so more pages are freed
1159 * off fuller lists instead of spinning excessively around empty
1164 if (++migratetype == MIGRATE_PCPTYPES)
1166 list = &src->lists[migratetype];
1167 } while (list_empty(list));
1169 /* This is the only non-empty list. Free them all. */
1170 if (batch_free == MIGRATE_PCPTYPES)
1174 page = list_last_entry(list, struct page, lru);
1175 list_del(&page->lru);
1177 list_add(&page->lru, dst);
1178 } while (--count && --batch_free && !list_empty(list));
1182 static void free_one_page(struct zone *zone,
1183 struct page *page, unsigned long pfn,
1187 unsigned long nr_scanned;
1188 unsigned long flags;
1190 spin_lock_irqsave(&zone->lock, flags);
1191 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1193 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1195 if (unlikely(has_isolate_pageblock(zone) ||
1196 is_migrate_isolate(migratetype))) {
1197 migratetype = get_pfnblock_migratetype(page, pfn);
1199 __free_one_page(page, pfn, zone, order, migratetype);
1200 spin_unlock_irqrestore(&zone->lock, flags);
1203 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1204 unsigned long zone, int nid)
1206 set_page_links(page, zone, nid, pfn);
1207 init_page_count(page);
1208 page_mapcount_reset(page);
1209 page_cpupid_reset_last(page);
1211 INIT_LIST_HEAD(&page->lru);
1212 #ifdef WANT_PAGE_VIRTUAL
1213 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1214 if (!is_highmem_idx(zone))
1215 set_page_address(page, __va(pfn << PAGE_SHIFT));
1219 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1222 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1225 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1226 static void init_reserved_page(unsigned long pfn)
1231 if (!early_page_uninitialised(pfn))
1234 nid = early_pfn_to_nid(pfn);
1235 pgdat = NODE_DATA(nid);
1237 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1238 struct zone *zone = &pgdat->node_zones[zid];
1240 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1243 __init_single_pfn(pfn, zid, nid);
1246 static inline void init_reserved_page(unsigned long pfn)
1249 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1252 * Initialised pages do not have PageReserved set. This function is
1253 * called for each range allocated by the bootmem allocator and
1254 * marks the pages PageReserved. The remaining valid pages are later
1255 * sent to the buddy page allocator.
1257 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1259 unsigned long start_pfn = PFN_DOWN(start);
1260 unsigned long end_pfn = PFN_UP(end);
1262 for (; start_pfn < end_pfn; start_pfn++) {
1263 if (pfn_valid(start_pfn)) {
1264 struct page *page = pfn_to_page(start_pfn);
1266 init_reserved_page(start_pfn);
1268 /* Avoid false-positive PageTail() */
1269 INIT_LIST_HEAD(&page->lru);
1271 SetPageReserved(page);
1276 static void __free_pages_ok(struct page *page, unsigned int order)
1278 unsigned long flags;
1280 unsigned long pfn = page_to_pfn(page);
1282 if (!free_pages_prepare(page, order, true))
1285 migratetype = get_pfnblock_migratetype(page, pfn);
1286 local_lock_irqsave(pa_lock, flags);
1287 __count_vm_events(PGFREE, 1 << order);
1288 free_one_page(page_zone(page), page, pfn, order, migratetype);
1289 local_unlock_irqrestore(pa_lock, flags);
1292 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1294 unsigned int nr_pages = 1 << order;
1295 struct page *p = page;
1299 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1304 __ClearPageReserved(p);
1305 set_page_count(p, 0);
1307 page_zone(page)->managed_pages += nr_pages;
1308 set_page_refcounted(page);
1309 __free_pages(page, order);
1312 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1313 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1315 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1317 int __meminit early_pfn_to_nid(unsigned long pfn)
1319 static DEFINE_SPINLOCK(early_pfn_lock);
1322 spin_lock(&early_pfn_lock);
1323 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1325 nid = first_online_node;
1326 spin_unlock(&early_pfn_lock);
1332 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1333 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1334 struct mminit_pfnnid_cache *state)
1338 nid = __early_pfn_to_nid(pfn, state);
1339 if (nid >= 0 && nid != node)
1344 /* Only safe to use early in boot when initialisation is single-threaded */
1345 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1347 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1352 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1356 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1357 struct mminit_pfnnid_cache *state)
1364 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1367 if (early_page_uninitialised(pfn))
1369 return __free_pages_boot_core(page, order);
1373 * Check that the whole (or subset of) a pageblock given by the interval of
1374 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1375 * with the migration of free compaction scanner. The scanners then need to
1376 * use only pfn_valid_within() check for arches that allow holes within
1379 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1381 * It's possible on some configurations to have a setup like node0 node1 node0
1382 * i.e. it's possible that all pages within a zones range of pages do not
1383 * belong to a single zone. We assume that a border between node0 and node1
1384 * can occur within a single pageblock, but not a node0 node1 node0
1385 * interleaving within a single pageblock. It is therefore sufficient to check
1386 * the first and last page of a pageblock and avoid checking each individual
1387 * page in a pageblock.
1389 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1390 unsigned long end_pfn, struct zone *zone)
1392 struct page *start_page;
1393 struct page *end_page;
1395 /* end_pfn is one past the range we are checking */
1398 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1401 start_page = pfn_to_page(start_pfn);
1403 if (page_zone(start_page) != zone)
1406 end_page = pfn_to_page(end_pfn);
1408 /* This gives a shorter code than deriving page_zone(end_page) */
1409 if (page_zone_id(start_page) != page_zone_id(end_page))
1415 void set_zone_contiguous(struct zone *zone)
1417 unsigned long block_start_pfn = zone->zone_start_pfn;
1418 unsigned long block_end_pfn;
1420 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1421 for (; block_start_pfn < zone_end_pfn(zone);
1422 block_start_pfn = block_end_pfn,
1423 block_end_pfn += pageblock_nr_pages) {
1425 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1427 if (!__pageblock_pfn_to_page(block_start_pfn,
1428 block_end_pfn, zone))
1432 /* We confirm that there is no hole */
1433 zone->contiguous = true;
1436 void clear_zone_contiguous(struct zone *zone)
1438 zone->contiguous = false;
1441 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1442 static void __init deferred_free_range(struct page *page,
1443 unsigned long pfn, int nr_pages)
1450 /* Free a large naturally-aligned chunk if possible */
1451 if (nr_pages == pageblock_nr_pages &&
1452 (pfn & (pageblock_nr_pages - 1)) == 0) {
1453 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1454 __free_pages_boot_core(page, pageblock_order);
1458 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1459 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1460 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1461 __free_pages_boot_core(page, 0);
1465 /* Completion tracking for deferred_init_memmap() threads */
1466 static atomic_t pgdat_init_n_undone __initdata;
1467 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1469 static inline void __init pgdat_init_report_one_done(void)
1471 if (atomic_dec_and_test(&pgdat_init_n_undone))
1472 complete(&pgdat_init_all_done_comp);
1475 /* Initialise remaining memory on a node */
1476 static int __init deferred_init_memmap(void *data)
1478 pg_data_t *pgdat = data;
1479 int nid = pgdat->node_id;
1480 struct mminit_pfnnid_cache nid_init_state = { };
1481 unsigned long start = jiffies;
1482 unsigned long nr_pages = 0;
1483 unsigned long walk_start, walk_end;
1486 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1487 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1489 if (first_init_pfn == ULONG_MAX) {
1490 pgdat_init_report_one_done();
1494 /* Bind memory initialisation thread to a local node if possible */
1495 if (!cpumask_empty(cpumask))
1496 set_cpus_allowed_ptr(current, cpumask);
1498 /* Sanity check boundaries */
1499 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1500 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1501 pgdat->first_deferred_pfn = ULONG_MAX;
1503 /* Only the highest zone is deferred so find it */
1504 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1505 zone = pgdat->node_zones + zid;
1506 if (first_init_pfn < zone_end_pfn(zone))
1510 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1511 unsigned long pfn, end_pfn;
1512 struct page *page = NULL;
1513 struct page *free_base_page = NULL;
1514 unsigned long free_base_pfn = 0;
1517 end_pfn = min(walk_end, zone_end_pfn(zone));
1518 pfn = first_init_pfn;
1519 if (pfn < walk_start)
1521 if (pfn < zone->zone_start_pfn)
1522 pfn = zone->zone_start_pfn;
1524 for (; pfn < end_pfn; pfn++) {
1525 if (!pfn_valid_within(pfn))
1529 * Ensure pfn_valid is checked every
1530 * pageblock_nr_pages for memory holes
1532 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1533 if (!pfn_valid(pfn)) {
1539 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1544 /* Minimise pfn page lookups and scheduler checks */
1545 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1548 nr_pages += nr_to_free;
1549 deferred_free_range(free_base_page,
1550 free_base_pfn, nr_to_free);
1551 free_base_page = NULL;
1552 free_base_pfn = nr_to_free = 0;
1554 page = pfn_to_page(pfn);
1559 VM_BUG_ON(page_zone(page) != zone);
1563 __init_single_page(page, pfn, zid, nid);
1564 if (!free_base_page) {
1565 free_base_page = page;
1566 free_base_pfn = pfn;
1571 /* Where possible, batch up pages for a single free */
1574 /* Free the current block of pages to allocator */
1575 nr_pages += nr_to_free;
1576 deferred_free_range(free_base_page, free_base_pfn,
1578 free_base_page = NULL;
1579 free_base_pfn = nr_to_free = 0;
1581 /* Free the last block of pages to allocator */
1582 nr_pages += nr_to_free;
1583 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1585 first_init_pfn = max(end_pfn, first_init_pfn);
1588 /* Sanity check that the next zone really is unpopulated */
1589 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1591 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1592 jiffies_to_msecs(jiffies - start));
1594 pgdat_init_report_one_done();
1597 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1599 void __init page_alloc_init_late(void)
1603 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1606 /* There will be num_node_state(N_MEMORY) threads */
1607 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1608 for_each_node_state(nid, N_MEMORY) {
1609 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1612 /* Block until all are initialised */
1613 wait_for_completion(&pgdat_init_all_done_comp);
1615 /* Reinit limits that are based on free pages after the kernel is up */
1616 files_maxfiles_init();
1619 for_each_populated_zone(zone)
1620 set_zone_contiguous(zone);
1624 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1625 void __init init_cma_reserved_pageblock(struct page *page)
1627 unsigned i = pageblock_nr_pages;
1628 struct page *p = page;
1631 __ClearPageReserved(p);
1632 set_page_count(p, 0);
1635 set_pageblock_migratetype(page, MIGRATE_CMA);
1637 if (pageblock_order >= MAX_ORDER) {
1638 i = pageblock_nr_pages;
1641 set_page_refcounted(p);
1642 __free_pages(p, MAX_ORDER - 1);
1643 p += MAX_ORDER_NR_PAGES;
1644 } while (i -= MAX_ORDER_NR_PAGES);
1646 set_page_refcounted(page);
1647 __free_pages(page, pageblock_order);
1650 adjust_managed_page_count(page, pageblock_nr_pages);
1655 * The order of subdivision here is critical for the IO subsystem.
1656 * Please do not alter this order without good reasons and regression
1657 * testing. Specifically, as large blocks of memory are subdivided,
1658 * the order in which smaller blocks are delivered depends on the order
1659 * they're subdivided in this function. This is the primary factor
1660 * influencing the order in which pages are delivered to the IO
1661 * subsystem according to empirical testing, and this is also justified
1662 * by considering the behavior of a buddy system containing a single
1663 * large block of memory acted on by a series of small allocations.
1664 * This behavior is a critical factor in sglist merging's success.
1668 static inline void expand(struct zone *zone, struct page *page,
1669 int low, int high, struct free_area *area,
1672 unsigned long size = 1 << high;
1674 while (high > low) {
1678 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1681 * Mark as guard pages (or page), that will allow to
1682 * merge back to allocator when buddy will be freed.
1683 * Corresponding page table entries will not be touched,
1684 * pages will stay not present in virtual address space
1686 if (set_page_guard(zone, &page[size], high, migratetype))
1689 list_add(&page[size].lru, &area->free_list[migratetype]);
1691 set_page_order(&page[size], high);
1695 static void check_new_page_bad(struct page *page)
1697 const char *bad_reason = NULL;
1698 unsigned long bad_flags = 0;
1700 if (unlikely(atomic_read(&page->_mapcount) != -1))
1701 bad_reason = "nonzero mapcount";
1702 if (unlikely(page->mapping != NULL))
1703 bad_reason = "non-NULL mapping";
1704 if (unlikely(page_ref_count(page) != 0))
1705 bad_reason = "nonzero _count";
1706 if (unlikely(page->flags & __PG_HWPOISON)) {
1707 bad_reason = "HWPoisoned (hardware-corrupted)";
1708 bad_flags = __PG_HWPOISON;
1709 /* Don't complain about hwpoisoned pages */
1710 page_mapcount_reset(page); /* remove PageBuddy */
1713 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1714 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1715 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1718 if (unlikely(page->mem_cgroup))
1719 bad_reason = "page still charged to cgroup";
1721 bad_page(page, bad_reason, bad_flags);
1725 * This page is about to be returned from the page allocator
1727 static inline int check_new_page(struct page *page)
1729 if (likely(page_expected_state(page,
1730 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1733 check_new_page_bad(page);
1737 static inline bool free_pages_prezeroed(bool poisoned)
1739 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1740 page_poisoning_enabled() && poisoned;
1743 #ifdef CONFIG_DEBUG_VM
1744 static bool check_pcp_refill(struct page *page)
1749 static bool check_new_pcp(struct page *page)
1751 return check_new_page(page);
1754 static bool check_pcp_refill(struct page *page)
1756 return check_new_page(page);
1758 static bool check_new_pcp(struct page *page)
1762 #endif /* CONFIG_DEBUG_VM */
1764 static bool check_new_pages(struct page *page, unsigned int order)
1767 for (i = 0; i < (1 << order); i++) {
1768 struct page *p = page + i;
1770 if (unlikely(check_new_page(p)))
1777 inline void post_alloc_hook(struct page *page, unsigned int order,
1780 set_page_private(page, 0);
1781 set_page_refcounted(page);
1783 arch_alloc_page(page, order);
1784 kernel_map_pages(page, 1 << order, 1);
1785 kernel_poison_pages(page, 1 << order, 1);
1786 kasan_alloc_pages(page, order);
1787 set_page_owner(page, order, gfp_flags);
1790 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1791 unsigned int alloc_flags)
1794 bool poisoned = true;
1796 for (i = 0; i < (1 << order); i++) {
1797 struct page *p = page + i;
1799 poisoned &= page_is_poisoned(p);
1802 post_alloc_hook(page, order, gfp_flags);
1804 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1805 for (i = 0; i < (1 << order); i++)
1806 clear_highpage(page + i);
1808 if (order && (gfp_flags & __GFP_COMP))
1809 prep_compound_page(page, order);
1812 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1813 * allocate the page. The expectation is that the caller is taking
1814 * steps that will free more memory. The caller should avoid the page
1815 * being used for !PFMEMALLOC purposes.
1817 if (alloc_flags & ALLOC_NO_WATERMARKS)
1818 set_page_pfmemalloc(page);
1820 clear_page_pfmemalloc(page);
1824 * Go through the free lists for the given migratetype and remove
1825 * the smallest available page from the freelists
1828 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1831 unsigned int current_order;
1832 struct free_area *area;
1835 /* Find a page of the appropriate size in the preferred list */
1836 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1837 area = &(zone->free_area[current_order]);
1838 page = list_first_entry_or_null(&area->free_list[migratetype],
1842 list_del(&page->lru);
1843 rmv_page_order(page);
1845 expand(zone, page, order, current_order, area, migratetype);
1846 set_pcppage_migratetype(page, migratetype);
1855 * This array describes the order lists are fallen back to when
1856 * the free lists for the desirable migrate type are depleted
1858 static int fallbacks[MIGRATE_TYPES][4] = {
1859 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1860 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1861 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1863 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1865 #ifdef CONFIG_MEMORY_ISOLATION
1866 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1871 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1874 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1877 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1878 unsigned int order) { return NULL; }
1882 * Move the free pages in a range to the free lists of the requested type.
1883 * Note that start_page and end_pages are not aligned on a pageblock
1884 * boundary. If alignment is required, use move_freepages_block()
1886 int move_freepages(struct zone *zone,
1887 struct page *start_page, struct page *end_page,
1892 int pages_moved = 0;
1894 #ifndef CONFIG_HOLES_IN_ZONE
1896 * page_zone is not safe to call in this context when
1897 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1898 * anyway as we check zone boundaries in move_freepages_block().
1899 * Remove at a later date when no bug reports exist related to
1900 * grouping pages by mobility
1902 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1905 for (page = start_page; page <= end_page;) {
1906 /* Make sure we are not inadvertently changing nodes */
1907 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1909 if (!pfn_valid_within(page_to_pfn(page))) {
1914 if (!PageBuddy(page)) {
1919 order = page_order(page);
1920 list_move(&page->lru,
1921 &zone->free_area[order].free_list[migratetype]);
1923 pages_moved += 1 << order;
1929 int move_freepages_block(struct zone *zone, struct page *page,
1932 unsigned long start_pfn, end_pfn;
1933 struct page *start_page, *end_page;
1935 start_pfn = page_to_pfn(page);
1936 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1937 start_page = pfn_to_page(start_pfn);
1938 end_page = start_page + pageblock_nr_pages - 1;
1939 end_pfn = start_pfn + pageblock_nr_pages - 1;
1941 /* Do not cross zone boundaries */
1942 if (!zone_spans_pfn(zone, start_pfn))
1944 if (!zone_spans_pfn(zone, end_pfn))
1947 return move_freepages(zone, start_page, end_page, migratetype);
1950 static void change_pageblock_range(struct page *pageblock_page,
1951 int start_order, int migratetype)
1953 int nr_pageblocks = 1 << (start_order - pageblock_order);
1955 while (nr_pageblocks--) {
1956 set_pageblock_migratetype(pageblock_page, migratetype);
1957 pageblock_page += pageblock_nr_pages;
1962 * When we are falling back to another migratetype during allocation, try to
1963 * steal extra free pages from the same pageblocks to satisfy further
1964 * allocations, instead of polluting multiple pageblocks.
1966 * If we are stealing a relatively large buddy page, it is likely there will
1967 * be more free pages in the pageblock, so try to steal them all. For
1968 * reclaimable and unmovable allocations, we steal regardless of page size,
1969 * as fragmentation caused by those allocations polluting movable pageblocks
1970 * is worse than movable allocations stealing from unmovable and reclaimable
1973 static bool can_steal_fallback(unsigned int order, int start_mt)
1976 * Leaving this order check is intended, although there is
1977 * relaxed order check in next check. The reason is that
1978 * we can actually steal whole pageblock if this condition met,
1979 * but, below check doesn't guarantee it and that is just heuristic
1980 * so could be changed anytime.
1982 if (order >= pageblock_order)
1985 if (order >= pageblock_order / 2 ||
1986 start_mt == MIGRATE_RECLAIMABLE ||
1987 start_mt == MIGRATE_UNMOVABLE ||
1988 page_group_by_mobility_disabled)
1995 * This function implements actual steal behaviour. If order is large enough,
1996 * we can steal whole pageblock. If not, we first move freepages in this
1997 * pageblock and check whether half of pages are moved or not. If half of
1998 * pages are moved, we can change migratetype of pageblock and permanently
1999 * use it's pages as requested migratetype in the future.
2001 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2004 unsigned int current_order = page_order(page);
2007 /* Take ownership for orders >= pageblock_order */
2008 if (current_order >= pageblock_order) {
2009 change_pageblock_range(page, current_order, start_type);
2013 pages = move_freepages_block(zone, page, start_type);
2015 /* Claim the whole block if over half of it is free */
2016 if (pages >= (1 << (pageblock_order-1)) ||
2017 page_group_by_mobility_disabled)
2018 set_pageblock_migratetype(page, start_type);
2022 * Check whether there is a suitable fallback freepage with requested order.
2023 * If only_stealable is true, this function returns fallback_mt only if
2024 * we can steal other freepages all together. This would help to reduce
2025 * fragmentation due to mixed migratetype pages in one pageblock.
2027 int find_suitable_fallback(struct free_area *area, unsigned int order,
2028 int migratetype, bool only_stealable, bool *can_steal)
2033 if (area->nr_free == 0)
2038 fallback_mt = fallbacks[migratetype][i];
2039 if (fallback_mt == MIGRATE_TYPES)
2042 if (list_empty(&area->free_list[fallback_mt]))
2045 if (can_steal_fallback(order, migratetype))
2048 if (!only_stealable)
2059 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2060 * there are no empty page blocks that contain a page with a suitable order
2062 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2063 unsigned int alloc_order)
2066 unsigned long max_managed, flags;
2069 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2070 * Check is race-prone but harmless.
2072 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2073 if (zone->nr_reserved_highatomic >= max_managed)
2076 spin_lock_irqsave(&zone->lock, flags);
2078 /* Recheck the nr_reserved_highatomic limit under the lock */
2079 if (zone->nr_reserved_highatomic >= max_managed)
2083 mt = get_pageblock_migratetype(page);
2084 if (mt != MIGRATE_HIGHATOMIC &&
2085 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2086 zone->nr_reserved_highatomic += pageblock_nr_pages;
2087 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2088 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2092 spin_unlock_irqrestore(&zone->lock, flags);
2096 * Used when an allocation is about to fail under memory pressure. This
2097 * potentially hurts the reliability of high-order allocations when under
2098 * intense memory pressure but failed atomic allocations should be easier
2099 * to recover from than an OOM.
2101 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2103 struct zonelist *zonelist = ac->zonelist;
2104 unsigned long flags;
2110 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2112 /* Preserve at least one pageblock */
2113 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2116 spin_lock_irqsave(&zone->lock, flags);
2117 for (order = 0; order < MAX_ORDER; order++) {
2118 struct free_area *area = &(zone->free_area[order]);
2120 page = list_first_entry_or_null(
2121 &area->free_list[MIGRATE_HIGHATOMIC],
2127 * It should never happen but changes to locking could
2128 * inadvertently allow a per-cpu drain to add pages
2129 * to MIGRATE_HIGHATOMIC while unreserving so be safe
2130 * and watch for underflows.
2132 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2133 zone->nr_reserved_highatomic);
2136 * Convert to ac->migratetype and avoid the normal
2137 * pageblock stealing heuristics. Minimally, the caller
2138 * is doing the work and needs the pages. More
2139 * importantly, if the block was always converted to
2140 * MIGRATE_UNMOVABLE or another type then the number
2141 * of pageblocks that cannot be completely freed
2144 set_pageblock_migratetype(page, ac->migratetype);
2145 move_freepages_block(zone, page, ac->migratetype);
2146 spin_unlock_irqrestore(&zone->lock, flags);
2149 spin_unlock_irqrestore(&zone->lock, flags);
2153 /* Remove an element from the buddy allocator from the fallback list */
2154 static inline struct page *
2155 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2157 struct free_area *area;
2158 unsigned int current_order;
2163 /* Find the largest possible block of pages in the other list */
2164 for (current_order = MAX_ORDER-1;
2165 current_order >= order && current_order <= MAX_ORDER-1;
2167 area = &(zone->free_area[current_order]);
2168 fallback_mt = find_suitable_fallback(area, current_order,
2169 start_migratetype, false, &can_steal);
2170 if (fallback_mt == -1)
2173 page = list_first_entry(&area->free_list[fallback_mt],
2176 steal_suitable_fallback(zone, page, start_migratetype);
2178 /* Remove the page from the freelists */
2180 list_del(&page->lru);
2181 rmv_page_order(page);
2183 expand(zone, page, order, current_order, area,
2186 * The pcppage_migratetype may differ from pageblock's
2187 * migratetype depending on the decisions in
2188 * find_suitable_fallback(). This is OK as long as it does not
2189 * differ for MIGRATE_CMA pageblocks. Those can be used as
2190 * fallback only via special __rmqueue_cma_fallback() function
2192 set_pcppage_migratetype(page, start_migratetype);
2194 trace_mm_page_alloc_extfrag(page, order, current_order,
2195 start_migratetype, fallback_mt);
2204 * Do the hard work of removing an element from the buddy allocator.
2205 * Call me with the zone->lock already held.
2207 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2212 page = __rmqueue_smallest(zone, order, migratetype);
2213 if (unlikely(!page)) {
2214 if (migratetype == MIGRATE_MOVABLE)
2215 page = __rmqueue_cma_fallback(zone, order);
2218 page = __rmqueue_fallback(zone, order, migratetype);
2221 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2226 * Obtain a specified number of elements from the buddy allocator, all under
2227 * a single hold of the lock, for efficiency. Add them to the supplied list.
2228 * Returns the number of new pages which were placed at *list.
2230 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2231 unsigned long count, struct list_head *list,
2232 int migratetype, bool cold)
2236 spin_lock(&zone->lock);
2237 for (i = 0; i < count; ++i) {
2238 struct page *page = __rmqueue(zone, order, migratetype);
2239 if (unlikely(page == NULL))
2242 if (unlikely(check_pcp_refill(page)))
2246 * Split buddy pages returned by expand() are received here
2247 * in physical page order. The page is added to the callers and
2248 * list and the list head then moves forward. From the callers
2249 * perspective, the linked list is ordered by page number in
2250 * some conditions. This is useful for IO devices that can
2251 * merge IO requests if the physical pages are ordered
2255 list_add(&page->lru, list);
2257 list_add_tail(&page->lru, list);
2259 if (is_migrate_cma(get_pcppage_migratetype(page)))
2260 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2263 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2264 spin_unlock(&zone->lock);
2270 * Called from the vmstat counter updater to drain pagesets of this
2271 * currently executing processor on remote nodes after they have
2274 * Note that this function must be called with the thread pinned to
2275 * a single processor.
2277 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2279 unsigned long flags;
2281 int to_drain, batch;
2283 local_lock_irqsave(pa_lock, flags);
2284 batch = READ_ONCE(pcp->batch);
2285 to_drain = min(pcp->count, batch);
2287 isolate_pcp_pages(to_drain, pcp, &dst);
2288 pcp->count -= to_drain;
2290 local_unlock_irqrestore(pa_lock, flags);
2291 free_pcppages_bulk(zone, to_drain, &dst);
2296 * Drain pcplists of the indicated processor and zone.
2298 * The processor must either be the current processor and the
2299 * thread pinned to the current processor or a processor that
2302 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2304 unsigned long flags;
2305 struct per_cpu_pageset *pset;
2306 struct per_cpu_pages *pcp;
2310 cpu_lock_irqsave(cpu, flags);
2311 pset = per_cpu_ptr(zone->pageset, cpu);
2316 isolate_pcp_pages(count, pcp, &dst);
2319 cpu_unlock_irqrestore(cpu, flags);
2321 free_pcppages_bulk(zone, count, &dst);
2325 * Drain pcplists of all zones on the indicated processor.
2327 * The processor must either be the current processor and the
2328 * thread pinned to the current processor or a processor that
2331 static void drain_pages(unsigned int cpu)
2335 for_each_populated_zone(zone) {
2336 drain_pages_zone(cpu, zone);
2341 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2343 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2344 * the single zone's pages.
2346 void drain_local_pages(struct zone *zone)
2348 int cpu = smp_processor_id();
2351 drain_pages_zone(cpu, zone);
2357 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2359 * When zone parameter is non-NULL, spill just the single zone's pages.
2361 * Note that this code is protected against sending an IPI to an offline
2362 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2363 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2364 * nothing keeps CPUs from showing up after we populated the cpumask and
2365 * before the call to on_each_cpu_mask().
2367 void drain_all_pages(struct zone *zone)
2372 * Allocate in the BSS so we wont require allocation in
2373 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2375 static cpumask_t cpus_with_pcps;
2378 * We don't care about racing with CPU hotplug event
2379 * as offline notification will cause the notified
2380 * cpu to drain that CPU pcps and on_each_cpu_mask
2381 * disables preemption as part of its processing
2383 for_each_online_cpu(cpu) {
2384 struct per_cpu_pageset *pcp;
2386 bool has_pcps = false;
2389 pcp = per_cpu_ptr(zone->pageset, cpu);
2393 for_each_populated_zone(z) {
2394 pcp = per_cpu_ptr(z->pageset, cpu);
2395 if (pcp->pcp.count) {
2403 cpumask_set_cpu(cpu, &cpus_with_pcps);
2405 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2407 #ifndef CONFIG_PREEMPT_RT_BASE
2408 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2411 for_each_cpu(cpu, &cpus_with_pcps) {
2413 drain_pages_zone(cpu, zone);
2420 #ifdef CONFIG_HIBERNATION
2422 void mark_free_pages(struct zone *zone)
2424 unsigned long pfn, max_zone_pfn;
2425 unsigned long flags;
2426 unsigned int order, t;
2429 if (zone_is_empty(zone))
2432 spin_lock_irqsave(&zone->lock, flags);
2434 max_zone_pfn = zone_end_pfn(zone);
2435 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2436 if (pfn_valid(pfn)) {
2437 page = pfn_to_page(pfn);
2439 if (page_zone(page) != zone)
2442 if (!swsusp_page_is_forbidden(page))
2443 swsusp_unset_page_free(page);
2446 for_each_migratetype_order(order, t) {
2447 list_for_each_entry(page,
2448 &zone->free_area[order].free_list[t], lru) {
2451 pfn = page_to_pfn(page);
2452 for (i = 0; i < (1UL << order); i++)
2453 swsusp_set_page_free(pfn_to_page(pfn + i));
2456 spin_unlock_irqrestore(&zone->lock, flags);
2458 #endif /* CONFIG_PM */
2461 * Free a 0-order page
2462 * cold == true ? free a cold page : free a hot page
2464 void free_hot_cold_page(struct page *page, bool cold)
2466 struct zone *zone = page_zone(page);
2467 struct per_cpu_pages *pcp;
2468 unsigned long flags;
2469 unsigned long pfn = page_to_pfn(page);
2472 if (!free_pcp_prepare(page))
2475 migratetype = get_pfnblock_migratetype(page, pfn);
2476 set_pcppage_migratetype(page, migratetype);
2477 local_lock_irqsave(pa_lock, flags);
2478 __count_vm_event(PGFREE);
2481 * We only track unmovable, reclaimable and movable on pcp lists.
2482 * Free ISOLATE pages back to the allocator because they are being
2483 * offlined but treat RESERVE as movable pages so we can get those
2484 * areas back if necessary. Otherwise, we may have to free
2485 * excessively into the page allocator
2487 if (migratetype >= MIGRATE_PCPTYPES) {
2488 if (unlikely(is_migrate_isolate(migratetype))) {
2489 free_one_page(zone, page, pfn, 0, migratetype);
2492 migratetype = MIGRATE_MOVABLE;
2495 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2497 list_add(&page->lru, &pcp->lists[migratetype]);
2499 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2501 if (pcp->count >= pcp->high) {
2502 unsigned long batch = READ_ONCE(pcp->batch);
2505 isolate_pcp_pages(batch, pcp, &dst);
2506 pcp->count -= batch;
2507 local_unlock_irqrestore(pa_lock, flags);
2508 free_pcppages_bulk(zone, batch, &dst);
2513 local_unlock_irqrestore(pa_lock, flags);
2517 * Free a list of 0-order pages
2519 void free_hot_cold_page_list(struct list_head *list, bool cold)
2521 struct page *page, *next;
2523 list_for_each_entry_safe(page, next, list, lru) {
2524 trace_mm_page_free_batched(page, cold);
2525 free_hot_cold_page(page, cold);
2530 * split_page takes a non-compound higher-order page, and splits it into
2531 * n (1<<order) sub-pages: page[0..n]
2532 * Each sub-page must be freed individually.
2534 * Note: this is probably too low level an operation for use in drivers.
2535 * Please consult with lkml before using this in your driver.
2537 void split_page(struct page *page, unsigned int order)
2541 VM_BUG_ON_PAGE(PageCompound(page), page);
2542 VM_BUG_ON_PAGE(!page_count(page), page);
2544 #ifdef CONFIG_KMEMCHECK
2546 * Split shadow pages too, because free(page[0]) would
2547 * otherwise free the whole shadow.
2549 if (kmemcheck_page_is_tracked(page))
2550 split_page(virt_to_page(page[0].shadow), order);
2553 for (i = 1; i < (1 << order); i++)
2554 set_page_refcounted(page + i);
2555 split_page_owner(page, order);
2557 EXPORT_SYMBOL_GPL(split_page);
2559 int __isolate_free_page(struct page *page, unsigned int order)
2561 unsigned long watermark;
2565 BUG_ON(!PageBuddy(page));
2567 zone = page_zone(page);
2568 mt = get_pageblock_migratetype(page);
2570 if (!is_migrate_isolate(mt)) {
2572 * Obey watermarks as if the page was being allocated. We can
2573 * emulate a high-order watermark check with a raised order-0
2574 * watermark, because we already know our high-order page
2577 watermark = min_wmark_pages(zone) + (1UL << order);
2578 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2581 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2584 /* Remove page from free list */
2585 list_del(&page->lru);
2586 zone->free_area[order].nr_free--;
2587 rmv_page_order(page);
2590 * Set the pageblock if the isolated page is at least half of a
2593 if (order >= pageblock_order - 1) {
2594 struct page *endpage = page + (1 << order) - 1;
2595 for (; page < endpage; page += pageblock_nr_pages) {
2596 int mt = get_pageblock_migratetype(page);
2597 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2598 set_pageblock_migratetype(page,
2604 return 1UL << order;
2608 * Update NUMA hit/miss statistics
2610 * Must be called with interrupts disabled.
2612 * When __GFP_OTHER_NODE is set assume the node of the preferred
2613 * zone is the local node. This is useful for daemons who allocate
2614 * memory on behalf of other processes.
2616 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2620 int local_nid = numa_node_id();
2621 enum zone_stat_item local_stat = NUMA_LOCAL;
2623 if (unlikely(flags & __GFP_OTHER_NODE)) {
2624 local_stat = NUMA_OTHER;
2625 local_nid = preferred_zone->node;
2628 if (z->node == local_nid) {
2629 __inc_zone_state(z, NUMA_HIT);
2630 __inc_zone_state(z, local_stat);
2632 __inc_zone_state(z, NUMA_MISS);
2633 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2639 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2642 struct page *buffered_rmqueue(struct zone *preferred_zone,
2643 struct zone *zone, unsigned int order,
2644 gfp_t gfp_flags, unsigned int alloc_flags,
2647 unsigned long flags;
2649 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2651 if (likely(order == 0)) {
2652 struct per_cpu_pages *pcp;
2653 struct list_head *list;
2655 local_lock_irqsave(pa_lock, flags);
2657 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2658 list = &pcp->lists[migratetype];
2659 if (list_empty(list)) {
2660 pcp->count += rmqueue_bulk(zone, 0,
2663 if (unlikely(list_empty(list)))
2668 page = list_last_entry(list, struct page, lru);
2670 page = list_first_entry(list, struct page, lru);
2672 list_del(&page->lru);
2675 } while (check_new_pcp(page));
2678 * We most definitely don't want callers attempting to
2679 * allocate greater than order-1 page units with __GFP_NOFAIL.
2681 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2682 local_spin_lock_irqsave(pa_lock, &zone->lock, flags);
2686 if (alloc_flags & ALLOC_HARDER) {
2687 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2689 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2692 page = __rmqueue(zone, order, migratetype);
2693 } while (page && check_new_pages(page, order));
2695 spin_unlock(&zone->lock);
2698 __mod_zone_freepage_state(zone, -(1 << order),
2699 get_pcppage_migratetype(page));
2700 spin_unlock(&zone->lock);
2703 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2704 zone_statistics(preferred_zone, zone, gfp_flags);
2705 local_unlock_irqrestore(pa_lock, flags);
2707 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2711 local_unlock_irqrestore(pa_lock, flags);
2715 #ifdef CONFIG_FAIL_PAGE_ALLOC
2718 struct fault_attr attr;
2720 bool ignore_gfp_highmem;
2721 bool ignore_gfp_reclaim;
2723 } fail_page_alloc = {
2724 .attr = FAULT_ATTR_INITIALIZER,
2725 .ignore_gfp_reclaim = true,
2726 .ignore_gfp_highmem = true,
2730 static int __init setup_fail_page_alloc(char *str)
2732 return setup_fault_attr(&fail_page_alloc.attr, str);
2734 __setup("fail_page_alloc=", setup_fail_page_alloc);
2736 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2738 if (order < fail_page_alloc.min_order)
2740 if (gfp_mask & __GFP_NOFAIL)
2742 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2744 if (fail_page_alloc.ignore_gfp_reclaim &&
2745 (gfp_mask & __GFP_DIRECT_RECLAIM))
2748 return should_fail(&fail_page_alloc.attr, 1 << order);
2751 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2753 static int __init fail_page_alloc_debugfs(void)
2755 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2758 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2759 &fail_page_alloc.attr);
2761 return PTR_ERR(dir);
2763 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2764 &fail_page_alloc.ignore_gfp_reclaim))
2766 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2767 &fail_page_alloc.ignore_gfp_highmem))
2769 if (!debugfs_create_u32("min-order", mode, dir,
2770 &fail_page_alloc.min_order))
2775 debugfs_remove_recursive(dir);
2780 late_initcall(fail_page_alloc_debugfs);
2782 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2784 #else /* CONFIG_FAIL_PAGE_ALLOC */
2786 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2791 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2794 * Return true if free base pages are above 'mark'. For high-order checks it
2795 * will return true of the order-0 watermark is reached and there is at least
2796 * one free page of a suitable size. Checking now avoids taking the zone lock
2797 * to check in the allocation paths if no pages are free.
2799 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2800 int classzone_idx, unsigned int alloc_flags,
2805 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2807 /* free_pages may go negative - that's OK */
2808 free_pages -= (1 << order) - 1;
2810 if (alloc_flags & ALLOC_HIGH)
2814 * If the caller does not have rights to ALLOC_HARDER then subtract
2815 * the high-atomic reserves. This will over-estimate the size of the
2816 * atomic reserve but it avoids a search.
2818 if (likely(!alloc_harder))
2819 free_pages -= z->nr_reserved_highatomic;
2824 /* If allocation can't use CMA areas don't use free CMA pages */
2825 if (!(alloc_flags & ALLOC_CMA))
2826 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2830 * Check watermarks for an order-0 allocation request. If these
2831 * are not met, then a high-order request also cannot go ahead
2832 * even if a suitable page happened to be free.
2834 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2837 /* If this is an order-0 request then the watermark is fine */
2841 /* For a high-order request, check at least one suitable page is free */
2842 for (o = order; o < MAX_ORDER; o++) {
2843 struct free_area *area = &z->free_area[o];
2852 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2853 if (!list_empty(&area->free_list[mt]))
2858 if ((alloc_flags & ALLOC_CMA) &&
2859 !list_empty(&area->free_list[MIGRATE_CMA])) {
2867 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2868 int classzone_idx, unsigned int alloc_flags)
2870 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2871 zone_page_state(z, NR_FREE_PAGES));
2874 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2875 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2877 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2881 /* If allocation can't use CMA areas don't use free CMA pages */
2882 if (!(alloc_flags & ALLOC_CMA))
2883 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2887 * Fast check for order-0 only. If this fails then the reserves
2888 * need to be calculated. There is a corner case where the check
2889 * passes but only the high-order atomic reserve are free. If
2890 * the caller is !atomic then it'll uselessly search the free
2891 * list. That corner case is then slower but it is harmless.
2893 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2896 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2900 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2901 unsigned long mark, int classzone_idx)
2903 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2905 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2906 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2908 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2913 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2915 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2918 #else /* CONFIG_NUMA */
2919 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2923 #endif /* CONFIG_NUMA */
2926 * get_page_from_freelist goes through the zonelist trying to allocate
2929 static struct page *
2930 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2931 const struct alloc_context *ac)
2933 struct zoneref *z = ac->preferred_zoneref;
2935 struct pglist_data *last_pgdat_dirty_limit = NULL;
2938 * Scan zonelist, looking for a zone with enough free.
2939 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2941 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2946 if (cpusets_enabled() &&
2947 (alloc_flags & ALLOC_CPUSET) &&
2948 !__cpuset_zone_allowed(zone, gfp_mask))
2951 * When allocating a page cache page for writing, we
2952 * want to get it from a node that is within its dirty
2953 * limit, such that no single node holds more than its
2954 * proportional share of globally allowed dirty pages.
2955 * The dirty limits take into account the node's
2956 * lowmem reserves and high watermark so that kswapd
2957 * should be able to balance it without having to
2958 * write pages from its LRU list.
2960 * XXX: For now, allow allocations to potentially
2961 * exceed the per-node dirty limit in the slowpath
2962 * (spread_dirty_pages unset) before going into reclaim,
2963 * which is important when on a NUMA setup the allowed
2964 * nodes are together not big enough to reach the
2965 * global limit. The proper fix for these situations
2966 * will require awareness of nodes in the
2967 * dirty-throttling and the flusher threads.
2969 if (ac->spread_dirty_pages) {
2970 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2973 if (!node_dirty_ok(zone->zone_pgdat)) {
2974 last_pgdat_dirty_limit = zone->zone_pgdat;
2979 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2980 if (!zone_watermark_fast(zone, order, mark,
2981 ac_classzone_idx(ac), alloc_flags)) {
2984 /* Checked here to keep the fast path fast */
2985 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2986 if (alloc_flags & ALLOC_NO_WATERMARKS)
2989 if (node_reclaim_mode == 0 ||
2990 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2993 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2995 case NODE_RECLAIM_NOSCAN:
2998 case NODE_RECLAIM_FULL:
2999 /* scanned but unreclaimable */
3002 /* did we reclaim enough */
3003 if (zone_watermark_ok(zone, order, mark,
3004 ac_classzone_idx(ac), alloc_flags))
3012 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
3013 gfp_mask, alloc_flags, ac->migratetype);
3015 prep_new_page(page, order, gfp_mask, alloc_flags);
3018 * If this is a high-order atomic allocation then check
3019 * if the pageblock should be reserved for the future
3021 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3022 reserve_highatomic_pageblock(page, zone, order);
3032 * Large machines with many possible nodes should not always dump per-node
3033 * meminfo in irq context.
3035 static inline bool should_suppress_show_mem(void)
3040 ret = in_interrupt();
3045 static DEFINE_RATELIMIT_STATE(nopage_rs,
3046 DEFAULT_RATELIMIT_INTERVAL,
3047 DEFAULT_RATELIMIT_BURST);
3049 void warn_alloc(gfp_t gfp_mask, const char *fmt, ...)
3051 unsigned int filter = SHOW_MEM_FILTER_NODES;
3052 struct va_format vaf;
3055 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3056 debug_guardpage_minorder() > 0)
3060 * This documents exceptions given to allocations in certain
3061 * contexts that are allowed to allocate outside current's set
3064 if (!(gfp_mask & __GFP_NOMEMALLOC))
3065 if (test_thread_flag(TIF_MEMDIE) ||
3066 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3067 filter &= ~SHOW_MEM_FILTER_NODES;
3068 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3069 filter &= ~SHOW_MEM_FILTER_NODES;
3071 pr_warn("%s: ", current->comm);
3073 va_start(args, fmt);
3076 pr_cont("%pV", &vaf);
3079 pr_cont(", mode:%#x(%pGg)\n", gfp_mask, &gfp_mask);
3082 if (!should_suppress_show_mem())
3086 static inline struct page *
3087 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3088 const struct alloc_context *ac, unsigned long *did_some_progress)
3090 struct oom_control oc = {
3091 .zonelist = ac->zonelist,
3092 .nodemask = ac->nodemask,
3094 .gfp_mask = gfp_mask,
3099 *did_some_progress = 0;
3102 * Acquire the oom lock. If that fails, somebody else is
3103 * making progress for us.
3105 if (!mutex_trylock(&oom_lock)) {
3106 *did_some_progress = 1;
3107 schedule_timeout_uninterruptible(1);
3112 * Go through the zonelist yet one more time, keep very high watermark
3113 * here, this is only to catch a parallel oom killing, we must fail if
3114 * we're still under heavy pressure.
3116 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3117 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3121 if (!(gfp_mask & __GFP_NOFAIL)) {
3122 /* Coredumps can quickly deplete all memory reserves */
3123 if (current->flags & PF_DUMPCORE)
3125 /* The OOM killer will not help higher order allocs */
3126 if (order > PAGE_ALLOC_COSTLY_ORDER)
3128 /* The OOM killer does not needlessly kill tasks for lowmem */
3129 if (ac->high_zoneidx < ZONE_NORMAL)
3131 if (pm_suspended_storage())
3134 * XXX: GFP_NOFS allocations should rather fail than rely on
3135 * other request to make a forward progress.
3136 * We are in an unfortunate situation where out_of_memory cannot
3137 * do much for this context but let's try it to at least get
3138 * access to memory reserved if the current task is killed (see
3139 * out_of_memory). Once filesystems are ready to handle allocation
3140 * failures more gracefully we should just bail out here.
3143 /* The OOM killer may not free memory on a specific node */
3144 if (gfp_mask & __GFP_THISNODE)
3147 /* Exhausted what can be done so it's blamo time */
3148 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3149 *did_some_progress = 1;
3151 if (gfp_mask & __GFP_NOFAIL) {
3152 page = get_page_from_freelist(gfp_mask, order,
3153 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3155 * fallback to ignore cpuset restriction if our nodes
3159 page = get_page_from_freelist(gfp_mask, order,
3160 ALLOC_NO_WATERMARKS, ac);
3164 mutex_unlock(&oom_lock);
3169 * Maximum number of compaction retries wit a progress before OOM
3170 * killer is consider as the only way to move forward.
3172 #define MAX_COMPACT_RETRIES 16
3174 #ifdef CONFIG_COMPACTION
3175 /* Try memory compaction for high-order allocations before reclaim */
3176 static struct page *
3177 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3178 unsigned int alloc_flags, const struct alloc_context *ac,
3179 enum compact_priority prio, enum compact_result *compact_result)
3186 current->flags |= PF_MEMALLOC;
3187 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3189 current->flags &= ~PF_MEMALLOC;
3191 if (*compact_result <= COMPACT_INACTIVE)
3195 * At least in one zone compaction wasn't deferred or skipped, so let's
3196 * count a compaction stall
3198 count_vm_event(COMPACTSTALL);
3200 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3203 struct zone *zone = page_zone(page);
3205 zone->compact_blockskip_flush = false;
3206 compaction_defer_reset(zone, order, true);
3207 count_vm_event(COMPACTSUCCESS);
3212 * It's bad if compaction run occurs and fails. The most likely reason
3213 * is that pages exist, but not enough to satisfy watermarks.
3215 count_vm_event(COMPACTFAIL);
3223 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3224 enum compact_result compact_result,
3225 enum compact_priority *compact_priority,
3226 int *compaction_retries)
3228 int max_retries = MAX_COMPACT_RETRIES;
3234 if (compaction_made_progress(compact_result))
3235 (*compaction_retries)++;
3238 * compaction considers all the zone as desperately out of memory
3239 * so it doesn't really make much sense to retry except when the
3240 * failure could be caused by insufficient priority
3242 if (compaction_failed(compact_result))
3243 goto check_priority;
3246 * make sure the compaction wasn't deferred or didn't bail out early
3247 * due to locks contention before we declare that we should give up.
3248 * But do not retry if the given zonelist is not suitable for
3251 if (compaction_withdrawn(compact_result))
3252 return compaction_zonelist_suitable(ac, order, alloc_flags);
3255 * !costly requests are much more important than __GFP_REPEAT
3256 * costly ones because they are de facto nofail and invoke OOM
3257 * killer to move on while costly can fail and users are ready
3258 * to cope with that. 1/4 retries is rather arbitrary but we
3259 * would need much more detailed feedback from compaction to
3260 * make a better decision.
3262 if (order > PAGE_ALLOC_COSTLY_ORDER)
3264 if (*compaction_retries <= max_retries)
3268 * Make sure there are attempts at the highest priority if we exhausted
3269 * all retries or failed at the lower priorities.
3272 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3273 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3274 if (*compact_priority > min_priority) {
3275 (*compact_priority)--;
3276 *compaction_retries = 0;
3282 static inline struct page *
3283 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3284 unsigned int alloc_flags, const struct alloc_context *ac,
3285 enum compact_priority prio, enum compact_result *compact_result)
3287 *compact_result = COMPACT_SKIPPED;
3292 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3293 enum compact_result compact_result,
3294 enum compact_priority *compact_priority,
3295 int *compaction_retries)
3300 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3304 * There are setups with compaction disabled which would prefer to loop
3305 * inside the allocator rather than hit the oom killer prematurely.
3306 * Let's give them a good hope and keep retrying while the order-0
3307 * watermarks are OK.
3309 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3311 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3312 ac_classzone_idx(ac), alloc_flags))
3317 #endif /* CONFIG_COMPACTION */
3319 /* Perform direct synchronous page reclaim */
3321 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3322 const struct alloc_context *ac)
3324 struct reclaim_state reclaim_state;
3329 /* We now go into synchronous reclaim */
3330 cpuset_memory_pressure_bump();
3331 current->flags |= PF_MEMALLOC;
3332 lockdep_set_current_reclaim_state(gfp_mask);
3333 reclaim_state.reclaimed_slab = 0;
3334 current->reclaim_state = &reclaim_state;
3336 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3339 current->reclaim_state = NULL;
3340 lockdep_clear_current_reclaim_state();
3341 current->flags &= ~PF_MEMALLOC;
3348 /* The really slow allocator path where we enter direct reclaim */
3349 static inline struct page *
3350 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3351 unsigned int alloc_flags, const struct alloc_context *ac,
3352 unsigned long *did_some_progress)
3354 struct page *page = NULL;
3355 bool drained = false;
3357 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3358 if (unlikely(!(*did_some_progress)))
3362 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3365 * If an allocation failed after direct reclaim, it could be because
3366 * pages are pinned on the per-cpu lists or in high alloc reserves.
3367 * Shrink them them and try again
3369 if (!page && !drained) {
3370 unreserve_highatomic_pageblock(ac);
3371 drain_all_pages(NULL);
3379 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3383 pg_data_t *last_pgdat = NULL;
3385 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3386 ac->high_zoneidx, ac->nodemask) {
3387 if (last_pgdat != zone->zone_pgdat)
3388 wakeup_kswapd(zone, order, ac->high_zoneidx);
3389 last_pgdat = zone->zone_pgdat;
3393 static inline unsigned int
3394 gfp_to_alloc_flags(gfp_t gfp_mask)
3396 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3398 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3399 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3402 * The caller may dip into page reserves a bit more if the caller
3403 * cannot run direct reclaim, or if the caller has realtime scheduling
3404 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3405 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3407 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3409 if (gfp_mask & __GFP_ATOMIC) {
3411 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3412 * if it can't schedule.
3414 if (!(gfp_mask & __GFP_NOMEMALLOC))
3415 alloc_flags |= ALLOC_HARDER;
3417 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3418 * comment for __cpuset_node_allowed().
3420 alloc_flags &= ~ALLOC_CPUSET;
3421 } else if (unlikely(rt_task(current)) && !in_interrupt())
3422 alloc_flags |= ALLOC_HARDER;
3425 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3426 alloc_flags |= ALLOC_CMA;
3431 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3433 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3436 if (gfp_mask & __GFP_MEMALLOC)
3438 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3440 if (!in_interrupt() &&
3441 ((current->flags & PF_MEMALLOC) ||
3442 unlikely(test_thread_flag(TIF_MEMDIE))))
3449 * Maximum number of reclaim retries without any progress before OOM killer
3450 * is consider as the only way to move forward.
3452 #define MAX_RECLAIM_RETRIES 16
3455 * Checks whether it makes sense to retry the reclaim to make a forward progress
3456 * for the given allocation request.
3457 * The reclaim feedback represented by did_some_progress (any progress during
3458 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3459 * any progress in a row) is considered as well as the reclaimable pages on the
3460 * applicable zone list (with a backoff mechanism which is a function of
3461 * no_progress_loops).
3463 * Returns true if a retry is viable or false to enter the oom path.
3466 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3467 struct alloc_context *ac, int alloc_flags,
3468 bool did_some_progress, int *no_progress_loops)
3474 * Costly allocations might have made a progress but this doesn't mean
3475 * their order will become available due to high fragmentation so
3476 * always increment the no progress counter for them
3478 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3479 *no_progress_loops = 0;
3481 (*no_progress_loops)++;
3484 * Make sure we converge to OOM if we cannot make any progress
3485 * several times in the row.
3487 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3491 * Keep reclaiming pages while there is a chance this will lead
3492 * somewhere. If none of the target zones can satisfy our allocation
3493 * request even if all reclaimable pages are considered then we are
3494 * screwed and have to go OOM.
3496 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3498 unsigned long available;
3499 unsigned long reclaimable;
3501 available = reclaimable = zone_reclaimable_pages(zone);
3502 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3503 MAX_RECLAIM_RETRIES);
3504 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3507 * Would the allocation succeed if we reclaimed the whole
3510 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3511 ac_classzone_idx(ac), alloc_flags, available)) {
3513 * If we didn't make any progress and have a lot of
3514 * dirty + writeback pages then we should wait for
3515 * an IO to complete to slow down the reclaim and
3516 * prevent from pre mature OOM
3518 if (!did_some_progress) {
3519 unsigned long write_pending;
3521 write_pending = zone_page_state_snapshot(zone,
3522 NR_ZONE_WRITE_PENDING);
3524 if (2 * write_pending > reclaimable) {
3525 congestion_wait(BLK_RW_ASYNC, HZ/10);
3531 * Memory allocation/reclaim might be called from a WQ
3532 * context and the current implementation of the WQ
3533 * concurrency control doesn't recognize that
3534 * a particular WQ is congested if the worker thread is
3535 * looping without ever sleeping. Therefore we have to
3536 * do a short sleep here rather than calling
3539 if (current->flags & PF_WQ_WORKER)
3540 schedule_timeout_uninterruptible(1);
3551 static inline struct page *
3552 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3553 struct alloc_context *ac)
3555 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3556 struct page *page = NULL;
3557 unsigned int alloc_flags;
3558 unsigned long did_some_progress;
3559 enum compact_priority compact_priority = DEF_COMPACT_PRIORITY;
3560 enum compact_result compact_result;
3561 int compaction_retries = 0;
3562 int no_progress_loops = 0;
3563 unsigned long alloc_start = jiffies;
3564 unsigned int stall_timeout = 10 * HZ;
3567 * In the slowpath, we sanity check order to avoid ever trying to
3568 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3569 * be using allocators in order of preference for an area that is
3572 if (order >= MAX_ORDER) {
3573 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3578 * We also sanity check to catch abuse of atomic reserves being used by
3579 * callers that are not in atomic context.
3581 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3582 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3583 gfp_mask &= ~__GFP_ATOMIC;
3586 * The fast path uses conservative alloc_flags to succeed only until
3587 * kswapd needs to be woken up, and to avoid the cost of setting up
3588 * alloc_flags precisely. So we do that now.
3590 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3592 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3593 wake_all_kswapds(order, ac);
3596 * The adjusted alloc_flags might result in immediate success, so try
3599 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3604 * For costly allocations, try direct compaction first, as it's likely
3605 * that we have enough base pages and don't need to reclaim. Don't try
3606 * that for allocations that are allowed to ignore watermarks, as the
3607 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3609 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3610 !gfp_pfmemalloc_allowed(gfp_mask)) {
3611 page = __alloc_pages_direct_compact(gfp_mask, order,
3613 INIT_COMPACT_PRIORITY,
3619 * Checks for costly allocations with __GFP_NORETRY, which
3620 * includes THP page fault allocations
3622 if (gfp_mask & __GFP_NORETRY) {
3624 * If compaction is deferred for high-order allocations,
3625 * it is because sync compaction recently failed. If
3626 * this is the case and the caller requested a THP
3627 * allocation, we do not want to heavily disrupt the
3628 * system, so we fail the allocation instead of entering
3631 if (compact_result == COMPACT_DEFERRED)
3635 * Looks like reclaim/compaction is worth trying, but
3636 * sync compaction could be very expensive, so keep
3637 * using async compaction.
3639 compact_priority = INIT_COMPACT_PRIORITY;
3644 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3645 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3646 wake_all_kswapds(order, ac);
3648 if (gfp_pfmemalloc_allowed(gfp_mask))
3649 alloc_flags = ALLOC_NO_WATERMARKS;
3652 * Reset the zonelist iterators if memory policies can be ignored.
3653 * These allocations are high priority and system rather than user
3656 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3657 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3658 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3659 ac->high_zoneidx, ac->nodemask);
3662 /* Attempt with potentially adjusted zonelist and alloc_flags */
3663 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3667 /* Caller is not willing to reclaim, we can't balance anything */
3668 if (!can_direct_reclaim) {
3670 * All existing users of the __GFP_NOFAIL are blockable, so warn
3671 * of any new users that actually allow this type of allocation
3674 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3678 /* Avoid recursion of direct reclaim */
3679 if (current->flags & PF_MEMALLOC) {
3681 * __GFP_NOFAIL request from this context is rather bizarre
3682 * because we cannot reclaim anything and only can loop waiting
3683 * for somebody to do a work for us.
3685 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3692 /* Avoid allocations with no watermarks from looping endlessly */
3693 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3697 /* Try direct reclaim and then allocating */
3698 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3699 &did_some_progress);
3703 /* Try direct compaction and then allocating */
3704 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3705 compact_priority, &compact_result);
3709 /* Do not loop if specifically requested */
3710 if (gfp_mask & __GFP_NORETRY)
3714 * Do not retry costly high order allocations unless they are
3717 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3720 /* Make sure we know about allocations which stall for too long */
3721 if (time_after(jiffies, alloc_start + stall_timeout)) {
3722 warn_alloc(gfp_mask,
3723 "page allocation stalls for %ums, order:%u",
3724 jiffies_to_msecs(jiffies-alloc_start), order);
3725 stall_timeout += 10 * HZ;
3728 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3729 did_some_progress > 0, &no_progress_loops))
3733 * It doesn't make any sense to retry for the compaction if the order-0
3734 * reclaim is not able to make any progress because the current
3735 * implementation of the compaction depends on the sufficient amount
3736 * of free memory (see __compaction_suitable)
3738 if (did_some_progress > 0 &&
3739 should_compact_retry(ac, order, alloc_flags,
3740 compact_result, &compact_priority,
3741 &compaction_retries))
3744 /* Reclaim has failed us, start killing things */
3745 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3749 /* Retry as long as the OOM killer is making progress */
3750 if (did_some_progress) {
3751 no_progress_loops = 0;
3756 warn_alloc(gfp_mask,
3757 "page allocation failure: order:%u", order);
3763 * This is the 'heart' of the zoned buddy allocator.
3766 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3767 struct zonelist *zonelist, nodemask_t *nodemask)
3770 unsigned int cpuset_mems_cookie;
3771 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3772 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3773 struct alloc_context ac = {
3774 .high_zoneidx = gfp_zone(gfp_mask),
3775 .zonelist = zonelist,
3776 .nodemask = nodemask,
3777 .migratetype = gfpflags_to_migratetype(gfp_mask),
3780 if (cpusets_enabled()) {
3781 alloc_mask |= __GFP_HARDWALL;
3782 alloc_flags |= ALLOC_CPUSET;
3784 ac.nodemask = &cpuset_current_mems_allowed;
3787 gfp_mask &= gfp_allowed_mask;
3789 lockdep_trace_alloc(gfp_mask);
3791 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3793 if (should_fail_alloc_page(gfp_mask, order))
3797 * Check the zones suitable for the gfp_mask contain at least one
3798 * valid zone. It's possible to have an empty zonelist as a result
3799 * of __GFP_THISNODE and a memoryless node
3801 if (unlikely(!zonelist->_zonerefs->zone))
3804 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3805 alloc_flags |= ALLOC_CMA;
3808 cpuset_mems_cookie = read_mems_allowed_begin();
3810 /* Dirty zone balancing only done in the fast path */
3811 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3814 * The preferred zone is used for statistics but crucially it is
3815 * also used as the starting point for the zonelist iterator. It
3816 * may get reset for allocations that ignore memory policies.
3818 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3819 ac.high_zoneidx, ac.nodemask);
3820 if (!ac.preferred_zoneref) {
3825 /* First allocation attempt */
3826 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3831 * Runtime PM, block IO and its error handling path can deadlock
3832 * because I/O on the device might not complete.
3834 alloc_mask = memalloc_noio_flags(gfp_mask);
3835 ac.spread_dirty_pages = false;
3838 * Restore the original nodemask if it was potentially replaced with
3839 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3841 if (cpusets_enabled())
3842 ac.nodemask = nodemask;
3843 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3847 * When updating a task's mems_allowed, it is possible to race with
3848 * parallel threads in such a way that an allocation can fail while
3849 * the mask is being updated. If a page allocation is about to fail,
3850 * check if the cpuset changed during allocation and if so, retry.
3852 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3853 alloc_mask = gfp_mask;
3858 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3859 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3860 __free_pages(page, order);
3864 if (kmemcheck_enabled && page)
3865 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3867 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3871 EXPORT_SYMBOL(__alloc_pages_nodemask);
3874 * Common helper functions.
3876 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3881 * __get_free_pages() returns a 32-bit address, which cannot represent
3884 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3886 page = alloc_pages(gfp_mask, order);
3889 return (unsigned long) page_address(page);
3891 EXPORT_SYMBOL(__get_free_pages);
3893 unsigned long get_zeroed_page(gfp_t gfp_mask)
3895 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3897 EXPORT_SYMBOL(get_zeroed_page);
3899 void __free_pages(struct page *page, unsigned int order)
3901 if (put_page_testzero(page)) {
3903 free_hot_cold_page(page, false);
3905 __free_pages_ok(page, order);
3909 EXPORT_SYMBOL(__free_pages);
3911 void free_pages(unsigned long addr, unsigned int order)
3914 VM_BUG_ON(!virt_addr_valid((void *)addr));
3915 __free_pages(virt_to_page((void *)addr), order);
3919 EXPORT_SYMBOL(free_pages);
3923 * An arbitrary-length arbitrary-offset area of memory which resides
3924 * within a 0 or higher order page. Multiple fragments within that page
3925 * are individually refcounted, in the page's reference counter.
3927 * The page_frag functions below provide a simple allocation framework for
3928 * page fragments. This is used by the network stack and network device
3929 * drivers to provide a backing region of memory for use as either an
3930 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3932 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3935 struct page *page = NULL;
3936 gfp_t gfp = gfp_mask;
3938 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3939 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3941 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3942 PAGE_FRAG_CACHE_MAX_ORDER);
3943 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3945 if (unlikely(!page))
3946 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3948 nc->va = page ? page_address(page) : NULL;
3953 void *__alloc_page_frag(struct page_frag_cache *nc,
3954 unsigned int fragsz, gfp_t gfp_mask)
3956 unsigned int size = PAGE_SIZE;
3960 if (unlikely(!nc->va)) {
3962 page = __page_frag_refill(nc, gfp_mask);
3966 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3967 /* if size can vary use size else just use PAGE_SIZE */
3970 /* Even if we own the page, we do not use atomic_set().
3971 * This would break get_page_unless_zero() users.
3973 page_ref_add(page, size - 1);
3975 /* reset page count bias and offset to start of new frag */
3976 nc->pfmemalloc = page_is_pfmemalloc(page);
3977 nc->pagecnt_bias = size;
3981 offset = nc->offset - fragsz;
3982 if (unlikely(offset < 0)) {
3983 page = virt_to_page(nc->va);
3985 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3988 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3989 /* if size can vary use size else just use PAGE_SIZE */
3992 /* OK, page count is 0, we can safely set it */
3993 set_page_count(page, size);
3995 /* reset page count bias and offset to start of new frag */
3996 nc->pagecnt_bias = size;
3997 offset = size - fragsz;
4001 nc->offset = offset;
4003 return nc->va + offset;
4005 EXPORT_SYMBOL(__alloc_page_frag);
4008 * Frees a page fragment allocated out of either a compound or order 0 page.
4010 void __free_page_frag(void *addr)
4012 struct page *page = virt_to_head_page(addr);
4014 if (unlikely(put_page_testzero(page)))
4015 __free_pages_ok(page, compound_order(page));
4017 EXPORT_SYMBOL(__free_page_frag);
4019 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4023 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4024 unsigned long used = addr + PAGE_ALIGN(size);
4026 split_page(virt_to_page((void *)addr), order);
4027 while (used < alloc_end) {
4032 return (void *)addr;
4036 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4037 * @size: the number of bytes to allocate
4038 * @gfp_mask: GFP flags for the allocation
4040 * This function is similar to alloc_pages(), except that it allocates the
4041 * minimum number of pages to satisfy the request. alloc_pages() can only
4042 * allocate memory in power-of-two pages.
4044 * This function is also limited by MAX_ORDER.
4046 * Memory allocated by this function must be released by free_pages_exact().
4048 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4050 unsigned int order = get_order(size);
4053 addr = __get_free_pages(gfp_mask, order);
4054 return make_alloc_exact(addr, order, size);
4056 EXPORT_SYMBOL(alloc_pages_exact);
4059 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4061 * @nid: the preferred node ID where memory should be allocated
4062 * @size: the number of bytes to allocate
4063 * @gfp_mask: GFP flags for the allocation
4065 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4068 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4070 unsigned int order = get_order(size);
4071 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4074 return make_alloc_exact((unsigned long)page_address(p), order, size);
4078 * free_pages_exact - release memory allocated via alloc_pages_exact()
4079 * @virt: the value returned by alloc_pages_exact.
4080 * @size: size of allocation, same value as passed to alloc_pages_exact().
4082 * Release the memory allocated by a previous call to alloc_pages_exact.
4084 void free_pages_exact(void *virt, size_t size)
4086 unsigned long addr = (unsigned long)virt;
4087 unsigned long end = addr + PAGE_ALIGN(size);
4089 while (addr < end) {
4094 EXPORT_SYMBOL(free_pages_exact);
4097 * nr_free_zone_pages - count number of pages beyond high watermark
4098 * @offset: The zone index of the highest zone
4100 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4101 * high watermark within all zones at or below a given zone index. For each
4102 * zone, the number of pages is calculated as:
4103 * managed_pages - high_pages
4105 static unsigned long nr_free_zone_pages(int offset)
4110 /* Just pick one node, since fallback list is circular */
4111 unsigned long sum = 0;
4113 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4115 for_each_zone_zonelist(zone, z, zonelist, offset) {
4116 unsigned long size = zone->managed_pages;
4117 unsigned long high = high_wmark_pages(zone);
4126 * nr_free_buffer_pages - count number of pages beyond high watermark
4128 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4129 * watermark within ZONE_DMA and ZONE_NORMAL.
4131 unsigned long nr_free_buffer_pages(void)
4133 return nr_free_zone_pages(gfp_zone(GFP_USER));
4135 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4138 * nr_free_pagecache_pages - count number of pages beyond high watermark
4140 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4141 * high watermark within all zones.
4143 unsigned long nr_free_pagecache_pages(void)
4145 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4148 static inline void show_node(struct zone *zone)
4150 if (IS_ENABLED(CONFIG_NUMA))
4151 printk("Node %d ", zone_to_nid(zone));
4154 long si_mem_available(void)
4157 unsigned long pagecache;
4158 unsigned long wmark_low = 0;
4159 unsigned long pages[NR_LRU_LISTS];
4163 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4164 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4167 wmark_low += zone->watermark[WMARK_LOW];
4170 * Estimate the amount of memory available for userspace allocations,
4171 * without causing swapping.
4173 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4176 * Not all the page cache can be freed, otherwise the system will
4177 * start swapping. Assume at least half of the page cache, or the
4178 * low watermark worth of cache, needs to stay.
4180 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4181 pagecache -= min(pagecache / 2, wmark_low);
4182 available += pagecache;
4185 * Part of the reclaimable slab consists of items that are in use,
4186 * and cannot be freed. Cap this estimate at the low watermark.
4188 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4189 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4195 EXPORT_SYMBOL_GPL(si_mem_available);
4197 void si_meminfo(struct sysinfo *val)
4199 val->totalram = totalram_pages;
4200 val->sharedram = global_node_page_state(NR_SHMEM);
4201 val->freeram = global_page_state(NR_FREE_PAGES);
4202 val->bufferram = nr_blockdev_pages();
4203 val->totalhigh = totalhigh_pages;
4204 val->freehigh = nr_free_highpages();
4205 val->mem_unit = PAGE_SIZE;
4208 EXPORT_SYMBOL(si_meminfo);
4211 void si_meminfo_node(struct sysinfo *val, int nid)
4213 int zone_type; /* needs to be signed */
4214 unsigned long managed_pages = 0;
4215 unsigned long managed_highpages = 0;
4216 unsigned long free_highpages = 0;
4217 pg_data_t *pgdat = NODE_DATA(nid);
4219 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4220 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4221 val->totalram = managed_pages;
4222 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4223 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4224 #ifdef CONFIG_HIGHMEM
4225 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4226 struct zone *zone = &pgdat->node_zones[zone_type];
4228 if (is_highmem(zone)) {
4229 managed_highpages += zone->managed_pages;
4230 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4233 val->totalhigh = managed_highpages;
4234 val->freehigh = free_highpages;
4236 val->totalhigh = managed_highpages;
4237 val->freehigh = free_highpages;
4239 val->mem_unit = PAGE_SIZE;
4244 * Determine whether the node should be displayed or not, depending on whether
4245 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4247 bool skip_free_areas_node(unsigned int flags, int nid)
4250 unsigned int cpuset_mems_cookie;
4252 if (!(flags & SHOW_MEM_FILTER_NODES))
4256 cpuset_mems_cookie = read_mems_allowed_begin();
4257 ret = !node_isset(nid, cpuset_current_mems_allowed);
4258 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4263 #define K(x) ((x) << (PAGE_SHIFT-10))
4265 static void show_migration_types(unsigned char type)
4267 static const char types[MIGRATE_TYPES] = {
4268 [MIGRATE_UNMOVABLE] = 'U',
4269 [MIGRATE_MOVABLE] = 'M',
4270 [MIGRATE_RECLAIMABLE] = 'E',
4271 [MIGRATE_HIGHATOMIC] = 'H',
4273 [MIGRATE_CMA] = 'C',
4275 #ifdef CONFIG_MEMORY_ISOLATION
4276 [MIGRATE_ISOLATE] = 'I',
4279 char tmp[MIGRATE_TYPES + 1];
4283 for (i = 0; i < MIGRATE_TYPES; i++) {
4284 if (type & (1 << i))
4289 printk(KERN_CONT "(%s) ", tmp);
4293 * Show free area list (used inside shift_scroll-lock stuff)
4294 * We also calculate the percentage fragmentation. We do this by counting the
4295 * memory on each free list with the exception of the first item on the list.
4298 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4301 void show_free_areas(unsigned int filter)
4303 unsigned long free_pcp = 0;
4308 for_each_populated_zone(zone) {
4309 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4312 for_each_online_cpu(cpu)
4313 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4316 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4317 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4318 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4319 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4320 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4321 " free:%lu free_pcp:%lu free_cma:%lu\n",
4322 global_node_page_state(NR_ACTIVE_ANON),
4323 global_node_page_state(NR_INACTIVE_ANON),
4324 global_node_page_state(NR_ISOLATED_ANON),
4325 global_node_page_state(NR_ACTIVE_FILE),
4326 global_node_page_state(NR_INACTIVE_FILE),
4327 global_node_page_state(NR_ISOLATED_FILE),
4328 global_node_page_state(NR_UNEVICTABLE),
4329 global_node_page_state(NR_FILE_DIRTY),
4330 global_node_page_state(NR_WRITEBACK),
4331 global_node_page_state(NR_UNSTABLE_NFS),
4332 global_page_state(NR_SLAB_RECLAIMABLE),
4333 global_page_state(NR_SLAB_UNRECLAIMABLE),
4334 global_node_page_state(NR_FILE_MAPPED),
4335 global_node_page_state(NR_SHMEM),
4336 global_page_state(NR_PAGETABLE),
4337 global_page_state(NR_BOUNCE),
4338 global_page_state(NR_FREE_PAGES),
4340 global_page_state(NR_FREE_CMA_PAGES));
4342 for_each_online_pgdat(pgdat) {
4344 " active_anon:%lukB"
4345 " inactive_anon:%lukB"
4346 " active_file:%lukB"
4347 " inactive_file:%lukB"
4348 " unevictable:%lukB"
4349 " isolated(anon):%lukB"
4350 " isolated(file):%lukB"
4355 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4357 " shmem_pmdmapped: %lukB"
4360 " writeback_tmp:%lukB"
4362 " pages_scanned:%lu"
4363 " all_unreclaimable? %s"
4366 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4367 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4368 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4369 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4370 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4371 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4372 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4373 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4374 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4375 K(node_page_state(pgdat, NR_WRITEBACK)),
4376 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4377 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4378 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4380 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4382 K(node_page_state(pgdat, NR_SHMEM)),
4383 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4384 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4385 node_page_state(pgdat, NR_PAGES_SCANNED),
4386 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4389 for_each_populated_zone(zone) {
4392 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4396 for_each_online_cpu(cpu)
4397 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4406 " active_anon:%lukB"
4407 " inactive_anon:%lukB"
4408 " active_file:%lukB"
4409 " inactive_file:%lukB"
4410 " unevictable:%lukB"
4411 " writepending:%lukB"
4415 " slab_reclaimable:%lukB"
4416 " slab_unreclaimable:%lukB"
4417 " kernel_stack:%lukB"
4425 K(zone_page_state(zone, NR_FREE_PAGES)),
4426 K(min_wmark_pages(zone)),
4427 K(low_wmark_pages(zone)),
4428 K(high_wmark_pages(zone)),
4429 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4430 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4431 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4432 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4433 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4434 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4435 K(zone->present_pages),
4436 K(zone->managed_pages),
4437 K(zone_page_state(zone, NR_MLOCK)),
4438 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4439 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4440 zone_page_state(zone, NR_KERNEL_STACK_KB),
4441 K(zone_page_state(zone, NR_PAGETABLE)),
4442 K(zone_page_state(zone, NR_BOUNCE)),
4444 K(this_cpu_read(zone->pageset->pcp.count)),
4445 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4446 printk("lowmem_reserve[]:");
4447 for (i = 0; i < MAX_NR_ZONES; i++)
4448 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4449 printk(KERN_CONT "\n");
4452 for_each_populated_zone(zone) {
4454 unsigned long nr[MAX_ORDER], flags, total = 0;
4455 unsigned char types[MAX_ORDER];
4457 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4460 printk(KERN_CONT "%s: ", zone->name);
4462 spin_lock_irqsave(&zone->lock, flags);
4463 for (order = 0; order < MAX_ORDER; order++) {
4464 struct free_area *area = &zone->free_area[order];
4467 nr[order] = area->nr_free;
4468 total += nr[order] << order;
4471 for (type = 0; type < MIGRATE_TYPES; type++) {
4472 if (!list_empty(&area->free_list[type]))
4473 types[order] |= 1 << type;
4476 spin_unlock_irqrestore(&zone->lock, flags);
4477 for (order = 0; order < MAX_ORDER; order++) {
4478 printk(KERN_CONT "%lu*%lukB ",
4479 nr[order], K(1UL) << order);
4481 show_migration_types(types[order]);
4483 printk(KERN_CONT "= %lukB\n", K(total));
4486 hugetlb_show_meminfo();
4488 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4490 show_swap_cache_info();
4493 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4495 zoneref->zone = zone;
4496 zoneref->zone_idx = zone_idx(zone);
4500 * Builds allocation fallback zone lists.
4502 * Add all populated zones of a node to the zonelist.
4504 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4508 enum zone_type zone_type = MAX_NR_ZONES;
4512 zone = pgdat->node_zones + zone_type;
4513 if (managed_zone(zone)) {
4514 zoneref_set_zone(zone,
4515 &zonelist->_zonerefs[nr_zones++]);
4516 check_highest_zone(zone_type);
4518 } while (zone_type);
4526 * 0 = automatic detection of better ordering.
4527 * 1 = order by ([node] distance, -zonetype)
4528 * 2 = order by (-zonetype, [node] distance)
4530 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4531 * the same zonelist. So only NUMA can configure this param.
4533 #define ZONELIST_ORDER_DEFAULT 0
4534 #define ZONELIST_ORDER_NODE 1
4535 #define ZONELIST_ORDER_ZONE 2
4537 /* zonelist order in the kernel.
4538 * set_zonelist_order() will set this to NODE or ZONE.
4540 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4541 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4545 /* The value user specified ....changed by config */
4546 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4547 /* string for sysctl */
4548 #define NUMA_ZONELIST_ORDER_LEN 16
4549 char numa_zonelist_order[16] = "default";
4552 * interface for configure zonelist ordering.
4553 * command line option "numa_zonelist_order"
4554 * = "[dD]efault - default, automatic configuration.
4555 * = "[nN]ode - order by node locality, then by zone within node
4556 * = "[zZ]one - order by zone, then by locality within zone
4559 static int __parse_numa_zonelist_order(char *s)
4561 if (*s == 'd' || *s == 'D') {
4562 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4563 } else if (*s == 'n' || *s == 'N') {
4564 user_zonelist_order = ZONELIST_ORDER_NODE;
4565 } else if (*s == 'z' || *s == 'Z') {
4566 user_zonelist_order = ZONELIST_ORDER_ZONE;
4568 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4574 static __init int setup_numa_zonelist_order(char *s)
4581 ret = __parse_numa_zonelist_order(s);
4583 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4587 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4590 * sysctl handler for numa_zonelist_order
4592 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4593 void __user *buffer, size_t *length,
4596 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4598 static DEFINE_MUTEX(zl_order_mutex);
4600 mutex_lock(&zl_order_mutex);
4602 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4606 strcpy(saved_string, (char *)table->data);
4608 ret = proc_dostring(table, write, buffer, length, ppos);
4612 int oldval = user_zonelist_order;
4614 ret = __parse_numa_zonelist_order((char *)table->data);
4617 * bogus value. restore saved string
4619 strncpy((char *)table->data, saved_string,
4620 NUMA_ZONELIST_ORDER_LEN);
4621 user_zonelist_order = oldval;
4622 } else if (oldval != user_zonelist_order) {
4623 mutex_lock(&zonelists_mutex);
4624 build_all_zonelists(NULL, NULL);
4625 mutex_unlock(&zonelists_mutex);
4629 mutex_unlock(&zl_order_mutex);
4634 #define MAX_NODE_LOAD (nr_online_nodes)
4635 static int node_load[MAX_NUMNODES];
4638 * find_next_best_node - find the next node that should appear in a given node's fallback list
4639 * @node: node whose fallback list we're appending
4640 * @used_node_mask: nodemask_t of already used nodes
4642 * We use a number of factors to determine which is the next node that should
4643 * appear on a given node's fallback list. The node should not have appeared
4644 * already in @node's fallback list, and it should be the next closest node
4645 * according to the distance array (which contains arbitrary distance values
4646 * from each node to each node in the system), and should also prefer nodes
4647 * with no CPUs, since presumably they'll have very little allocation pressure
4648 * on them otherwise.
4649 * It returns -1 if no node is found.
4651 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4654 int min_val = INT_MAX;
4655 int best_node = NUMA_NO_NODE;
4656 const struct cpumask *tmp = cpumask_of_node(0);
4658 /* Use the local node if we haven't already */
4659 if (!node_isset(node, *used_node_mask)) {
4660 node_set(node, *used_node_mask);
4664 for_each_node_state(n, N_MEMORY) {
4666 /* Don't want a node to appear more than once */
4667 if (node_isset(n, *used_node_mask))
4670 /* Use the distance array to find the distance */
4671 val = node_distance(node, n);
4673 /* Penalize nodes under us ("prefer the next node") */
4676 /* Give preference to headless and unused nodes */
4677 tmp = cpumask_of_node(n);
4678 if (!cpumask_empty(tmp))
4679 val += PENALTY_FOR_NODE_WITH_CPUS;
4681 /* Slight preference for less loaded node */
4682 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4683 val += node_load[n];
4685 if (val < min_val) {
4692 node_set(best_node, *used_node_mask);
4699 * Build zonelists ordered by node and zones within node.
4700 * This results in maximum locality--normal zone overflows into local
4701 * DMA zone, if any--but risks exhausting DMA zone.
4703 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4706 struct zonelist *zonelist;
4708 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4709 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4711 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4712 zonelist->_zonerefs[j].zone = NULL;
4713 zonelist->_zonerefs[j].zone_idx = 0;
4717 * Build gfp_thisnode zonelists
4719 static void build_thisnode_zonelists(pg_data_t *pgdat)
4722 struct zonelist *zonelist;
4724 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4725 j = build_zonelists_node(pgdat, zonelist, 0);
4726 zonelist->_zonerefs[j].zone = NULL;
4727 zonelist->_zonerefs[j].zone_idx = 0;
4731 * Build zonelists ordered by zone and nodes within zones.
4732 * This results in conserving DMA zone[s] until all Normal memory is
4733 * exhausted, but results in overflowing to remote node while memory
4734 * may still exist in local DMA zone.
4736 static int node_order[MAX_NUMNODES];
4738 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4741 int zone_type; /* needs to be signed */
4743 struct zonelist *zonelist;
4745 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4747 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4748 for (j = 0; j < nr_nodes; j++) {
4749 node = node_order[j];
4750 z = &NODE_DATA(node)->node_zones[zone_type];
4751 if (managed_zone(z)) {
4753 &zonelist->_zonerefs[pos++]);
4754 check_highest_zone(zone_type);
4758 zonelist->_zonerefs[pos].zone = NULL;
4759 zonelist->_zonerefs[pos].zone_idx = 0;
4762 #if defined(CONFIG_64BIT)
4764 * Devices that require DMA32/DMA are relatively rare and do not justify a
4765 * penalty to every machine in case the specialised case applies. Default
4766 * to Node-ordering on 64-bit NUMA machines
4768 static int default_zonelist_order(void)
4770 return ZONELIST_ORDER_NODE;
4774 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4775 * by the kernel. If processes running on node 0 deplete the low memory zone
4776 * then reclaim will occur more frequency increasing stalls and potentially
4777 * be easier to OOM if a large percentage of the zone is under writeback or
4778 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4779 * Hence, default to zone ordering on 32-bit.
4781 static int default_zonelist_order(void)
4783 return ZONELIST_ORDER_ZONE;
4785 #endif /* CONFIG_64BIT */
4787 static void set_zonelist_order(void)
4789 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4790 current_zonelist_order = default_zonelist_order();
4792 current_zonelist_order = user_zonelist_order;
4795 static void build_zonelists(pg_data_t *pgdat)
4798 nodemask_t used_mask;
4799 int local_node, prev_node;
4800 struct zonelist *zonelist;
4801 unsigned int order = current_zonelist_order;
4803 /* initialize zonelists */
4804 for (i = 0; i < MAX_ZONELISTS; i++) {
4805 zonelist = pgdat->node_zonelists + i;
4806 zonelist->_zonerefs[0].zone = NULL;
4807 zonelist->_zonerefs[0].zone_idx = 0;
4810 /* NUMA-aware ordering of nodes */
4811 local_node = pgdat->node_id;
4812 load = nr_online_nodes;
4813 prev_node = local_node;
4814 nodes_clear(used_mask);
4816 memset(node_order, 0, sizeof(node_order));
4819 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4821 * We don't want to pressure a particular node.
4822 * So adding penalty to the first node in same
4823 * distance group to make it round-robin.
4825 if (node_distance(local_node, node) !=
4826 node_distance(local_node, prev_node))
4827 node_load[node] = load;
4831 if (order == ZONELIST_ORDER_NODE)
4832 build_zonelists_in_node_order(pgdat, node);
4834 node_order[i++] = node; /* remember order */
4837 if (order == ZONELIST_ORDER_ZONE) {
4838 /* calculate node order -- i.e., DMA last! */
4839 build_zonelists_in_zone_order(pgdat, i);
4842 build_thisnode_zonelists(pgdat);
4845 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4847 * Return node id of node used for "local" allocations.
4848 * I.e., first node id of first zone in arg node's generic zonelist.
4849 * Used for initializing percpu 'numa_mem', which is used primarily
4850 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4852 int local_memory_node(int node)
4856 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4857 gfp_zone(GFP_KERNEL),
4859 return z->zone->node;
4863 static void setup_min_unmapped_ratio(void);
4864 static void setup_min_slab_ratio(void);
4865 #else /* CONFIG_NUMA */
4867 static void set_zonelist_order(void)
4869 current_zonelist_order = ZONELIST_ORDER_ZONE;
4872 static void build_zonelists(pg_data_t *pgdat)
4874 int node, local_node;
4876 struct zonelist *zonelist;
4878 local_node = pgdat->node_id;
4880 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4881 j = build_zonelists_node(pgdat, zonelist, 0);
4884 * Now we build the zonelist so that it contains the zones
4885 * of all the other nodes.
4886 * We don't want to pressure a particular node, so when
4887 * building the zones for node N, we make sure that the
4888 * zones coming right after the local ones are those from
4889 * node N+1 (modulo N)
4891 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4892 if (!node_online(node))
4894 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4896 for (node = 0; node < local_node; node++) {
4897 if (!node_online(node))
4899 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4902 zonelist->_zonerefs[j].zone = NULL;
4903 zonelist->_zonerefs[j].zone_idx = 0;
4906 #endif /* CONFIG_NUMA */
4909 * Boot pageset table. One per cpu which is going to be used for all
4910 * zones and all nodes. The parameters will be set in such a way
4911 * that an item put on a list will immediately be handed over to
4912 * the buddy list. This is safe since pageset manipulation is done
4913 * with interrupts disabled.
4915 * The boot_pagesets must be kept even after bootup is complete for
4916 * unused processors and/or zones. They do play a role for bootstrapping
4917 * hotplugged processors.
4919 * zoneinfo_show() and maybe other functions do
4920 * not check if the processor is online before following the pageset pointer.
4921 * Other parts of the kernel may not check if the zone is available.
4923 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4924 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4925 static void setup_zone_pageset(struct zone *zone);
4928 * Global mutex to protect against size modification of zonelists
4929 * as well as to serialize pageset setup for the new populated zone.
4931 DEFINE_MUTEX(zonelists_mutex);
4933 /* return values int ....just for stop_machine() */
4934 static int __build_all_zonelists(void *data)
4938 pg_data_t *self = data;
4941 memset(node_load, 0, sizeof(node_load));
4944 if (self && !node_online(self->node_id)) {
4945 build_zonelists(self);
4948 for_each_online_node(nid) {
4949 pg_data_t *pgdat = NODE_DATA(nid);
4951 build_zonelists(pgdat);
4955 * Initialize the boot_pagesets that are going to be used
4956 * for bootstrapping processors. The real pagesets for
4957 * each zone will be allocated later when the per cpu
4958 * allocator is available.
4960 * boot_pagesets are used also for bootstrapping offline
4961 * cpus if the system is already booted because the pagesets
4962 * are needed to initialize allocators on a specific cpu too.
4963 * F.e. the percpu allocator needs the page allocator which
4964 * needs the percpu allocator in order to allocate its pagesets
4965 * (a chicken-egg dilemma).
4967 for_each_possible_cpu(cpu) {
4968 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4970 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4972 * We now know the "local memory node" for each node--
4973 * i.e., the node of the first zone in the generic zonelist.
4974 * Set up numa_mem percpu variable for on-line cpus. During
4975 * boot, only the boot cpu should be on-line; we'll init the
4976 * secondary cpus' numa_mem as they come on-line. During
4977 * node/memory hotplug, we'll fixup all on-line cpus.
4979 if (cpu_online(cpu))
4980 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4987 static noinline void __init
4988 build_all_zonelists_init(void)
4990 __build_all_zonelists(NULL);
4991 mminit_verify_zonelist();
4992 cpuset_init_current_mems_allowed();
4996 * Called with zonelists_mutex held always
4997 * unless system_state == SYSTEM_BOOTING.
4999 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5000 * [we're only called with non-NULL zone through __meminit paths] and
5001 * (2) call of __init annotated helper build_all_zonelists_init
5002 * [protected by SYSTEM_BOOTING].
5004 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5006 set_zonelist_order();
5008 if (system_state == SYSTEM_BOOTING) {
5009 build_all_zonelists_init();
5011 #ifdef CONFIG_MEMORY_HOTPLUG
5013 setup_zone_pageset(zone);
5015 /* we have to stop all cpus to guarantee there is no user
5017 stop_machine(__build_all_zonelists, pgdat, NULL);
5018 /* cpuset refresh routine should be here */
5020 vm_total_pages = nr_free_pagecache_pages();
5022 * Disable grouping by mobility if the number of pages in the
5023 * system is too low to allow the mechanism to work. It would be
5024 * more accurate, but expensive to check per-zone. This check is
5025 * made on memory-hotadd so a system can start with mobility
5026 * disabled and enable it later
5028 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5029 page_group_by_mobility_disabled = 1;
5031 page_group_by_mobility_disabled = 0;
5033 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5035 zonelist_order_name[current_zonelist_order],
5036 page_group_by_mobility_disabled ? "off" : "on",
5039 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5044 * Initially all pages are reserved - free ones are freed
5045 * up by free_all_bootmem() once the early boot process is
5046 * done. Non-atomic initialization, single-pass.
5048 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5049 unsigned long start_pfn, enum memmap_context context)
5051 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5052 unsigned long end_pfn = start_pfn + size;
5053 pg_data_t *pgdat = NODE_DATA(nid);
5055 unsigned long nr_initialised = 0;
5056 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5057 struct memblock_region *r = NULL, *tmp;
5060 if (highest_memmap_pfn < end_pfn - 1)
5061 highest_memmap_pfn = end_pfn - 1;
5064 * Honor reservation requested by the driver for this ZONE_DEVICE
5067 if (altmap && start_pfn == altmap->base_pfn)
5068 start_pfn += altmap->reserve;
5070 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5072 * There can be holes in boot-time mem_map[]s handed to this
5073 * function. They do not exist on hotplugged memory.
5075 if (context != MEMMAP_EARLY)
5078 if (!early_pfn_valid(pfn))
5080 if (!early_pfn_in_nid(pfn, nid))
5082 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5085 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5087 * Check given memblock attribute by firmware which can affect
5088 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5089 * mirrored, it's an overlapped memmap init. skip it.
5091 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5092 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5093 for_each_memblock(memory, tmp)
5094 if (pfn < memblock_region_memory_end_pfn(tmp))
5098 if (pfn >= memblock_region_memory_base_pfn(r) &&
5099 memblock_is_mirror(r)) {
5100 /* already initialized as NORMAL */
5101 pfn = memblock_region_memory_end_pfn(r);
5109 * Mark the block movable so that blocks are reserved for
5110 * movable at startup. This will force kernel allocations
5111 * to reserve their blocks rather than leaking throughout
5112 * the address space during boot when many long-lived
5113 * kernel allocations are made.
5115 * bitmap is created for zone's valid pfn range. but memmap
5116 * can be created for invalid pages (for alignment)
5117 * check here not to call set_pageblock_migratetype() against
5120 if (!(pfn & (pageblock_nr_pages - 1))) {
5121 struct page *page = pfn_to_page(pfn);
5123 __init_single_page(page, pfn, zone, nid);
5124 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5126 __init_single_pfn(pfn, zone, nid);
5131 static void __meminit zone_init_free_lists(struct zone *zone)
5133 unsigned int order, t;
5134 for_each_migratetype_order(order, t) {
5135 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5136 zone->free_area[order].nr_free = 0;
5140 #ifndef __HAVE_ARCH_MEMMAP_INIT
5141 #define memmap_init(size, nid, zone, start_pfn) \
5142 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5145 static int zone_batchsize(struct zone *zone)
5151 * The per-cpu-pages pools are set to around 1000th of the
5152 * size of the zone. But no more than 1/2 of a meg.
5154 * OK, so we don't know how big the cache is. So guess.
5156 batch = zone->managed_pages / 1024;
5157 if (batch * PAGE_SIZE > 512 * 1024)
5158 batch = (512 * 1024) / PAGE_SIZE;
5159 batch /= 4; /* We effectively *= 4 below */
5164 * Clamp the batch to a 2^n - 1 value. Having a power
5165 * of 2 value was found to be more likely to have
5166 * suboptimal cache aliasing properties in some cases.
5168 * For example if 2 tasks are alternately allocating
5169 * batches of pages, one task can end up with a lot
5170 * of pages of one half of the possible page colors
5171 * and the other with pages of the other colors.
5173 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5178 /* The deferral and batching of frees should be suppressed under NOMMU
5181 * The problem is that NOMMU needs to be able to allocate large chunks
5182 * of contiguous memory as there's no hardware page translation to
5183 * assemble apparent contiguous memory from discontiguous pages.
5185 * Queueing large contiguous runs of pages for batching, however,
5186 * causes the pages to actually be freed in smaller chunks. As there
5187 * can be a significant delay between the individual batches being
5188 * recycled, this leads to the once large chunks of space being
5189 * fragmented and becoming unavailable for high-order allocations.
5196 * pcp->high and pcp->batch values are related and dependent on one another:
5197 * ->batch must never be higher then ->high.
5198 * The following function updates them in a safe manner without read side
5201 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5202 * those fields changing asynchronously (acording the the above rule).
5204 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5205 * outside of boot time (or some other assurance that no concurrent updaters
5208 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5209 unsigned long batch)
5211 /* start with a fail safe value for batch */
5215 /* Update high, then batch, in order */
5222 /* a companion to pageset_set_high() */
5223 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5225 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5228 static void pageset_init(struct per_cpu_pageset *p)
5230 struct per_cpu_pages *pcp;
5233 memset(p, 0, sizeof(*p));
5237 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5238 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5241 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5244 pageset_set_batch(p, batch);
5248 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5249 * to the value high for the pageset p.
5251 static void pageset_set_high(struct per_cpu_pageset *p,
5254 unsigned long batch = max(1UL, high / 4);
5255 if ((high / 4) > (PAGE_SHIFT * 8))
5256 batch = PAGE_SHIFT * 8;
5258 pageset_update(&p->pcp, high, batch);
5261 static void pageset_set_high_and_batch(struct zone *zone,
5262 struct per_cpu_pageset *pcp)
5264 if (percpu_pagelist_fraction)
5265 pageset_set_high(pcp,
5266 (zone->managed_pages /
5267 percpu_pagelist_fraction));
5269 pageset_set_batch(pcp, zone_batchsize(zone));
5272 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5274 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5277 pageset_set_high_and_batch(zone, pcp);
5280 static void __meminit setup_zone_pageset(struct zone *zone)
5283 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5284 for_each_possible_cpu(cpu)
5285 zone_pageset_init(zone, cpu);
5289 * Allocate per cpu pagesets and initialize them.
5290 * Before this call only boot pagesets were available.
5292 void __init setup_per_cpu_pageset(void)
5294 struct pglist_data *pgdat;
5297 for_each_populated_zone(zone)
5298 setup_zone_pageset(zone);
5300 for_each_online_pgdat(pgdat)
5301 pgdat->per_cpu_nodestats =
5302 alloc_percpu(struct per_cpu_nodestat);
5305 static __meminit void zone_pcp_init(struct zone *zone)
5308 * per cpu subsystem is not up at this point. The following code
5309 * relies on the ability of the linker to provide the
5310 * offset of a (static) per cpu variable into the per cpu area.
5312 zone->pageset = &boot_pageset;
5314 if (populated_zone(zone))
5315 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5316 zone->name, zone->present_pages,
5317 zone_batchsize(zone));
5320 int __meminit init_currently_empty_zone(struct zone *zone,
5321 unsigned long zone_start_pfn,
5324 struct pglist_data *pgdat = zone->zone_pgdat;
5326 pgdat->nr_zones = zone_idx(zone) + 1;
5328 zone->zone_start_pfn = zone_start_pfn;
5330 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5331 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5333 (unsigned long)zone_idx(zone),
5334 zone_start_pfn, (zone_start_pfn + size));
5336 zone_init_free_lists(zone);
5337 zone->initialized = 1;
5342 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5343 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5346 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5348 int __meminit __early_pfn_to_nid(unsigned long pfn,
5349 struct mminit_pfnnid_cache *state)
5351 unsigned long start_pfn, end_pfn;
5354 if (state->last_start <= pfn && pfn < state->last_end)
5355 return state->last_nid;
5357 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5359 state->last_start = start_pfn;
5360 state->last_end = end_pfn;
5361 state->last_nid = nid;
5366 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5369 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5370 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5371 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5373 * If an architecture guarantees that all ranges registered contain no holes
5374 * and may be freed, this this function may be used instead of calling
5375 * memblock_free_early_nid() manually.
5377 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5379 unsigned long start_pfn, end_pfn;
5382 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5383 start_pfn = min(start_pfn, max_low_pfn);
5384 end_pfn = min(end_pfn, max_low_pfn);
5386 if (start_pfn < end_pfn)
5387 memblock_free_early_nid(PFN_PHYS(start_pfn),
5388 (end_pfn - start_pfn) << PAGE_SHIFT,
5394 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5395 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5397 * If an architecture guarantees that all ranges registered contain no holes and may
5398 * be freed, this function may be used instead of calling memory_present() manually.
5400 void __init sparse_memory_present_with_active_regions(int nid)
5402 unsigned long start_pfn, end_pfn;
5405 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5406 memory_present(this_nid, start_pfn, end_pfn);
5410 * get_pfn_range_for_nid - Return the start and end page frames for a node
5411 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5412 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5413 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5415 * It returns the start and end page frame of a node based on information
5416 * provided by memblock_set_node(). If called for a node
5417 * with no available memory, a warning is printed and the start and end
5420 void __meminit get_pfn_range_for_nid(unsigned int nid,
5421 unsigned long *start_pfn, unsigned long *end_pfn)
5423 unsigned long this_start_pfn, this_end_pfn;
5429 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5430 *start_pfn = min(*start_pfn, this_start_pfn);
5431 *end_pfn = max(*end_pfn, this_end_pfn);
5434 if (*start_pfn == -1UL)
5439 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5440 * assumption is made that zones within a node are ordered in monotonic
5441 * increasing memory addresses so that the "highest" populated zone is used
5443 static void __init find_usable_zone_for_movable(void)
5446 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5447 if (zone_index == ZONE_MOVABLE)
5450 if (arch_zone_highest_possible_pfn[zone_index] >
5451 arch_zone_lowest_possible_pfn[zone_index])
5455 VM_BUG_ON(zone_index == -1);
5456 movable_zone = zone_index;
5460 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5461 * because it is sized independent of architecture. Unlike the other zones,
5462 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5463 * in each node depending on the size of each node and how evenly kernelcore
5464 * is distributed. This helper function adjusts the zone ranges
5465 * provided by the architecture for a given node by using the end of the
5466 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5467 * zones within a node are in order of monotonic increases memory addresses
5469 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5470 unsigned long zone_type,
5471 unsigned long node_start_pfn,
5472 unsigned long node_end_pfn,
5473 unsigned long *zone_start_pfn,
5474 unsigned long *zone_end_pfn)
5476 /* Only adjust if ZONE_MOVABLE is on this node */
5477 if (zone_movable_pfn[nid]) {
5478 /* Size ZONE_MOVABLE */
5479 if (zone_type == ZONE_MOVABLE) {
5480 *zone_start_pfn = zone_movable_pfn[nid];
5481 *zone_end_pfn = min(node_end_pfn,
5482 arch_zone_highest_possible_pfn[movable_zone]);
5484 /* Adjust for ZONE_MOVABLE starting within this range */
5485 } else if (!mirrored_kernelcore &&
5486 *zone_start_pfn < zone_movable_pfn[nid] &&
5487 *zone_end_pfn > zone_movable_pfn[nid]) {
5488 *zone_end_pfn = zone_movable_pfn[nid];
5490 /* Check if this whole range is within ZONE_MOVABLE */
5491 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5492 *zone_start_pfn = *zone_end_pfn;
5497 * Return the number of pages a zone spans in a node, including holes
5498 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5500 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5501 unsigned long zone_type,
5502 unsigned long node_start_pfn,
5503 unsigned long node_end_pfn,
5504 unsigned long *zone_start_pfn,
5505 unsigned long *zone_end_pfn,
5506 unsigned long *ignored)
5508 /* When hotadd a new node from cpu_up(), the node should be empty */
5509 if (!node_start_pfn && !node_end_pfn)
5512 /* Get the start and end of the zone */
5513 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5514 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5515 adjust_zone_range_for_zone_movable(nid, zone_type,
5516 node_start_pfn, node_end_pfn,
5517 zone_start_pfn, zone_end_pfn);
5519 /* Check that this node has pages within the zone's required range */
5520 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5523 /* Move the zone boundaries inside the node if necessary */
5524 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5525 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5527 /* Return the spanned pages */
5528 return *zone_end_pfn - *zone_start_pfn;
5532 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5533 * then all holes in the requested range will be accounted for.
5535 unsigned long __meminit __absent_pages_in_range(int nid,
5536 unsigned long range_start_pfn,
5537 unsigned long range_end_pfn)
5539 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5540 unsigned long start_pfn, end_pfn;
5543 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5544 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5545 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5546 nr_absent -= end_pfn - start_pfn;
5552 * absent_pages_in_range - Return number of page frames in holes within a range
5553 * @start_pfn: The start PFN to start searching for holes
5554 * @end_pfn: The end PFN to stop searching for holes
5556 * It returns the number of pages frames in memory holes within a range.
5558 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5559 unsigned long end_pfn)
5561 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5564 /* Return the number of page frames in holes in a zone on a node */
5565 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5566 unsigned long zone_type,
5567 unsigned long node_start_pfn,
5568 unsigned long node_end_pfn,
5569 unsigned long *ignored)
5571 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5572 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5573 unsigned long zone_start_pfn, zone_end_pfn;
5574 unsigned long nr_absent;
5576 /* When hotadd a new node from cpu_up(), the node should be empty */
5577 if (!node_start_pfn && !node_end_pfn)
5580 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5581 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5583 adjust_zone_range_for_zone_movable(nid, zone_type,
5584 node_start_pfn, node_end_pfn,
5585 &zone_start_pfn, &zone_end_pfn);
5586 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5589 * ZONE_MOVABLE handling.
5590 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5593 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5594 unsigned long start_pfn, end_pfn;
5595 struct memblock_region *r;
5597 for_each_memblock(memory, r) {
5598 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5599 zone_start_pfn, zone_end_pfn);
5600 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5601 zone_start_pfn, zone_end_pfn);
5603 if (zone_type == ZONE_MOVABLE &&
5604 memblock_is_mirror(r))
5605 nr_absent += end_pfn - start_pfn;
5607 if (zone_type == ZONE_NORMAL &&
5608 !memblock_is_mirror(r))
5609 nr_absent += end_pfn - start_pfn;
5616 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5617 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5618 unsigned long zone_type,
5619 unsigned long node_start_pfn,
5620 unsigned long node_end_pfn,
5621 unsigned long *zone_start_pfn,
5622 unsigned long *zone_end_pfn,
5623 unsigned long *zones_size)
5627 *zone_start_pfn = node_start_pfn;
5628 for (zone = 0; zone < zone_type; zone++)
5629 *zone_start_pfn += zones_size[zone];
5631 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5633 return zones_size[zone_type];
5636 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5637 unsigned long zone_type,
5638 unsigned long node_start_pfn,
5639 unsigned long node_end_pfn,
5640 unsigned long *zholes_size)
5645 return zholes_size[zone_type];
5648 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5650 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5651 unsigned long node_start_pfn,
5652 unsigned long node_end_pfn,
5653 unsigned long *zones_size,
5654 unsigned long *zholes_size)
5656 unsigned long realtotalpages = 0, totalpages = 0;
5659 for (i = 0; i < MAX_NR_ZONES; i++) {
5660 struct zone *zone = pgdat->node_zones + i;
5661 unsigned long zone_start_pfn, zone_end_pfn;
5662 unsigned long size, real_size;
5664 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5670 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5671 node_start_pfn, node_end_pfn,
5674 zone->zone_start_pfn = zone_start_pfn;
5676 zone->zone_start_pfn = 0;
5677 zone->spanned_pages = size;
5678 zone->present_pages = real_size;
5681 realtotalpages += real_size;
5684 pgdat->node_spanned_pages = totalpages;
5685 pgdat->node_present_pages = realtotalpages;
5686 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5690 #ifndef CONFIG_SPARSEMEM
5692 * Calculate the size of the zone->blockflags rounded to an unsigned long
5693 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5694 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5695 * round what is now in bits to nearest long in bits, then return it in
5698 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5700 unsigned long usemapsize;
5702 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5703 usemapsize = roundup(zonesize, pageblock_nr_pages);
5704 usemapsize = usemapsize >> pageblock_order;
5705 usemapsize *= NR_PAGEBLOCK_BITS;
5706 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5708 return usemapsize / 8;
5711 static void __init setup_usemap(struct pglist_data *pgdat,
5713 unsigned long zone_start_pfn,
5714 unsigned long zonesize)
5716 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5717 zone->pageblock_flags = NULL;
5719 zone->pageblock_flags =
5720 memblock_virt_alloc_node_nopanic(usemapsize,
5724 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5725 unsigned long zone_start_pfn, unsigned long zonesize) {}
5726 #endif /* CONFIG_SPARSEMEM */
5728 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5730 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5731 void __paginginit set_pageblock_order(void)
5735 /* Check that pageblock_nr_pages has not already been setup */
5736 if (pageblock_order)
5739 if (HPAGE_SHIFT > PAGE_SHIFT)
5740 order = HUGETLB_PAGE_ORDER;
5742 order = MAX_ORDER - 1;
5745 * Assume the largest contiguous order of interest is a huge page.
5746 * This value may be variable depending on boot parameters on IA64 and
5749 pageblock_order = order;
5751 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5754 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5755 * is unused as pageblock_order is set at compile-time. See
5756 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5759 void __paginginit set_pageblock_order(void)
5763 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5765 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5766 unsigned long present_pages)
5768 unsigned long pages = spanned_pages;
5771 * Provide a more accurate estimation if there are holes within
5772 * the zone and SPARSEMEM is in use. If there are holes within the
5773 * zone, each populated memory region may cost us one or two extra
5774 * memmap pages due to alignment because memmap pages for each
5775 * populated regions may not naturally algined on page boundary.
5776 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5778 if (spanned_pages > present_pages + (present_pages >> 4) &&
5779 IS_ENABLED(CONFIG_SPARSEMEM))
5780 pages = present_pages;
5782 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5786 * Set up the zone data structures:
5787 * - mark all pages reserved
5788 * - mark all memory queues empty
5789 * - clear the memory bitmaps
5791 * NOTE: pgdat should get zeroed by caller.
5793 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5796 int nid = pgdat->node_id;
5799 pgdat_resize_init(pgdat);
5800 #ifdef CONFIG_NUMA_BALANCING
5801 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5802 pgdat->numabalancing_migrate_nr_pages = 0;
5803 pgdat->numabalancing_migrate_next_window = jiffies;
5805 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5806 spin_lock_init(&pgdat->split_queue_lock);
5807 INIT_LIST_HEAD(&pgdat->split_queue);
5808 pgdat->split_queue_len = 0;
5810 init_waitqueue_head(&pgdat->kswapd_wait);
5811 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5812 #ifdef CONFIG_COMPACTION
5813 init_waitqueue_head(&pgdat->kcompactd_wait);
5815 pgdat_page_ext_init(pgdat);
5816 spin_lock_init(&pgdat->lru_lock);
5817 lruvec_init(node_lruvec(pgdat));
5819 for (j = 0; j < MAX_NR_ZONES; j++) {
5820 struct zone *zone = pgdat->node_zones + j;
5821 unsigned long size, realsize, freesize, memmap_pages;
5822 unsigned long zone_start_pfn = zone->zone_start_pfn;
5824 size = zone->spanned_pages;
5825 realsize = freesize = zone->present_pages;
5828 * Adjust freesize so that it accounts for how much memory
5829 * is used by this zone for memmap. This affects the watermark
5830 * and per-cpu initialisations
5832 memmap_pages = calc_memmap_size(size, realsize);
5833 if (!is_highmem_idx(j)) {
5834 if (freesize >= memmap_pages) {
5835 freesize -= memmap_pages;
5838 " %s zone: %lu pages used for memmap\n",
5839 zone_names[j], memmap_pages);
5841 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5842 zone_names[j], memmap_pages, freesize);
5845 /* Account for reserved pages */
5846 if (j == 0 && freesize > dma_reserve) {
5847 freesize -= dma_reserve;
5848 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5849 zone_names[0], dma_reserve);
5852 if (!is_highmem_idx(j))
5853 nr_kernel_pages += freesize;
5854 /* Charge for highmem memmap if there are enough kernel pages */
5855 else if (nr_kernel_pages > memmap_pages * 2)
5856 nr_kernel_pages -= memmap_pages;
5857 nr_all_pages += freesize;
5860 * Set an approximate value for lowmem here, it will be adjusted
5861 * when the bootmem allocator frees pages into the buddy system.
5862 * And all highmem pages will be managed by the buddy system.
5864 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5868 zone->name = zone_names[j];
5869 zone->zone_pgdat = pgdat;
5870 spin_lock_init(&zone->lock);
5871 zone_seqlock_init(zone);
5872 zone_pcp_init(zone);
5877 set_pageblock_order();
5878 setup_usemap(pgdat, zone, zone_start_pfn, size);
5879 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5881 memmap_init(size, nid, j, zone_start_pfn);
5885 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5887 unsigned long __maybe_unused start = 0;
5888 unsigned long __maybe_unused offset = 0;
5890 /* Skip empty nodes */
5891 if (!pgdat->node_spanned_pages)
5894 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5895 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5896 offset = pgdat->node_start_pfn - start;
5897 /* ia64 gets its own node_mem_map, before this, without bootmem */
5898 if (!pgdat->node_mem_map) {
5899 unsigned long size, end;
5903 * The zone's endpoints aren't required to be MAX_ORDER
5904 * aligned but the node_mem_map endpoints must be in order
5905 * for the buddy allocator to function correctly.
5907 end = pgdat_end_pfn(pgdat);
5908 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5909 size = (end - start) * sizeof(struct page);
5910 map = alloc_remap(pgdat->node_id, size);
5912 map = memblock_virt_alloc_node_nopanic(size,
5914 pgdat->node_mem_map = map + offset;
5916 #ifndef CONFIG_NEED_MULTIPLE_NODES
5918 * With no DISCONTIG, the global mem_map is just set as node 0's
5920 if (pgdat == NODE_DATA(0)) {
5921 mem_map = NODE_DATA(0)->node_mem_map;
5922 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5923 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5925 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5928 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5931 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5932 unsigned long node_start_pfn, unsigned long *zholes_size)
5934 pg_data_t *pgdat = NODE_DATA(nid);
5935 unsigned long start_pfn = 0;
5936 unsigned long end_pfn = 0;
5938 /* pg_data_t should be reset to zero when it's allocated */
5939 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5941 reset_deferred_meminit(pgdat);
5942 pgdat->node_id = nid;
5943 pgdat->node_start_pfn = node_start_pfn;
5944 pgdat->per_cpu_nodestats = NULL;
5945 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5946 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5947 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5948 (u64)start_pfn << PAGE_SHIFT,
5949 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5951 start_pfn = node_start_pfn;
5953 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5954 zones_size, zholes_size);
5956 alloc_node_mem_map(pgdat);
5957 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5958 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5959 nid, (unsigned long)pgdat,
5960 (unsigned long)pgdat->node_mem_map);
5963 free_area_init_core(pgdat);
5966 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5968 #if MAX_NUMNODES > 1
5970 * Figure out the number of possible node ids.
5972 void __init setup_nr_node_ids(void)
5974 unsigned int highest;
5976 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5977 nr_node_ids = highest + 1;
5982 * node_map_pfn_alignment - determine the maximum internode alignment
5984 * This function should be called after node map is populated and sorted.
5985 * It calculates the maximum power of two alignment which can distinguish
5988 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5989 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5990 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5991 * shifted, 1GiB is enough and this function will indicate so.
5993 * This is used to test whether pfn -> nid mapping of the chosen memory
5994 * model has fine enough granularity to avoid incorrect mapping for the
5995 * populated node map.
5997 * Returns the determined alignment in pfn's. 0 if there is no alignment
5998 * requirement (single node).
6000 unsigned long __init node_map_pfn_alignment(void)
6002 unsigned long accl_mask = 0, last_end = 0;
6003 unsigned long start, end, mask;
6007 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6008 if (!start || last_nid < 0 || last_nid == nid) {
6015 * Start with a mask granular enough to pin-point to the
6016 * start pfn and tick off bits one-by-one until it becomes
6017 * too coarse to separate the current node from the last.
6019 mask = ~((1 << __ffs(start)) - 1);
6020 while (mask && last_end <= (start & (mask << 1)))
6023 /* accumulate all internode masks */
6027 /* convert mask to number of pages */
6028 return ~accl_mask + 1;
6031 /* Find the lowest pfn for a node */
6032 static unsigned long __init find_min_pfn_for_node(int nid)
6034 unsigned long min_pfn = ULONG_MAX;
6035 unsigned long start_pfn;
6038 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6039 min_pfn = min(min_pfn, start_pfn);
6041 if (min_pfn == ULONG_MAX) {
6042 pr_warn("Could not find start_pfn for node %d\n", nid);
6050 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6052 * It returns the minimum PFN based on information provided via
6053 * memblock_set_node().
6055 unsigned long __init find_min_pfn_with_active_regions(void)
6057 return find_min_pfn_for_node(MAX_NUMNODES);
6061 * early_calculate_totalpages()
6062 * Sum pages in active regions for movable zone.
6063 * Populate N_MEMORY for calculating usable_nodes.
6065 static unsigned long __init early_calculate_totalpages(void)
6067 unsigned long totalpages = 0;
6068 unsigned long start_pfn, end_pfn;
6071 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6072 unsigned long pages = end_pfn - start_pfn;
6074 totalpages += pages;
6076 node_set_state(nid, N_MEMORY);
6082 * Find the PFN the Movable zone begins in each node. Kernel memory
6083 * is spread evenly between nodes as long as the nodes have enough
6084 * memory. When they don't, some nodes will have more kernelcore than
6087 static void __init find_zone_movable_pfns_for_nodes(void)
6090 unsigned long usable_startpfn;
6091 unsigned long kernelcore_node, kernelcore_remaining;
6092 /* save the state before borrow the nodemask */
6093 nodemask_t saved_node_state = node_states[N_MEMORY];
6094 unsigned long totalpages = early_calculate_totalpages();
6095 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6096 struct memblock_region *r;
6098 /* Need to find movable_zone earlier when movable_node is specified. */
6099 find_usable_zone_for_movable();
6102 * If movable_node is specified, ignore kernelcore and movablecore
6105 if (movable_node_is_enabled()) {
6106 for_each_memblock(memory, r) {
6107 if (!memblock_is_hotpluggable(r))
6112 usable_startpfn = PFN_DOWN(r->base);
6113 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6114 min(usable_startpfn, zone_movable_pfn[nid]) :
6122 * If kernelcore=mirror is specified, ignore movablecore option
6124 if (mirrored_kernelcore) {
6125 bool mem_below_4gb_not_mirrored = false;
6127 for_each_memblock(memory, r) {
6128 if (memblock_is_mirror(r))
6133 usable_startpfn = memblock_region_memory_base_pfn(r);
6135 if (usable_startpfn < 0x100000) {
6136 mem_below_4gb_not_mirrored = true;
6140 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6141 min(usable_startpfn, zone_movable_pfn[nid]) :
6145 if (mem_below_4gb_not_mirrored)
6146 pr_warn("This configuration results in unmirrored kernel memory.");
6152 * If movablecore=nn[KMG] was specified, calculate what size of
6153 * kernelcore that corresponds so that memory usable for
6154 * any allocation type is evenly spread. If both kernelcore
6155 * and movablecore are specified, then the value of kernelcore
6156 * will be used for required_kernelcore if it's greater than
6157 * what movablecore would have allowed.
6159 if (required_movablecore) {
6160 unsigned long corepages;
6163 * Round-up so that ZONE_MOVABLE is at least as large as what
6164 * was requested by the user
6166 required_movablecore =
6167 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6168 required_movablecore = min(totalpages, required_movablecore);
6169 corepages = totalpages - required_movablecore;
6171 required_kernelcore = max(required_kernelcore, corepages);
6175 * If kernelcore was not specified or kernelcore size is larger
6176 * than totalpages, there is no ZONE_MOVABLE.
6178 if (!required_kernelcore || required_kernelcore >= totalpages)
6181 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6182 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6185 /* Spread kernelcore memory as evenly as possible throughout nodes */
6186 kernelcore_node = required_kernelcore / usable_nodes;
6187 for_each_node_state(nid, N_MEMORY) {
6188 unsigned long start_pfn, end_pfn;
6191 * Recalculate kernelcore_node if the division per node
6192 * now exceeds what is necessary to satisfy the requested
6193 * amount of memory for the kernel
6195 if (required_kernelcore < kernelcore_node)
6196 kernelcore_node = required_kernelcore / usable_nodes;
6199 * As the map is walked, we track how much memory is usable
6200 * by the kernel using kernelcore_remaining. When it is
6201 * 0, the rest of the node is usable by ZONE_MOVABLE
6203 kernelcore_remaining = kernelcore_node;
6205 /* Go through each range of PFNs within this node */
6206 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6207 unsigned long size_pages;
6209 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6210 if (start_pfn >= end_pfn)
6213 /* Account for what is only usable for kernelcore */
6214 if (start_pfn < usable_startpfn) {
6215 unsigned long kernel_pages;
6216 kernel_pages = min(end_pfn, usable_startpfn)
6219 kernelcore_remaining -= min(kernel_pages,
6220 kernelcore_remaining);
6221 required_kernelcore -= min(kernel_pages,
6222 required_kernelcore);
6224 /* Continue if range is now fully accounted */
6225 if (end_pfn <= usable_startpfn) {
6228 * Push zone_movable_pfn to the end so
6229 * that if we have to rebalance
6230 * kernelcore across nodes, we will
6231 * not double account here
6233 zone_movable_pfn[nid] = end_pfn;
6236 start_pfn = usable_startpfn;
6240 * The usable PFN range for ZONE_MOVABLE is from
6241 * start_pfn->end_pfn. Calculate size_pages as the
6242 * number of pages used as kernelcore
6244 size_pages = end_pfn - start_pfn;
6245 if (size_pages > kernelcore_remaining)
6246 size_pages = kernelcore_remaining;
6247 zone_movable_pfn[nid] = start_pfn + size_pages;
6250 * Some kernelcore has been met, update counts and
6251 * break if the kernelcore for this node has been
6254 required_kernelcore -= min(required_kernelcore,
6256 kernelcore_remaining -= size_pages;
6257 if (!kernelcore_remaining)
6263 * If there is still required_kernelcore, we do another pass with one
6264 * less node in the count. This will push zone_movable_pfn[nid] further
6265 * along on the nodes that still have memory until kernelcore is
6269 if (usable_nodes && required_kernelcore > usable_nodes)
6273 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6274 for (nid = 0; nid < MAX_NUMNODES; nid++)
6275 zone_movable_pfn[nid] =
6276 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6279 /* restore the node_state */
6280 node_states[N_MEMORY] = saved_node_state;
6283 /* Any regular or high memory on that node ? */
6284 static void check_for_memory(pg_data_t *pgdat, int nid)
6286 enum zone_type zone_type;
6288 if (N_MEMORY == N_NORMAL_MEMORY)
6291 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6292 struct zone *zone = &pgdat->node_zones[zone_type];
6293 if (populated_zone(zone)) {
6294 node_set_state(nid, N_HIGH_MEMORY);
6295 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6296 zone_type <= ZONE_NORMAL)
6297 node_set_state(nid, N_NORMAL_MEMORY);
6304 * free_area_init_nodes - Initialise all pg_data_t and zone data
6305 * @max_zone_pfn: an array of max PFNs for each zone
6307 * This will call free_area_init_node() for each active node in the system.
6308 * Using the page ranges provided by memblock_set_node(), the size of each
6309 * zone in each node and their holes is calculated. If the maximum PFN
6310 * between two adjacent zones match, it is assumed that the zone is empty.
6311 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6312 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6313 * starts where the previous one ended. For example, ZONE_DMA32 starts
6314 * at arch_max_dma_pfn.
6316 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6318 unsigned long start_pfn, end_pfn;
6321 /* Record where the zone boundaries are */
6322 memset(arch_zone_lowest_possible_pfn, 0,
6323 sizeof(arch_zone_lowest_possible_pfn));
6324 memset(arch_zone_highest_possible_pfn, 0,
6325 sizeof(arch_zone_highest_possible_pfn));
6327 start_pfn = find_min_pfn_with_active_regions();
6329 for (i = 0; i < MAX_NR_ZONES; i++) {
6330 if (i == ZONE_MOVABLE)
6333 end_pfn = max(max_zone_pfn[i], start_pfn);
6334 arch_zone_lowest_possible_pfn[i] = start_pfn;
6335 arch_zone_highest_possible_pfn[i] = end_pfn;
6337 start_pfn = end_pfn;
6339 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6340 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6342 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6343 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6344 find_zone_movable_pfns_for_nodes();
6346 /* Print out the zone ranges */
6347 pr_info("Zone ranges:\n");
6348 for (i = 0; i < MAX_NR_ZONES; i++) {
6349 if (i == ZONE_MOVABLE)
6351 pr_info(" %-8s ", zone_names[i]);
6352 if (arch_zone_lowest_possible_pfn[i] ==
6353 arch_zone_highest_possible_pfn[i])
6356 pr_cont("[mem %#018Lx-%#018Lx]\n",
6357 (u64)arch_zone_lowest_possible_pfn[i]
6359 ((u64)arch_zone_highest_possible_pfn[i]
6360 << PAGE_SHIFT) - 1);
6363 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6364 pr_info("Movable zone start for each node\n");
6365 for (i = 0; i < MAX_NUMNODES; i++) {
6366 if (zone_movable_pfn[i])
6367 pr_info(" Node %d: %#018Lx\n", i,
6368 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6371 /* Print out the early node map */
6372 pr_info("Early memory node ranges\n");
6373 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6374 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6375 (u64)start_pfn << PAGE_SHIFT,
6376 ((u64)end_pfn << PAGE_SHIFT) - 1);
6378 /* Initialise every node */
6379 mminit_verify_pageflags_layout();
6380 setup_nr_node_ids();
6381 for_each_online_node(nid) {
6382 pg_data_t *pgdat = NODE_DATA(nid);
6383 free_area_init_node(nid, NULL,
6384 find_min_pfn_for_node(nid), NULL);
6386 /* Any memory on that node */
6387 if (pgdat->node_present_pages)
6388 node_set_state(nid, N_MEMORY);
6389 check_for_memory(pgdat, nid);
6393 static int __init cmdline_parse_core(char *p, unsigned long *core)
6395 unsigned long long coremem;
6399 coremem = memparse(p, &p);
6400 *core = coremem >> PAGE_SHIFT;
6402 /* Paranoid check that UL is enough for the coremem value */
6403 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6409 * kernelcore=size sets the amount of memory for use for allocations that
6410 * cannot be reclaimed or migrated.
6412 static int __init cmdline_parse_kernelcore(char *p)
6414 /* parse kernelcore=mirror */
6415 if (parse_option_str(p, "mirror")) {
6416 mirrored_kernelcore = true;
6420 return cmdline_parse_core(p, &required_kernelcore);
6424 * movablecore=size sets the amount of memory for use for allocations that
6425 * can be reclaimed or migrated.
6427 static int __init cmdline_parse_movablecore(char *p)
6429 return cmdline_parse_core(p, &required_movablecore);
6432 early_param("kernelcore", cmdline_parse_kernelcore);
6433 early_param("movablecore", cmdline_parse_movablecore);
6435 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6437 void adjust_managed_page_count(struct page *page, long count)
6439 spin_lock(&managed_page_count_lock);
6440 page_zone(page)->managed_pages += count;
6441 totalram_pages += count;
6442 #ifdef CONFIG_HIGHMEM
6443 if (PageHighMem(page))
6444 totalhigh_pages += count;
6446 spin_unlock(&managed_page_count_lock);
6448 EXPORT_SYMBOL(adjust_managed_page_count);
6450 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6453 unsigned long pages = 0;
6455 start = (void *)PAGE_ALIGN((unsigned long)start);
6456 end = (void *)((unsigned long)end & PAGE_MASK);
6457 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6458 if ((unsigned int)poison <= 0xFF)
6459 memset(pos, poison, PAGE_SIZE);
6460 free_reserved_page(virt_to_page(pos));
6464 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6465 s, pages << (PAGE_SHIFT - 10), start, end);
6469 EXPORT_SYMBOL(free_reserved_area);
6471 #ifdef CONFIG_HIGHMEM
6472 void free_highmem_page(struct page *page)
6474 __free_reserved_page(page);
6476 page_zone(page)->managed_pages++;
6482 void __init mem_init_print_info(const char *str)
6484 unsigned long physpages, codesize, datasize, rosize, bss_size;
6485 unsigned long init_code_size, init_data_size;
6487 physpages = get_num_physpages();
6488 codesize = _etext - _stext;
6489 datasize = _edata - _sdata;
6490 rosize = __end_rodata - __start_rodata;
6491 bss_size = __bss_stop - __bss_start;
6492 init_data_size = __init_end - __init_begin;
6493 init_code_size = _einittext - _sinittext;
6496 * Detect special cases and adjust section sizes accordingly:
6497 * 1) .init.* may be embedded into .data sections
6498 * 2) .init.text.* may be out of [__init_begin, __init_end],
6499 * please refer to arch/tile/kernel/vmlinux.lds.S.
6500 * 3) .rodata.* may be embedded into .text or .data sections.
6502 #define adj_init_size(start, end, size, pos, adj) \
6504 if (start <= pos && pos < end && size > adj) \
6508 adj_init_size(__init_begin, __init_end, init_data_size,
6509 _sinittext, init_code_size);
6510 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6511 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6512 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6513 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6515 #undef adj_init_size
6517 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6518 #ifdef CONFIG_HIGHMEM
6522 nr_free_pages() << (PAGE_SHIFT - 10),
6523 physpages << (PAGE_SHIFT - 10),
6524 codesize >> 10, datasize >> 10, rosize >> 10,
6525 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6526 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6527 totalcma_pages << (PAGE_SHIFT - 10),
6528 #ifdef CONFIG_HIGHMEM
6529 totalhigh_pages << (PAGE_SHIFT - 10),
6531 str ? ", " : "", str ? str : "");
6535 * set_dma_reserve - set the specified number of pages reserved in the first zone
6536 * @new_dma_reserve: The number of pages to mark reserved
6538 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6539 * In the DMA zone, a significant percentage may be consumed by kernel image
6540 * and other unfreeable allocations which can skew the watermarks badly. This
6541 * function may optionally be used to account for unfreeable pages in the
6542 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6543 * smaller per-cpu batchsize.
6545 void __init set_dma_reserve(unsigned long new_dma_reserve)
6547 dma_reserve = new_dma_reserve;
6550 void __init free_area_init(unsigned long *zones_size)
6552 free_area_init_node(0, zones_size,
6553 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6556 static int page_alloc_cpu_notify(struct notifier_block *self,
6557 unsigned long action, void *hcpu)
6559 int cpu = (unsigned long)hcpu;
6561 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6562 local_lock_irq_on(swapvec_lock, cpu);
6563 lru_add_drain_cpu(cpu);
6564 local_unlock_irq_on(swapvec_lock, cpu);
6568 * Spill the event counters of the dead processor
6569 * into the current processors event counters.
6570 * This artificially elevates the count of the current
6573 vm_events_fold_cpu(cpu);
6576 * Zero the differential counters of the dead processor
6577 * so that the vm statistics are consistent.
6579 * This is only okay since the processor is dead and cannot
6580 * race with what we are doing.
6582 cpu_vm_stats_fold(cpu);
6587 void __init page_alloc_init(void)
6589 hotcpu_notifier(page_alloc_cpu_notify, 0);
6590 local_irq_lock_init(pa_lock);
6594 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6595 * or min_free_kbytes changes.
6597 static void calculate_totalreserve_pages(void)
6599 struct pglist_data *pgdat;
6600 unsigned long reserve_pages = 0;
6601 enum zone_type i, j;
6603 for_each_online_pgdat(pgdat) {
6605 pgdat->totalreserve_pages = 0;
6607 for (i = 0; i < MAX_NR_ZONES; i++) {
6608 struct zone *zone = pgdat->node_zones + i;
6611 /* Find valid and maximum lowmem_reserve in the zone */
6612 for (j = i; j < MAX_NR_ZONES; j++) {
6613 if (zone->lowmem_reserve[j] > max)
6614 max = zone->lowmem_reserve[j];
6617 /* we treat the high watermark as reserved pages. */
6618 max += high_wmark_pages(zone);
6620 if (max > zone->managed_pages)
6621 max = zone->managed_pages;
6623 pgdat->totalreserve_pages += max;
6625 reserve_pages += max;
6628 totalreserve_pages = reserve_pages;
6632 * setup_per_zone_lowmem_reserve - called whenever
6633 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6634 * has a correct pages reserved value, so an adequate number of
6635 * pages are left in the zone after a successful __alloc_pages().
6637 static void setup_per_zone_lowmem_reserve(void)
6639 struct pglist_data *pgdat;
6640 enum zone_type j, idx;
6642 for_each_online_pgdat(pgdat) {
6643 for (j = 0; j < MAX_NR_ZONES; j++) {
6644 struct zone *zone = pgdat->node_zones + j;
6645 unsigned long managed_pages = zone->managed_pages;
6647 zone->lowmem_reserve[j] = 0;
6651 struct zone *lower_zone;
6655 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6656 sysctl_lowmem_reserve_ratio[idx] = 1;
6658 lower_zone = pgdat->node_zones + idx;
6659 lower_zone->lowmem_reserve[j] = managed_pages /
6660 sysctl_lowmem_reserve_ratio[idx];
6661 managed_pages += lower_zone->managed_pages;
6666 /* update totalreserve_pages */
6667 calculate_totalreserve_pages();
6670 static void __setup_per_zone_wmarks(void)
6672 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6673 unsigned long lowmem_pages = 0;
6675 unsigned long flags;
6677 /* Calculate total number of !ZONE_HIGHMEM pages */
6678 for_each_zone(zone) {
6679 if (!is_highmem(zone))
6680 lowmem_pages += zone->managed_pages;
6683 for_each_zone(zone) {
6686 spin_lock_irqsave(&zone->lock, flags);
6687 tmp = (u64)pages_min * zone->managed_pages;
6688 do_div(tmp, lowmem_pages);
6689 if (is_highmem(zone)) {
6691 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6692 * need highmem pages, so cap pages_min to a small
6695 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6696 * deltas control asynch page reclaim, and so should
6697 * not be capped for highmem.
6699 unsigned long min_pages;
6701 min_pages = zone->managed_pages / 1024;
6702 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6703 zone->watermark[WMARK_MIN] = min_pages;
6706 * If it's a lowmem zone, reserve a number of pages
6707 * proportionate to the zone's size.
6709 zone->watermark[WMARK_MIN] = tmp;
6713 * Set the kswapd watermarks distance according to the
6714 * scale factor in proportion to available memory, but
6715 * ensure a minimum size on small systems.
6717 tmp = max_t(u64, tmp >> 2,
6718 mult_frac(zone->managed_pages,
6719 watermark_scale_factor, 10000));
6721 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6722 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6724 spin_unlock_irqrestore(&zone->lock, flags);
6727 /* update totalreserve_pages */
6728 calculate_totalreserve_pages();
6732 * setup_per_zone_wmarks - called when min_free_kbytes changes
6733 * or when memory is hot-{added|removed}
6735 * Ensures that the watermark[min,low,high] values for each zone are set
6736 * correctly with respect to min_free_kbytes.
6738 void setup_per_zone_wmarks(void)
6740 mutex_lock(&zonelists_mutex);
6741 __setup_per_zone_wmarks();
6742 mutex_unlock(&zonelists_mutex);
6746 * Initialise min_free_kbytes.
6748 * For small machines we want it small (128k min). For large machines
6749 * we want it large (64MB max). But it is not linear, because network
6750 * bandwidth does not increase linearly with machine size. We use
6752 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6753 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6769 int __meminit init_per_zone_wmark_min(void)
6771 unsigned long lowmem_kbytes;
6772 int new_min_free_kbytes;
6774 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6775 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6777 if (new_min_free_kbytes > user_min_free_kbytes) {
6778 min_free_kbytes = new_min_free_kbytes;
6779 if (min_free_kbytes < 128)
6780 min_free_kbytes = 128;
6781 if (min_free_kbytes > 65536)
6782 min_free_kbytes = 65536;
6784 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6785 new_min_free_kbytes, user_min_free_kbytes);
6787 setup_per_zone_wmarks();
6788 refresh_zone_stat_thresholds();
6789 setup_per_zone_lowmem_reserve();
6792 setup_min_unmapped_ratio();
6793 setup_min_slab_ratio();
6798 core_initcall(init_per_zone_wmark_min)
6801 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6802 * that we can call two helper functions whenever min_free_kbytes
6805 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6806 void __user *buffer, size_t *length, loff_t *ppos)
6810 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6815 user_min_free_kbytes = min_free_kbytes;
6816 setup_per_zone_wmarks();
6821 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6822 void __user *buffer, size_t *length, loff_t *ppos)
6826 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6831 setup_per_zone_wmarks();
6837 static void setup_min_unmapped_ratio(void)
6842 for_each_online_pgdat(pgdat)
6843 pgdat->min_unmapped_pages = 0;
6846 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6847 sysctl_min_unmapped_ratio) / 100;
6851 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6852 void __user *buffer, size_t *length, loff_t *ppos)
6856 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6860 setup_min_unmapped_ratio();
6865 static void setup_min_slab_ratio(void)
6870 for_each_online_pgdat(pgdat)
6871 pgdat->min_slab_pages = 0;
6874 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6875 sysctl_min_slab_ratio) / 100;
6878 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6879 void __user *buffer, size_t *length, loff_t *ppos)
6883 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6887 setup_min_slab_ratio();
6894 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6895 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6896 * whenever sysctl_lowmem_reserve_ratio changes.
6898 * The reserve ratio obviously has absolutely no relation with the
6899 * minimum watermarks. The lowmem reserve ratio can only make sense
6900 * if in function of the boot time zone sizes.
6902 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6903 void __user *buffer, size_t *length, loff_t *ppos)
6905 proc_dointvec_minmax(table, write, buffer, length, ppos);
6906 setup_per_zone_lowmem_reserve();
6911 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6912 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6913 * pagelist can have before it gets flushed back to buddy allocator.
6915 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6916 void __user *buffer, size_t *length, loff_t *ppos)
6919 int old_percpu_pagelist_fraction;
6922 mutex_lock(&pcp_batch_high_lock);
6923 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6925 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6926 if (!write || ret < 0)
6929 /* Sanity checking to avoid pcp imbalance */
6930 if (percpu_pagelist_fraction &&
6931 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6932 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6938 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6941 for_each_populated_zone(zone) {
6944 for_each_possible_cpu(cpu)
6945 pageset_set_high_and_batch(zone,
6946 per_cpu_ptr(zone->pageset, cpu));
6949 mutex_unlock(&pcp_batch_high_lock);
6954 int hashdist = HASHDIST_DEFAULT;
6956 static int __init set_hashdist(char *str)
6960 hashdist = simple_strtoul(str, &str, 0);
6963 __setup("hashdist=", set_hashdist);
6966 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6968 * Returns the number of pages that arch has reserved but
6969 * is not known to alloc_large_system_hash().
6971 static unsigned long __init arch_reserved_kernel_pages(void)
6978 * allocate a large system hash table from bootmem
6979 * - it is assumed that the hash table must contain an exact power-of-2
6980 * quantity of entries
6981 * - limit is the number of hash buckets, not the total allocation size
6983 void *__init alloc_large_system_hash(const char *tablename,
6984 unsigned long bucketsize,
6985 unsigned long numentries,
6988 unsigned int *_hash_shift,
6989 unsigned int *_hash_mask,
6990 unsigned long low_limit,
6991 unsigned long high_limit)
6993 unsigned long long max = high_limit;
6994 unsigned long log2qty, size;
6997 /* allow the kernel cmdline to have a say */
6999 /* round applicable memory size up to nearest megabyte */
7000 numentries = nr_kernel_pages;
7001 numentries -= arch_reserved_kernel_pages();
7003 /* It isn't necessary when PAGE_SIZE >= 1MB */
7004 if (PAGE_SHIFT < 20)
7005 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7007 /* limit to 1 bucket per 2^scale bytes of low memory */
7008 if (scale > PAGE_SHIFT)
7009 numentries >>= (scale - PAGE_SHIFT);
7011 numentries <<= (PAGE_SHIFT - scale);
7013 /* Make sure we've got at least a 0-order allocation.. */
7014 if (unlikely(flags & HASH_SMALL)) {
7015 /* Makes no sense without HASH_EARLY */
7016 WARN_ON(!(flags & HASH_EARLY));
7017 if (!(numentries >> *_hash_shift)) {
7018 numentries = 1UL << *_hash_shift;
7019 BUG_ON(!numentries);
7021 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7022 numentries = PAGE_SIZE / bucketsize;
7024 numentries = roundup_pow_of_two(numentries);
7026 /* limit allocation size to 1/16 total memory by default */
7028 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7029 do_div(max, bucketsize);
7031 max = min(max, 0x80000000ULL);
7033 if (numentries < low_limit)
7034 numentries = low_limit;
7035 if (numentries > max)
7038 log2qty = ilog2(numentries);
7041 size = bucketsize << log2qty;
7042 if (flags & HASH_EARLY)
7043 table = memblock_virt_alloc_nopanic(size, 0);
7045 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7048 * If bucketsize is not a power-of-two, we may free
7049 * some pages at the end of hash table which
7050 * alloc_pages_exact() automatically does
7052 if (get_order(size) < MAX_ORDER) {
7053 table = alloc_pages_exact(size, GFP_ATOMIC);
7054 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7057 } while (!table && size > PAGE_SIZE && --log2qty);
7060 panic("Failed to allocate %s hash table\n", tablename);
7062 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7063 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7066 *_hash_shift = log2qty;
7068 *_hash_mask = (1 << log2qty) - 1;
7074 * This function checks whether pageblock includes unmovable pages or not.
7075 * If @count is not zero, it is okay to include less @count unmovable pages
7077 * PageLRU check without isolation or lru_lock could race so that
7078 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7079 * expect this function should be exact.
7081 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7082 bool skip_hwpoisoned_pages)
7084 unsigned long pfn, iter, found;
7088 * For avoiding noise data, lru_add_drain_all() should be called
7089 * If ZONE_MOVABLE, the zone never contains unmovable pages
7091 if (zone_idx(zone) == ZONE_MOVABLE)
7093 mt = get_pageblock_migratetype(page);
7094 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7097 pfn = page_to_pfn(page);
7098 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7099 unsigned long check = pfn + iter;
7101 if (!pfn_valid_within(check))
7104 page = pfn_to_page(check);
7107 * Hugepages are not in LRU lists, but they're movable.
7108 * We need not scan over tail pages bacause we don't
7109 * handle each tail page individually in migration.
7111 if (PageHuge(page)) {
7112 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7117 * We can't use page_count without pin a page
7118 * because another CPU can free compound page.
7119 * This check already skips compound tails of THP
7120 * because their page->_refcount is zero at all time.
7122 if (!page_ref_count(page)) {
7123 if (PageBuddy(page))
7124 iter += (1 << page_order(page)) - 1;
7129 * The HWPoisoned page may be not in buddy system, and
7130 * page_count() is not 0.
7132 if (skip_hwpoisoned_pages && PageHWPoison(page))
7138 * If there are RECLAIMABLE pages, we need to check
7139 * it. But now, memory offline itself doesn't call
7140 * shrink_node_slabs() and it still to be fixed.
7143 * If the page is not RAM, page_count()should be 0.
7144 * we don't need more check. This is an _used_ not-movable page.
7146 * The problematic thing here is PG_reserved pages. PG_reserved
7147 * is set to both of a memory hole page and a _used_ kernel
7156 bool is_pageblock_removable_nolock(struct page *page)
7162 * We have to be careful here because we are iterating over memory
7163 * sections which are not zone aware so we might end up outside of
7164 * the zone but still within the section.
7165 * We have to take care about the node as well. If the node is offline
7166 * its NODE_DATA will be NULL - see page_zone.
7168 if (!node_online(page_to_nid(page)))
7171 zone = page_zone(page);
7172 pfn = page_to_pfn(page);
7173 if (!zone_spans_pfn(zone, pfn))
7176 return !has_unmovable_pages(zone, page, 0, true);
7179 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7181 static unsigned long pfn_max_align_down(unsigned long pfn)
7183 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7184 pageblock_nr_pages) - 1);
7187 static unsigned long pfn_max_align_up(unsigned long pfn)
7189 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7190 pageblock_nr_pages));
7193 /* [start, end) must belong to a single zone. */
7194 static int __alloc_contig_migrate_range(struct compact_control *cc,
7195 unsigned long start, unsigned long end)
7197 /* This function is based on compact_zone() from compaction.c. */
7198 unsigned long nr_reclaimed;
7199 unsigned long pfn = start;
7200 unsigned int tries = 0;
7205 while (pfn < end || !list_empty(&cc->migratepages)) {
7206 if (fatal_signal_pending(current)) {
7211 if (list_empty(&cc->migratepages)) {
7212 cc->nr_migratepages = 0;
7213 pfn = isolate_migratepages_range(cc, pfn, end);
7219 } else if (++tries == 5) {
7220 ret = ret < 0 ? ret : -EBUSY;
7224 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7226 cc->nr_migratepages -= nr_reclaimed;
7228 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7229 NULL, 0, cc->mode, MR_CMA);
7232 putback_movable_pages(&cc->migratepages);
7239 * alloc_contig_range() -- tries to allocate given range of pages
7240 * @start: start PFN to allocate
7241 * @end: one-past-the-last PFN to allocate
7242 * @migratetype: migratetype of the underlaying pageblocks (either
7243 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7244 * in range must have the same migratetype and it must
7245 * be either of the two.
7247 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7248 * aligned, however it's the caller's responsibility to guarantee that
7249 * we are the only thread that changes migrate type of pageblocks the
7252 * The PFN range must belong to a single zone.
7254 * Returns zero on success or negative error code. On success all
7255 * pages which PFN is in [start, end) are allocated for the caller and
7256 * need to be freed with free_contig_range().
7258 int alloc_contig_range(unsigned long start, unsigned long end,
7259 unsigned migratetype)
7261 unsigned long outer_start, outer_end;
7265 struct compact_control cc = {
7266 .nr_migratepages = 0,
7268 .zone = page_zone(pfn_to_page(start)),
7269 .mode = MIGRATE_SYNC,
7270 .ignore_skip_hint = true,
7272 INIT_LIST_HEAD(&cc.migratepages);
7275 * What we do here is we mark all pageblocks in range as
7276 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7277 * have different sizes, and due to the way page allocator
7278 * work, we align the range to biggest of the two pages so
7279 * that page allocator won't try to merge buddies from
7280 * different pageblocks and change MIGRATE_ISOLATE to some
7281 * other migration type.
7283 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7284 * migrate the pages from an unaligned range (ie. pages that
7285 * we are interested in). This will put all the pages in
7286 * range back to page allocator as MIGRATE_ISOLATE.
7288 * When this is done, we take the pages in range from page
7289 * allocator removing them from the buddy system. This way
7290 * page allocator will never consider using them.
7292 * This lets us mark the pageblocks back as
7293 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7294 * aligned range but not in the unaligned, original range are
7295 * put back to page allocator so that buddy can use them.
7298 ret = start_isolate_page_range(pfn_max_align_down(start),
7299 pfn_max_align_up(end), migratetype,
7305 * In case of -EBUSY, we'd like to know which page causes problem.
7306 * So, just fall through. We will check it in test_pages_isolated().
7308 ret = __alloc_contig_migrate_range(&cc, start, end);
7309 if (ret && ret != -EBUSY)
7313 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7314 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7315 * more, all pages in [start, end) are free in page allocator.
7316 * What we are going to do is to allocate all pages from
7317 * [start, end) (that is remove them from page allocator).
7319 * The only problem is that pages at the beginning and at the
7320 * end of interesting range may be not aligned with pages that
7321 * page allocator holds, ie. they can be part of higher order
7322 * pages. Because of this, we reserve the bigger range and
7323 * once this is done free the pages we are not interested in.
7325 * We don't have to hold zone->lock here because the pages are
7326 * isolated thus they won't get removed from buddy.
7329 lru_add_drain_all();
7330 drain_all_pages(cc.zone);
7333 outer_start = start;
7334 while (!PageBuddy(pfn_to_page(outer_start))) {
7335 if (++order >= MAX_ORDER) {
7336 outer_start = start;
7339 outer_start &= ~0UL << order;
7342 if (outer_start != start) {
7343 order = page_order(pfn_to_page(outer_start));
7346 * outer_start page could be small order buddy page and
7347 * it doesn't include start page. Adjust outer_start
7348 * in this case to report failed page properly
7349 * on tracepoint in test_pages_isolated()
7351 if (outer_start + (1UL << order) <= start)
7352 outer_start = start;
7355 /* Make sure the range is really isolated. */
7356 if (test_pages_isolated(outer_start, end, false)) {
7357 pr_debug("%s: [%lx, %lx) PFNs busy\n",
7358 __func__, outer_start, end);
7363 /* Grab isolated pages from freelists. */
7364 outer_end = isolate_freepages_range(&cc, outer_start, end);
7370 /* Free head and tail (if any) */
7371 if (start != outer_start)
7372 free_contig_range(outer_start, start - outer_start);
7373 if (end != outer_end)
7374 free_contig_range(end, outer_end - end);
7377 undo_isolate_page_range(pfn_max_align_down(start),
7378 pfn_max_align_up(end), migratetype);
7382 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7384 unsigned int count = 0;
7386 for (; nr_pages--; pfn++) {
7387 struct page *page = pfn_to_page(pfn);
7389 count += page_count(page) != 1;
7392 WARN(count != 0, "%d pages are still in use!\n", count);
7396 #ifdef CONFIG_MEMORY_HOTPLUG
7398 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7399 * page high values need to be recalulated.
7401 void __meminit zone_pcp_update(struct zone *zone)
7404 mutex_lock(&pcp_batch_high_lock);
7405 for_each_possible_cpu(cpu)
7406 pageset_set_high_and_batch(zone,
7407 per_cpu_ptr(zone->pageset, cpu));
7408 mutex_unlock(&pcp_batch_high_lock);
7412 void zone_pcp_reset(struct zone *zone)
7414 unsigned long flags;
7416 struct per_cpu_pageset *pset;
7418 /* avoid races with drain_pages() */
7419 local_lock_irqsave(pa_lock, flags);
7420 if (zone->pageset != &boot_pageset) {
7421 for_each_online_cpu(cpu) {
7422 pset = per_cpu_ptr(zone->pageset, cpu);
7423 drain_zonestat(zone, pset);
7425 free_percpu(zone->pageset);
7426 zone->pageset = &boot_pageset;
7428 local_unlock_irqrestore(pa_lock, flags);
7431 #ifdef CONFIG_MEMORY_HOTREMOVE
7433 * All pages in the range must be in a single zone and isolated
7434 * before calling this.
7437 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7441 unsigned int order, i;
7443 unsigned long flags;
7444 /* find the first valid pfn */
7445 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7450 zone = page_zone(pfn_to_page(pfn));
7451 spin_lock_irqsave(&zone->lock, flags);
7453 while (pfn < end_pfn) {
7454 if (!pfn_valid(pfn)) {
7458 page = pfn_to_page(pfn);
7460 * The HWPoisoned page may be not in buddy system, and
7461 * page_count() is not 0.
7463 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7465 SetPageReserved(page);
7469 BUG_ON(page_count(page));
7470 BUG_ON(!PageBuddy(page));
7471 order = page_order(page);
7472 #ifdef CONFIG_DEBUG_VM
7473 pr_info("remove from free list %lx %d %lx\n",
7474 pfn, 1 << order, end_pfn);
7476 list_del(&page->lru);
7477 rmv_page_order(page);
7478 zone->free_area[order].nr_free--;
7479 for (i = 0; i < (1 << order); i++)
7480 SetPageReserved((page+i));
7481 pfn += (1 << order);
7483 spin_unlock_irqrestore(&zone->lock, flags);
7487 bool is_free_buddy_page(struct page *page)
7489 struct zone *zone = page_zone(page);
7490 unsigned long pfn = page_to_pfn(page);
7491 unsigned long flags;
7494 spin_lock_irqsave(&zone->lock, flags);
7495 for (order = 0; order < MAX_ORDER; order++) {
7496 struct page *page_head = page - (pfn & ((1 << order) - 1));
7498 if (PageBuddy(page_head) && page_order(page_head) >= order)
7501 spin_unlock_irqrestore(&zone->lock, flags);
7503 return order < MAX_ORDER;