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 <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/page_ext.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.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 int page_group_by_mobility_disabled __read_mostly;
287 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
288 static inline void reset_deferred_meminit(pg_data_t *pgdat)
290 pgdat->first_deferred_pfn = ULONG_MAX;
293 /* Returns true if the struct page for the pfn is uninitialised */
294 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
296 int nid = early_pfn_to_nid(pfn);
298 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
305 * Returns false when the remaining initialisation should be deferred until
306 * later in the boot cycle when it can be parallelised.
308 static inline bool update_defer_init(pg_data_t *pgdat,
309 unsigned long pfn, unsigned long zone_end,
310 unsigned long *nr_initialised)
312 unsigned long max_initialise;
314 /* Always populate low zones for address-contrained allocations */
315 if (zone_end < pgdat_end_pfn(pgdat))
318 * Initialise at least 2G of a node but also take into account that
319 * two large system hashes that can take up 1GB for 0.25TB/node.
321 max_initialise = max(2UL << (30 - PAGE_SHIFT),
322 (pgdat->node_spanned_pages >> 8));
325 if ((*nr_initialised > max_initialise) &&
326 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
327 pgdat->first_deferred_pfn = pfn;
334 static inline void reset_deferred_meminit(pg_data_t *pgdat)
338 static inline bool early_page_uninitialised(unsigned long pfn)
343 static inline bool update_defer_init(pg_data_t *pgdat,
344 unsigned long pfn, unsigned long zone_end,
345 unsigned long *nr_initialised)
351 /* Return a pointer to the bitmap storing bits affecting a block of pages */
352 static inline unsigned long *get_pageblock_bitmap(struct page *page,
355 #ifdef CONFIG_SPARSEMEM
356 return __pfn_to_section(pfn)->pageblock_flags;
358 return page_zone(page)->pageblock_flags;
359 #endif /* CONFIG_SPARSEMEM */
362 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
364 #ifdef CONFIG_SPARSEMEM
365 pfn &= (PAGES_PER_SECTION-1);
366 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
368 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
369 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
370 #endif /* CONFIG_SPARSEMEM */
374 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
375 * @page: The page within the block of interest
376 * @pfn: The target page frame number
377 * @end_bitidx: The last bit of interest to retrieve
378 * @mask: mask of bits that the caller is interested in
380 * Return: pageblock_bits flags
382 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
384 unsigned long end_bitidx,
387 unsigned long *bitmap;
388 unsigned long bitidx, word_bitidx;
391 bitmap = get_pageblock_bitmap(page, pfn);
392 bitidx = pfn_to_bitidx(page, pfn);
393 word_bitidx = bitidx / BITS_PER_LONG;
394 bitidx &= (BITS_PER_LONG-1);
396 word = bitmap[word_bitidx];
397 bitidx += end_bitidx;
398 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
401 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
402 unsigned long end_bitidx,
405 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
408 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
410 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
414 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
415 * @page: The page within the block of interest
416 * @flags: The flags to set
417 * @pfn: The target page frame number
418 * @end_bitidx: The last bit of interest
419 * @mask: mask of bits that the caller is interested in
421 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
423 unsigned long end_bitidx,
426 unsigned long *bitmap;
427 unsigned long bitidx, word_bitidx;
428 unsigned long old_word, word;
430 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
432 bitmap = get_pageblock_bitmap(page, pfn);
433 bitidx = pfn_to_bitidx(page, pfn);
434 word_bitidx = bitidx / BITS_PER_LONG;
435 bitidx &= (BITS_PER_LONG-1);
437 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
439 bitidx += end_bitidx;
440 mask <<= (BITS_PER_LONG - bitidx - 1);
441 flags <<= (BITS_PER_LONG - bitidx - 1);
443 word = READ_ONCE(bitmap[word_bitidx]);
445 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
446 if (word == old_word)
452 void set_pageblock_migratetype(struct page *page, int migratetype)
454 if (unlikely(page_group_by_mobility_disabled &&
455 migratetype < MIGRATE_PCPTYPES))
456 migratetype = MIGRATE_UNMOVABLE;
458 set_pageblock_flags_group(page, (unsigned long)migratetype,
459 PB_migrate, PB_migrate_end);
462 #ifdef CONFIG_DEBUG_VM
463 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
467 unsigned long pfn = page_to_pfn(page);
468 unsigned long sp, start_pfn;
471 seq = zone_span_seqbegin(zone);
472 start_pfn = zone->zone_start_pfn;
473 sp = zone->spanned_pages;
474 if (!zone_spans_pfn(zone, pfn))
476 } while (zone_span_seqretry(zone, seq));
479 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
480 pfn, zone_to_nid(zone), zone->name,
481 start_pfn, start_pfn + sp);
486 static int page_is_consistent(struct zone *zone, struct page *page)
488 if (!pfn_valid_within(page_to_pfn(page)))
490 if (zone != page_zone(page))
496 * Temporary debugging check for pages not lying within a given zone.
498 static int bad_range(struct zone *zone, struct page *page)
500 if (page_outside_zone_boundaries(zone, page))
502 if (!page_is_consistent(zone, page))
508 static inline int bad_range(struct zone *zone, struct page *page)
514 static void bad_page(struct page *page, const char *reason,
515 unsigned long bad_flags)
517 static unsigned long resume;
518 static unsigned long nr_shown;
519 static unsigned long nr_unshown;
522 * Allow a burst of 60 reports, then keep quiet for that minute;
523 * or allow a steady drip of one report per second.
525 if (nr_shown == 60) {
526 if (time_before(jiffies, resume)) {
532 "BUG: Bad page state: %lu messages suppressed\n",
539 resume = jiffies + 60 * HZ;
541 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
542 current->comm, page_to_pfn(page));
543 __dump_page(page, reason);
544 bad_flags &= page->flags;
546 pr_alert("bad because of flags: %#lx(%pGp)\n",
547 bad_flags, &bad_flags);
548 dump_page_owner(page);
553 /* Leave bad fields for debug, except PageBuddy could make trouble */
554 page_mapcount_reset(page); /* remove PageBuddy */
555 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
559 * Higher-order pages are called "compound pages". They are structured thusly:
561 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
563 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
564 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
566 * The first tail page's ->compound_dtor holds the offset in array of compound
567 * page destructors. See compound_page_dtors.
569 * The first tail page's ->compound_order holds the order of allocation.
570 * This usage means that zero-order pages may not be compound.
573 void free_compound_page(struct page *page)
575 __free_pages_ok(page, compound_order(page));
578 void prep_compound_page(struct page *page, unsigned int order)
581 int nr_pages = 1 << order;
583 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
584 set_compound_order(page, order);
586 for (i = 1; i < nr_pages; i++) {
587 struct page *p = page + i;
588 set_page_count(p, 0);
589 p->mapping = TAIL_MAPPING;
590 set_compound_head(p, page);
592 atomic_set(compound_mapcount_ptr(page), -1);
595 #ifdef CONFIG_DEBUG_PAGEALLOC
596 unsigned int _debug_guardpage_minorder;
597 bool _debug_pagealloc_enabled __read_mostly
598 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
599 EXPORT_SYMBOL(_debug_pagealloc_enabled);
600 bool _debug_guardpage_enabled __read_mostly;
602 static int __init early_debug_pagealloc(char *buf)
606 return kstrtobool(buf, &_debug_pagealloc_enabled);
608 early_param("debug_pagealloc", early_debug_pagealloc);
610 static bool need_debug_guardpage(void)
612 /* If we don't use debug_pagealloc, we don't need guard page */
613 if (!debug_pagealloc_enabled())
616 if (!debug_guardpage_minorder())
622 static void init_debug_guardpage(void)
624 if (!debug_pagealloc_enabled())
627 if (!debug_guardpage_minorder())
630 _debug_guardpage_enabled = true;
633 struct page_ext_operations debug_guardpage_ops = {
634 .need = need_debug_guardpage,
635 .init = init_debug_guardpage,
638 static int __init debug_guardpage_minorder_setup(char *buf)
642 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
643 pr_err("Bad debug_guardpage_minorder value\n");
646 _debug_guardpage_minorder = res;
647 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
650 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
652 static inline bool set_page_guard(struct zone *zone, struct page *page,
653 unsigned int order, int migratetype)
655 struct page_ext *page_ext;
657 if (!debug_guardpage_enabled())
660 if (order >= debug_guardpage_minorder())
663 page_ext = lookup_page_ext(page);
664 if (unlikely(!page_ext))
667 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
669 INIT_LIST_HEAD(&page->lru);
670 set_page_private(page, order);
671 /* Guard pages are not available for any usage */
672 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
677 static inline void clear_page_guard(struct zone *zone, struct page *page,
678 unsigned int order, int migratetype)
680 struct page_ext *page_ext;
682 if (!debug_guardpage_enabled())
685 page_ext = lookup_page_ext(page);
686 if (unlikely(!page_ext))
689 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
691 set_page_private(page, 0);
692 if (!is_migrate_isolate(migratetype))
693 __mod_zone_freepage_state(zone, (1 << order), migratetype);
696 struct page_ext_operations debug_guardpage_ops;
697 static inline bool set_page_guard(struct zone *zone, struct page *page,
698 unsigned int order, int migratetype) { return false; }
699 static inline void clear_page_guard(struct zone *zone, struct page *page,
700 unsigned int order, int migratetype) {}
703 static inline void set_page_order(struct page *page, unsigned int order)
705 set_page_private(page, order);
706 __SetPageBuddy(page);
709 static inline void rmv_page_order(struct page *page)
711 __ClearPageBuddy(page);
712 set_page_private(page, 0);
716 * This function checks whether a page is free && is the buddy
717 * we can do coalesce a page and its buddy if
718 * (a) the buddy is not in a hole (check before calling!) &&
719 * (b) the buddy is in the buddy system &&
720 * (c) a page and its buddy have the same order &&
721 * (d) a page and its buddy are in the same zone.
723 * For recording whether a page is in the buddy system, we set ->_mapcount
724 * PAGE_BUDDY_MAPCOUNT_VALUE.
725 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
726 * serialized by zone->lock.
728 * For recording page's order, we use page_private(page).
730 static inline int page_is_buddy(struct page *page, struct page *buddy,
733 if (page_is_guard(buddy) && page_order(buddy) == order) {
734 if (page_zone_id(page) != page_zone_id(buddy))
737 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
742 if (PageBuddy(buddy) && page_order(buddy) == order) {
744 * zone check is done late to avoid uselessly
745 * calculating zone/node ids for pages that could
748 if (page_zone_id(page) != page_zone_id(buddy))
751 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
759 * Freeing function for a buddy system allocator.
761 * The concept of a buddy system is to maintain direct-mapped table
762 * (containing bit values) for memory blocks of various "orders".
763 * The bottom level table contains the map for the smallest allocatable
764 * units of memory (here, pages), and each level above it describes
765 * pairs of units from the levels below, hence, "buddies".
766 * At a high level, all that happens here is marking the table entry
767 * at the bottom level available, and propagating the changes upward
768 * as necessary, plus some accounting needed to play nicely with other
769 * parts of the VM system.
770 * At each level, we keep a list of pages, which are heads of continuous
771 * free pages of length of (1 << order) and marked with _mapcount
772 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
774 * So when we are allocating or freeing one, we can derive the state of the
775 * other. That is, if we allocate a small block, and both were
776 * free, the remainder of the region must be split into blocks.
777 * If a block is freed, and its buddy is also free, then this
778 * triggers coalescing into a block of larger size.
783 static inline void __free_one_page(struct page *page,
785 struct zone *zone, unsigned int order,
788 unsigned long combined_pfn;
789 unsigned long uninitialized_var(buddy_pfn);
791 unsigned int max_order;
793 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
795 VM_BUG_ON(!zone_is_initialized(zone));
796 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
798 VM_BUG_ON(migratetype == -1);
799 if (likely(!is_migrate_isolate(migratetype)))
800 __mod_zone_freepage_state(zone, 1 << order, migratetype);
802 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
803 VM_BUG_ON_PAGE(bad_range(zone, page), page);
806 while (order < max_order - 1) {
807 buddy_pfn = __find_buddy_pfn(pfn, order);
808 buddy = page + (buddy_pfn - pfn);
810 if (!pfn_valid_within(buddy_pfn))
812 if (!page_is_buddy(page, buddy, order))
815 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
816 * merge with it and move up one order.
818 if (page_is_guard(buddy)) {
819 clear_page_guard(zone, buddy, order, migratetype);
821 list_del(&buddy->lru);
822 zone->free_area[order].nr_free--;
823 rmv_page_order(buddy);
825 combined_pfn = buddy_pfn & pfn;
826 page = page + (combined_pfn - pfn);
830 if (max_order < MAX_ORDER) {
831 /* If we are here, it means order is >= pageblock_order.
832 * We want to prevent merge between freepages on isolate
833 * pageblock and normal pageblock. Without this, pageblock
834 * isolation could cause incorrect freepage or CMA accounting.
836 * We don't want to hit this code for the more frequent
839 if (unlikely(has_isolate_pageblock(zone))) {
842 buddy_pfn = __find_buddy_pfn(pfn, order);
843 buddy = page + (buddy_pfn - pfn);
844 buddy_mt = get_pageblock_migratetype(buddy);
846 if (migratetype != buddy_mt
847 && (is_migrate_isolate(migratetype) ||
848 is_migrate_isolate(buddy_mt)))
852 goto continue_merging;
856 set_page_order(page, order);
859 * If this is not the largest possible page, check if the buddy
860 * of the next-highest order is free. If it is, it's possible
861 * that pages are being freed that will coalesce soon. In case,
862 * that is happening, add the free page to the tail of the list
863 * so it's less likely to be used soon and more likely to be merged
864 * as a higher order page
866 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
867 struct page *higher_page, *higher_buddy;
868 combined_pfn = buddy_pfn & pfn;
869 higher_page = page + (combined_pfn - pfn);
870 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
871 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
872 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
873 list_add_tail(&page->lru,
874 &zone->free_area[order].free_list[migratetype]);
879 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
881 zone->free_area[order].nr_free++;
885 * A bad page could be due to a number of fields. Instead of multiple branches,
886 * try and check multiple fields with one check. The caller must do a detailed
887 * check if necessary.
889 static inline bool page_expected_state(struct page *page,
890 unsigned long check_flags)
892 if (unlikely(atomic_read(&page->_mapcount) != -1))
895 if (unlikely((unsigned long)page->mapping |
896 page_ref_count(page) |
898 (unsigned long)page->mem_cgroup |
900 (page->flags & check_flags)))
906 static void free_pages_check_bad(struct page *page)
908 const char *bad_reason;
909 unsigned long bad_flags;
914 if (unlikely(atomic_read(&page->_mapcount) != -1))
915 bad_reason = "nonzero mapcount";
916 if (unlikely(page->mapping != NULL))
917 bad_reason = "non-NULL mapping";
918 if (unlikely(page_ref_count(page) != 0))
919 bad_reason = "nonzero _refcount";
920 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
921 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
922 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
925 if (unlikely(page->mem_cgroup))
926 bad_reason = "page still charged to cgroup";
928 bad_page(page, bad_reason, bad_flags);
931 static inline int free_pages_check(struct page *page)
933 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
936 /* Something has gone sideways, find it */
937 free_pages_check_bad(page);
941 static int free_tail_pages_check(struct page *head_page, struct page *page)
946 * We rely page->lru.next never has bit 0 set, unless the page
947 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
949 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
951 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
955 switch (page - head_page) {
957 /* the first tail page: ->mapping is compound_mapcount() */
958 if (unlikely(compound_mapcount(page))) {
959 bad_page(page, "nonzero compound_mapcount", 0);
965 * the second tail page: ->mapping is
966 * page_deferred_list().next -- ignore value.
970 if (page->mapping != TAIL_MAPPING) {
971 bad_page(page, "corrupted mapping in tail page", 0);
976 if (unlikely(!PageTail(page))) {
977 bad_page(page, "PageTail not set", 0);
980 if (unlikely(compound_head(page) != head_page)) {
981 bad_page(page, "compound_head not consistent", 0);
986 page->mapping = NULL;
987 clear_compound_head(page);
991 static __always_inline bool free_pages_prepare(struct page *page,
992 unsigned int order, bool check_free)
996 VM_BUG_ON_PAGE(PageTail(page), page);
998 trace_mm_page_free(page, order);
999 kmemcheck_free_shadow(page, order);
1002 * Check tail pages before head page information is cleared to
1003 * avoid checking PageCompound for order-0 pages.
1005 if (unlikely(order)) {
1006 bool compound = PageCompound(page);
1009 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1012 ClearPageDoubleMap(page);
1013 for (i = 1; i < (1 << order); i++) {
1015 bad += free_tail_pages_check(page, page + i);
1016 if (unlikely(free_pages_check(page + i))) {
1020 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 if (PageMappingFlags(page))
1024 page->mapping = NULL;
1025 if (memcg_kmem_enabled() && PageKmemcg(page))
1026 memcg_kmem_uncharge(page, order);
1028 bad += free_pages_check(page);
1032 page_cpupid_reset_last(page);
1033 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1034 reset_page_owner(page, order);
1036 if (!PageHighMem(page)) {
1037 debug_check_no_locks_freed(page_address(page),
1038 PAGE_SIZE << order);
1039 debug_check_no_obj_freed(page_address(page),
1040 PAGE_SIZE << order);
1042 arch_free_page(page, order);
1043 kernel_poison_pages(page, 1 << order, 0);
1044 kernel_map_pages(page, 1 << order, 0);
1045 kasan_free_pages(page, order);
1050 #ifdef CONFIG_DEBUG_VM
1051 static inline bool free_pcp_prepare(struct page *page)
1053 return free_pages_prepare(page, 0, true);
1056 static inline bool bulkfree_pcp_prepare(struct page *page)
1061 static bool free_pcp_prepare(struct page *page)
1063 return free_pages_prepare(page, 0, false);
1066 static bool bulkfree_pcp_prepare(struct page *page)
1068 return free_pages_check(page);
1070 #endif /* CONFIG_DEBUG_VM */
1073 * Frees a number of pages from the PCP lists
1074 * Assumes all pages on list are in same zone, and of same order.
1075 * count is the number of pages to free.
1077 * If the zone was previously in an "all pages pinned" state then look to
1078 * see if this freeing clears that state.
1080 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1081 * pinned" detection logic.
1083 static void free_pcppages_bulk(struct zone *zone, int count,
1084 struct per_cpu_pages *pcp)
1086 int migratetype = 0;
1088 unsigned long nr_scanned;
1089 bool isolated_pageblocks;
1091 spin_lock(&zone->lock);
1092 isolated_pageblocks = has_isolate_pageblock(zone);
1093 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1095 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1099 struct list_head *list;
1102 * Remove pages from lists in a round-robin fashion. A
1103 * batch_free count is maintained that is incremented when an
1104 * empty list is encountered. This is so more pages are freed
1105 * off fuller lists instead of spinning excessively around empty
1110 if (++migratetype == MIGRATE_PCPTYPES)
1112 list = &pcp->lists[migratetype];
1113 } while (list_empty(list));
1115 /* This is the only non-empty list. Free them all. */
1116 if (batch_free == MIGRATE_PCPTYPES)
1120 int mt; /* migratetype of the to-be-freed page */
1122 page = list_last_entry(list, struct page, lru);
1123 /* must delete as __free_one_page list manipulates */
1124 list_del(&page->lru);
1126 mt = get_pcppage_migratetype(page);
1127 /* MIGRATE_ISOLATE page should not go to pcplists */
1128 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1129 /* Pageblock could have been isolated meanwhile */
1130 if (unlikely(isolated_pageblocks))
1131 mt = get_pageblock_migratetype(page);
1133 if (bulkfree_pcp_prepare(page))
1136 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1137 trace_mm_page_pcpu_drain(page, 0, mt);
1138 } while (--count && --batch_free && !list_empty(list));
1140 spin_unlock(&zone->lock);
1143 static void free_one_page(struct zone *zone,
1144 struct page *page, unsigned long pfn,
1148 unsigned long nr_scanned;
1149 spin_lock(&zone->lock);
1150 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1152 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1154 if (unlikely(has_isolate_pageblock(zone) ||
1155 is_migrate_isolate(migratetype))) {
1156 migratetype = get_pfnblock_migratetype(page, pfn);
1158 __free_one_page(page, pfn, zone, order, migratetype);
1159 spin_unlock(&zone->lock);
1162 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1163 unsigned long zone, int nid)
1165 set_page_links(page, zone, nid, pfn);
1166 init_page_count(page);
1167 page_mapcount_reset(page);
1168 page_cpupid_reset_last(page);
1170 INIT_LIST_HEAD(&page->lru);
1171 #ifdef WANT_PAGE_VIRTUAL
1172 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1173 if (!is_highmem_idx(zone))
1174 set_page_address(page, __va(pfn << PAGE_SHIFT));
1178 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1181 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1184 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1185 static void init_reserved_page(unsigned long pfn)
1190 if (!early_page_uninitialised(pfn))
1193 nid = early_pfn_to_nid(pfn);
1194 pgdat = NODE_DATA(nid);
1196 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1197 struct zone *zone = &pgdat->node_zones[zid];
1199 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1202 __init_single_pfn(pfn, zid, nid);
1205 static inline void init_reserved_page(unsigned long pfn)
1208 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211 * Initialised pages do not have PageReserved set. This function is
1212 * called for each range allocated by the bootmem allocator and
1213 * marks the pages PageReserved. The remaining valid pages are later
1214 * sent to the buddy page allocator.
1216 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1218 unsigned long start_pfn = PFN_DOWN(start);
1219 unsigned long end_pfn = PFN_UP(end);
1221 for (; start_pfn < end_pfn; start_pfn++) {
1222 if (pfn_valid(start_pfn)) {
1223 struct page *page = pfn_to_page(start_pfn);
1225 init_reserved_page(start_pfn);
1227 /* Avoid false-positive PageTail() */
1228 INIT_LIST_HEAD(&page->lru);
1230 SetPageReserved(page);
1235 static void __free_pages_ok(struct page *page, unsigned int order)
1237 unsigned long flags;
1239 unsigned long pfn = page_to_pfn(page);
1241 if (!free_pages_prepare(page, order, true))
1244 migratetype = get_pfnblock_migratetype(page, pfn);
1245 local_irq_save(flags);
1246 __count_vm_events(PGFREE, 1 << order);
1247 free_one_page(page_zone(page), page, pfn, order, migratetype);
1248 local_irq_restore(flags);
1251 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1253 unsigned int nr_pages = 1 << order;
1254 struct page *p = page;
1258 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1260 __ClearPageReserved(p);
1261 set_page_count(p, 0);
1263 __ClearPageReserved(p);
1264 set_page_count(p, 0);
1266 page_zone(page)->managed_pages += nr_pages;
1267 set_page_refcounted(page);
1268 __free_pages(page, order);
1271 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1272 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1274 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1276 int __meminit early_pfn_to_nid(unsigned long pfn)
1278 static DEFINE_SPINLOCK(early_pfn_lock);
1281 spin_lock(&early_pfn_lock);
1282 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1284 nid = first_online_node;
1285 spin_unlock(&early_pfn_lock);
1291 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1292 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1293 struct mminit_pfnnid_cache *state)
1297 nid = __early_pfn_to_nid(pfn, state);
1298 if (nid >= 0 && nid != node)
1303 /* Only safe to use early in boot when initialisation is single-threaded */
1304 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1306 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1315 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1316 struct mminit_pfnnid_cache *state)
1323 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1326 if (early_page_uninitialised(pfn))
1328 return __free_pages_boot_core(page, order);
1332 * Check that the whole (or subset of) a pageblock given by the interval of
1333 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1334 * with the migration of free compaction scanner. The scanners then need to
1335 * use only pfn_valid_within() check for arches that allow holes within
1338 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1340 * It's possible on some configurations to have a setup like node0 node1 node0
1341 * i.e. it's possible that all pages within a zones range of pages do not
1342 * belong to a single zone. We assume that a border between node0 and node1
1343 * can occur within a single pageblock, but not a node0 node1 node0
1344 * interleaving within a single pageblock. It is therefore sufficient to check
1345 * the first and last page of a pageblock and avoid checking each individual
1346 * page in a pageblock.
1348 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1349 unsigned long end_pfn, struct zone *zone)
1351 struct page *start_page;
1352 struct page *end_page;
1354 /* end_pfn is one past the range we are checking */
1357 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1360 start_page = pfn_to_page(start_pfn);
1362 if (page_zone(start_page) != zone)
1365 end_page = pfn_to_page(end_pfn);
1367 /* This gives a shorter code than deriving page_zone(end_page) */
1368 if (page_zone_id(start_page) != page_zone_id(end_page))
1374 void set_zone_contiguous(struct zone *zone)
1376 unsigned long block_start_pfn = zone->zone_start_pfn;
1377 unsigned long block_end_pfn;
1379 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1380 for (; block_start_pfn < zone_end_pfn(zone);
1381 block_start_pfn = block_end_pfn,
1382 block_end_pfn += pageblock_nr_pages) {
1384 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1386 if (!__pageblock_pfn_to_page(block_start_pfn,
1387 block_end_pfn, zone))
1391 /* We confirm that there is no hole */
1392 zone->contiguous = true;
1395 void clear_zone_contiguous(struct zone *zone)
1397 zone->contiguous = false;
1400 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1401 static void __init deferred_free_range(struct page *page,
1402 unsigned long pfn, int nr_pages)
1409 /* Free a large naturally-aligned chunk if possible */
1410 if (nr_pages == pageblock_nr_pages &&
1411 (pfn & (pageblock_nr_pages - 1)) == 0) {
1412 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1413 __free_pages_boot_core(page, pageblock_order);
1417 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1418 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1419 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1420 __free_pages_boot_core(page, 0);
1424 /* Completion tracking for deferred_init_memmap() threads */
1425 static atomic_t pgdat_init_n_undone __initdata;
1426 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1428 static inline void __init pgdat_init_report_one_done(void)
1430 if (atomic_dec_and_test(&pgdat_init_n_undone))
1431 complete(&pgdat_init_all_done_comp);
1434 /* Initialise remaining memory on a node */
1435 static int __init deferred_init_memmap(void *data)
1437 pg_data_t *pgdat = data;
1438 int nid = pgdat->node_id;
1439 struct mminit_pfnnid_cache nid_init_state = { };
1440 unsigned long start = jiffies;
1441 unsigned long nr_pages = 0;
1442 unsigned long walk_start, walk_end;
1445 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1446 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1448 if (first_init_pfn == ULONG_MAX) {
1449 pgdat_init_report_one_done();
1453 /* Bind memory initialisation thread to a local node if possible */
1454 if (!cpumask_empty(cpumask))
1455 set_cpus_allowed_ptr(current, cpumask);
1457 /* Sanity check boundaries */
1458 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1459 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1460 pgdat->first_deferred_pfn = ULONG_MAX;
1462 /* Only the highest zone is deferred so find it */
1463 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1464 zone = pgdat->node_zones + zid;
1465 if (first_init_pfn < zone_end_pfn(zone))
1469 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1470 unsigned long pfn, end_pfn;
1471 struct page *page = NULL;
1472 struct page *free_base_page = NULL;
1473 unsigned long free_base_pfn = 0;
1476 end_pfn = min(walk_end, zone_end_pfn(zone));
1477 pfn = first_init_pfn;
1478 if (pfn < walk_start)
1480 if (pfn < zone->zone_start_pfn)
1481 pfn = zone->zone_start_pfn;
1483 for (; pfn < end_pfn; pfn++) {
1484 if (!pfn_valid_within(pfn))
1488 * Ensure pfn_valid is checked every
1489 * pageblock_nr_pages for memory holes
1491 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1492 if (!pfn_valid(pfn)) {
1498 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1503 /* Minimise pfn page lookups and scheduler checks */
1504 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1507 nr_pages += nr_to_free;
1508 deferred_free_range(free_base_page,
1509 free_base_pfn, nr_to_free);
1510 free_base_page = NULL;
1511 free_base_pfn = nr_to_free = 0;
1513 page = pfn_to_page(pfn);
1518 VM_BUG_ON(page_zone(page) != zone);
1522 __init_single_page(page, pfn, zid, nid);
1523 if (!free_base_page) {
1524 free_base_page = page;
1525 free_base_pfn = pfn;
1530 /* Where possible, batch up pages for a single free */
1533 /* Free the current block of pages to allocator */
1534 nr_pages += nr_to_free;
1535 deferred_free_range(free_base_page, free_base_pfn,
1537 free_base_page = NULL;
1538 free_base_pfn = nr_to_free = 0;
1540 /* Free the last block of pages to allocator */
1541 nr_pages += nr_to_free;
1542 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1544 first_init_pfn = max(end_pfn, first_init_pfn);
1547 /* Sanity check that the next zone really is unpopulated */
1548 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1550 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1551 jiffies_to_msecs(jiffies - start));
1553 pgdat_init_report_one_done();
1556 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1558 void __init page_alloc_init_late(void)
1562 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1565 /* There will be num_node_state(N_MEMORY) threads */
1566 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1567 for_each_node_state(nid, N_MEMORY) {
1568 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1571 /* Block until all are initialised */
1572 wait_for_completion(&pgdat_init_all_done_comp);
1574 /* Reinit limits that are based on free pages after the kernel is up */
1575 files_maxfiles_init();
1578 for_each_populated_zone(zone)
1579 set_zone_contiguous(zone);
1583 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1584 void __init init_cma_reserved_pageblock(struct page *page)
1586 unsigned i = pageblock_nr_pages;
1587 struct page *p = page;
1590 __ClearPageReserved(p);
1591 set_page_count(p, 0);
1594 set_pageblock_migratetype(page, MIGRATE_CMA);
1596 if (pageblock_order >= MAX_ORDER) {
1597 i = pageblock_nr_pages;
1600 set_page_refcounted(p);
1601 __free_pages(p, MAX_ORDER - 1);
1602 p += MAX_ORDER_NR_PAGES;
1603 } while (i -= MAX_ORDER_NR_PAGES);
1605 set_page_refcounted(page);
1606 __free_pages(page, pageblock_order);
1609 adjust_managed_page_count(page, pageblock_nr_pages);
1614 * The order of subdivision here is critical for the IO subsystem.
1615 * Please do not alter this order without good reasons and regression
1616 * testing. Specifically, as large blocks of memory are subdivided,
1617 * the order in which smaller blocks are delivered depends on the order
1618 * they're subdivided in this function. This is the primary factor
1619 * influencing the order in which pages are delivered to the IO
1620 * subsystem according to empirical testing, and this is also justified
1621 * by considering the behavior of a buddy system containing a single
1622 * large block of memory acted on by a series of small allocations.
1623 * This behavior is a critical factor in sglist merging's success.
1627 static inline void expand(struct zone *zone, struct page *page,
1628 int low, int high, struct free_area *area,
1631 unsigned long size = 1 << high;
1633 while (high > low) {
1637 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1640 * Mark as guard pages (or page), that will allow to
1641 * merge back to allocator when buddy will be freed.
1642 * Corresponding page table entries will not be touched,
1643 * pages will stay not present in virtual address space
1645 if (set_page_guard(zone, &page[size], high, migratetype))
1648 list_add(&page[size].lru, &area->free_list[migratetype]);
1650 set_page_order(&page[size], high);
1654 static void check_new_page_bad(struct page *page)
1656 const char *bad_reason = NULL;
1657 unsigned long bad_flags = 0;
1659 if (unlikely(atomic_read(&page->_mapcount) != -1))
1660 bad_reason = "nonzero mapcount";
1661 if (unlikely(page->mapping != NULL))
1662 bad_reason = "non-NULL mapping";
1663 if (unlikely(page_ref_count(page) != 0))
1664 bad_reason = "nonzero _count";
1665 if (unlikely(page->flags & __PG_HWPOISON)) {
1666 bad_reason = "HWPoisoned (hardware-corrupted)";
1667 bad_flags = __PG_HWPOISON;
1668 /* Don't complain about hwpoisoned pages */
1669 page_mapcount_reset(page); /* remove PageBuddy */
1672 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1673 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1674 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1677 if (unlikely(page->mem_cgroup))
1678 bad_reason = "page still charged to cgroup";
1680 bad_page(page, bad_reason, bad_flags);
1684 * This page is about to be returned from the page allocator
1686 static inline int check_new_page(struct page *page)
1688 if (likely(page_expected_state(page,
1689 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1692 check_new_page_bad(page);
1696 static inline bool free_pages_prezeroed(bool poisoned)
1698 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1699 page_poisoning_enabled() && poisoned;
1702 #ifdef CONFIG_DEBUG_VM
1703 static bool check_pcp_refill(struct page *page)
1708 static bool check_new_pcp(struct page *page)
1710 return check_new_page(page);
1713 static bool check_pcp_refill(struct page *page)
1715 return check_new_page(page);
1717 static bool check_new_pcp(struct page *page)
1721 #endif /* CONFIG_DEBUG_VM */
1723 static bool check_new_pages(struct page *page, unsigned int order)
1726 for (i = 0; i < (1 << order); i++) {
1727 struct page *p = page + i;
1729 if (unlikely(check_new_page(p)))
1736 inline void post_alloc_hook(struct page *page, unsigned int order,
1739 set_page_private(page, 0);
1740 set_page_refcounted(page);
1742 arch_alloc_page(page, order);
1743 kernel_map_pages(page, 1 << order, 1);
1744 kernel_poison_pages(page, 1 << order, 1);
1745 kasan_alloc_pages(page, order);
1746 set_page_owner(page, order, gfp_flags);
1749 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1750 unsigned int alloc_flags)
1753 bool poisoned = true;
1755 for (i = 0; i < (1 << order); i++) {
1756 struct page *p = page + i;
1758 poisoned &= page_is_poisoned(p);
1761 post_alloc_hook(page, order, gfp_flags);
1763 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1764 for (i = 0; i < (1 << order); i++)
1765 clear_highpage(page + i);
1767 if (order && (gfp_flags & __GFP_COMP))
1768 prep_compound_page(page, order);
1771 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1772 * allocate the page. The expectation is that the caller is taking
1773 * steps that will free more memory. The caller should avoid the page
1774 * being used for !PFMEMALLOC purposes.
1776 if (alloc_flags & ALLOC_NO_WATERMARKS)
1777 set_page_pfmemalloc(page);
1779 clear_page_pfmemalloc(page);
1783 * Go through the free lists for the given migratetype and remove
1784 * the smallest available page from the freelists
1787 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1790 unsigned int current_order;
1791 struct free_area *area;
1794 /* Find a page of the appropriate size in the preferred list */
1795 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1796 area = &(zone->free_area[current_order]);
1797 page = list_first_entry_or_null(&area->free_list[migratetype],
1801 list_del(&page->lru);
1802 rmv_page_order(page);
1804 expand(zone, page, order, current_order, area, migratetype);
1805 set_pcppage_migratetype(page, migratetype);
1814 * This array describes the order lists are fallen back to when
1815 * the free lists for the desirable migrate type are depleted
1817 static int fallbacks[MIGRATE_TYPES][4] = {
1818 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1819 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1820 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1822 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1824 #ifdef CONFIG_MEMORY_ISOLATION
1825 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1830 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1833 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1836 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1837 unsigned int order) { return NULL; }
1841 * Move the free pages in a range to the free lists of the requested type.
1842 * Note that start_page and end_pages are not aligned on a pageblock
1843 * boundary. If alignment is required, use move_freepages_block()
1845 int move_freepages(struct zone *zone,
1846 struct page *start_page, struct page *end_page,
1851 int pages_moved = 0;
1853 #ifndef CONFIG_HOLES_IN_ZONE
1855 * page_zone is not safe to call in this context when
1856 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1857 * anyway as we check zone boundaries in move_freepages_block().
1858 * Remove at a later date when no bug reports exist related to
1859 * grouping pages by mobility
1861 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1864 for (page = start_page; page <= end_page;) {
1865 if (!pfn_valid_within(page_to_pfn(page))) {
1870 /* Make sure we are not inadvertently changing nodes */
1871 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1873 if (!PageBuddy(page)) {
1878 order = page_order(page);
1879 list_move(&page->lru,
1880 &zone->free_area[order].free_list[migratetype]);
1882 pages_moved += 1 << order;
1888 int move_freepages_block(struct zone *zone, struct page *page,
1891 unsigned long start_pfn, end_pfn;
1892 struct page *start_page, *end_page;
1894 start_pfn = page_to_pfn(page);
1895 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1896 start_page = pfn_to_page(start_pfn);
1897 end_page = start_page + pageblock_nr_pages - 1;
1898 end_pfn = start_pfn + pageblock_nr_pages - 1;
1900 /* Do not cross zone boundaries */
1901 if (!zone_spans_pfn(zone, start_pfn))
1903 if (!zone_spans_pfn(zone, end_pfn))
1906 return move_freepages(zone, start_page, end_page, migratetype);
1909 static void change_pageblock_range(struct page *pageblock_page,
1910 int start_order, int migratetype)
1912 int nr_pageblocks = 1 << (start_order - pageblock_order);
1914 while (nr_pageblocks--) {
1915 set_pageblock_migratetype(pageblock_page, migratetype);
1916 pageblock_page += pageblock_nr_pages;
1921 * When we are falling back to another migratetype during allocation, try to
1922 * steal extra free pages from the same pageblocks to satisfy further
1923 * allocations, instead of polluting multiple pageblocks.
1925 * If we are stealing a relatively large buddy page, it is likely there will
1926 * be more free pages in the pageblock, so try to steal them all. For
1927 * reclaimable and unmovable allocations, we steal regardless of page size,
1928 * as fragmentation caused by those allocations polluting movable pageblocks
1929 * is worse than movable allocations stealing from unmovable and reclaimable
1932 static bool can_steal_fallback(unsigned int order, int start_mt)
1935 * Leaving this order check is intended, although there is
1936 * relaxed order check in next check. The reason is that
1937 * we can actually steal whole pageblock if this condition met,
1938 * but, below check doesn't guarantee it and that is just heuristic
1939 * so could be changed anytime.
1941 if (order >= pageblock_order)
1944 if (order >= pageblock_order / 2 ||
1945 start_mt == MIGRATE_RECLAIMABLE ||
1946 start_mt == MIGRATE_UNMOVABLE ||
1947 page_group_by_mobility_disabled)
1954 * This function implements actual steal behaviour. If order is large enough,
1955 * we can steal whole pageblock. If not, we first move freepages in this
1956 * pageblock and check whether half of pages are moved or not. If half of
1957 * pages are moved, we can change migratetype of pageblock and permanently
1958 * use it's pages as requested migratetype in the future.
1960 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1963 unsigned int current_order = page_order(page);
1966 /* Take ownership for orders >= pageblock_order */
1967 if (current_order >= pageblock_order) {
1968 change_pageblock_range(page, current_order, start_type);
1972 pages = move_freepages_block(zone, page, start_type);
1974 /* Claim the whole block if over half of it is free */
1975 if (pages >= (1 << (pageblock_order-1)) ||
1976 page_group_by_mobility_disabled)
1977 set_pageblock_migratetype(page, start_type);
1981 * Check whether there is a suitable fallback freepage with requested order.
1982 * If only_stealable is true, this function returns fallback_mt only if
1983 * we can steal other freepages all together. This would help to reduce
1984 * fragmentation due to mixed migratetype pages in one pageblock.
1986 int find_suitable_fallback(struct free_area *area, unsigned int order,
1987 int migratetype, bool only_stealable, bool *can_steal)
1992 if (area->nr_free == 0)
1997 fallback_mt = fallbacks[migratetype][i];
1998 if (fallback_mt == MIGRATE_TYPES)
2001 if (list_empty(&area->free_list[fallback_mt]))
2004 if (can_steal_fallback(order, migratetype))
2007 if (!only_stealable)
2018 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2019 * there are no empty page blocks that contain a page with a suitable order
2021 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2022 unsigned int alloc_order)
2025 unsigned long max_managed, flags;
2028 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2029 * Check is race-prone but harmless.
2031 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2032 if (zone->nr_reserved_highatomic >= max_managed)
2035 spin_lock_irqsave(&zone->lock, flags);
2037 /* Recheck the nr_reserved_highatomic limit under the lock */
2038 if (zone->nr_reserved_highatomic >= max_managed)
2042 mt = get_pageblock_migratetype(page);
2043 if (mt != MIGRATE_HIGHATOMIC &&
2044 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2045 zone->nr_reserved_highatomic += pageblock_nr_pages;
2046 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2047 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2051 spin_unlock_irqrestore(&zone->lock, flags);
2055 * Used when an allocation is about to fail under memory pressure. This
2056 * potentially hurts the reliability of high-order allocations when under
2057 * intense memory pressure but failed atomic allocations should be easier
2058 * to recover from than an OOM.
2060 * If @force is true, try to unreserve a pageblock even though highatomic
2061 * pageblock is exhausted.
2063 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2066 struct zonelist *zonelist = ac->zonelist;
2067 unsigned long flags;
2074 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2077 * Preserve at least one pageblock unless memory pressure
2080 if (!force && zone->nr_reserved_highatomic <=
2084 spin_lock_irqsave(&zone->lock, flags);
2085 for (order = 0; order < MAX_ORDER; order++) {
2086 struct free_area *area = &(zone->free_area[order]);
2088 page = list_first_entry_or_null(
2089 &area->free_list[MIGRATE_HIGHATOMIC],
2095 * In page freeing path, migratetype change is racy so
2096 * we can counter several free pages in a pageblock
2097 * in this loop althoug we changed the pageblock type
2098 * from highatomic to ac->migratetype. So we should
2099 * adjust the count once.
2101 if (get_pageblock_migratetype(page) ==
2102 MIGRATE_HIGHATOMIC) {
2104 * It should never happen but changes to
2105 * locking could inadvertently allow a per-cpu
2106 * drain to add pages to MIGRATE_HIGHATOMIC
2107 * while unreserving so be safe and watch for
2110 zone->nr_reserved_highatomic -= min(
2112 zone->nr_reserved_highatomic);
2116 * Convert to ac->migratetype and avoid the normal
2117 * pageblock stealing heuristics. Minimally, the caller
2118 * is doing the work and needs the pages. More
2119 * importantly, if the block was always converted to
2120 * MIGRATE_UNMOVABLE or another type then the number
2121 * of pageblocks that cannot be completely freed
2124 set_pageblock_migratetype(page, ac->migratetype);
2125 ret = move_freepages_block(zone, page, ac->migratetype);
2127 spin_unlock_irqrestore(&zone->lock, flags);
2131 spin_unlock_irqrestore(&zone->lock, flags);
2137 /* Remove an element from the buddy allocator from the fallback list */
2138 static inline struct page *
2139 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2141 struct free_area *area;
2142 unsigned int current_order;
2147 /* Find the largest possible block of pages in the other list */
2148 for (current_order = MAX_ORDER-1;
2149 current_order >= order && current_order <= MAX_ORDER-1;
2151 area = &(zone->free_area[current_order]);
2152 fallback_mt = find_suitable_fallback(area, current_order,
2153 start_migratetype, false, &can_steal);
2154 if (fallback_mt == -1)
2157 page = list_first_entry(&area->free_list[fallback_mt],
2160 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC)
2161 steal_suitable_fallback(zone, page, start_migratetype);
2163 /* Remove the page from the freelists */
2165 list_del(&page->lru);
2166 rmv_page_order(page);
2168 expand(zone, page, order, current_order, area,
2171 * The pcppage_migratetype may differ from pageblock's
2172 * migratetype depending on the decisions in
2173 * find_suitable_fallback(). This is OK as long as it does not
2174 * differ for MIGRATE_CMA pageblocks. Those can be used as
2175 * fallback only via special __rmqueue_cma_fallback() function
2177 set_pcppage_migratetype(page, start_migratetype);
2179 trace_mm_page_alloc_extfrag(page, order, current_order,
2180 start_migratetype, fallback_mt);
2189 * Do the hard work of removing an element from the buddy allocator.
2190 * Call me with the zone->lock already held.
2192 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2197 page = __rmqueue_smallest(zone, order, migratetype);
2198 if (unlikely(!page)) {
2199 if (migratetype == MIGRATE_MOVABLE)
2200 page = __rmqueue_cma_fallback(zone, order);
2203 page = __rmqueue_fallback(zone, order, migratetype);
2206 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2211 * Obtain a specified number of elements from the buddy allocator, all under
2212 * a single hold of the lock, for efficiency. Add them to the supplied list.
2213 * Returns the number of new pages which were placed at *list.
2215 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2216 unsigned long count, struct list_head *list,
2217 int migratetype, bool cold)
2221 spin_lock(&zone->lock);
2222 for (i = 0; i < count; ++i) {
2223 struct page *page = __rmqueue(zone, order, migratetype);
2224 if (unlikely(page == NULL))
2227 if (unlikely(check_pcp_refill(page)))
2231 * Split buddy pages returned by expand() are received here
2232 * in physical page order. The page is added to the callers and
2233 * list and the list head then moves forward. From the callers
2234 * perspective, the linked list is ordered by page number in
2235 * some conditions. This is useful for IO devices that can
2236 * merge IO requests if the physical pages are ordered
2240 list_add(&page->lru, list);
2242 list_add_tail(&page->lru, list);
2245 if (is_migrate_cma(get_pcppage_migratetype(page)))
2246 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2251 * i pages were removed from the buddy list even if some leak due
2252 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2253 * on i. Do not confuse with 'alloced' which is the number of
2254 * pages added to the pcp list.
2256 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2257 spin_unlock(&zone->lock);
2263 * Called from the vmstat counter updater to drain pagesets of this
2264 * currently executing processor on remote nodes after they have
2267 * Note that this function must be called with the thread pinned to
2268 * a single processor.
2270 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2272 unsigned long flags;
2273 int to_drain, batch;
2275 local_irq_save(flags);
2276 batch = READ_ONCE(pcp->batch);
2277 to_drain = min(pcp->count, batch);
2279 free_pcppages_bulk(zone, to_drain, pcp);
2280 pcp->count -= to_drain;
2282 local_irq_restore(flags);
2287 * Drain pcplists of the indicated processor and zone.
2289 * The processor must either be the current processor and the
2290 * thread pinned to the current processor or a processor that
2293 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2295 unsigned long flags;
2296 struct per_cpu_pageset *pset;
2297 struct per_cpu_pages *pcp;
2299 local_irq_save(flags);
2300 pset = per_cpu_ptr(zone->pageset, cpu);
2304 free_pcppages_bulk(zone, pcp->count, pcp);
2307 local_irq_restore(flags);
2311 * Drain pcplists of all zones on the indicated processor.
2313 * The processor must either be the current processor and the
2314 * thread pinned to the current processor or a processor that
2317 static void drain_pages(unsigned int cpu)
2321 for_each_populated_zone(zone) {
2322 drain_pages_zone(cpu, zone);
2327 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2329 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2330 * the single zone's pages.
2332 void drain_local_pages(struct zone *zone)
2334 int cpu = smp_processor_id();
2337 drain_pages_zone(cpu, zone);
2343 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2345 * When zone parameter is non-NULL, spill just the single zone's pages.
2347 * Note that this code is protected against sending an IPI to an offline
2348 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2349 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2350 * nothing keeps CPUs from showing up after we populated the cpumask and
2351 * before the call to on_each_cpu_mask().
2353 void drain_all_pages(struct zone *zone)
2358 * Allocate in the BSS so we wont require allocation in
2359 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2361 static cpumask_t cpus_with_pcps;
2364 * We don't care about racing with CPU hotplug event
2365 * as offline notification will cause the notified
2366 * cpu to drain that CPU pcps and on_each_cpu_mask
2367 * disables preemption as part of its processing
2369 for_each_online_cpu(cpu) {
2370 struct per_cpu_pageset *pcp;
2372 bool has_pcps = false;
2375 pcp = per_cpu_ptr(zone->pageset, cpu);
2379 for_each_populated_zone(z) {
2380 pcp = per_cpu_ptr(z->pageset, cpu);
2381 if (pcp->pcp.count) {
2389 cpumask_set_cpu(cpu, &cpus_with_pcps);
2391 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2393 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2397 #ifdef CONFIG_HIBERNATION
2399 void mark_free_pages(struct zone *zone)
2401 unsigned long pfn, max_zone_pfn;
2402 unsigned long flags;
2403 unsigned int order, t;
2406 if (zone_is_empty(zone))
2409 spin_lock_irqsave(&zone->lock, flags);
2411 max_zone_pfn = zone_end_pfn(zone);
2412 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2413 if (pfn_valid(pfn)) {
2414 page = pfn_to_page(pfn);
2416 if (page_zone(page) != zone)
2419 if (!swsusp_page_is_forbidden(page))
2420 swsusp_unset_page_free(page);
2423 for_each_migratetype_order(order, t) {
2424 list_for_each_entry(page,
2425 &zone->free_area[order].free_list[t], lru) {
2428 pfn = page_to_pfn(page);
2429 for (i = 0; i < (1UL << order); i++)
2430 swsusp_set_page_free(pfn_to_page(pfn + i));
2433 spin_unlock_irqrestore(&zone->lock, flags);
2435 #endif /* CONFIG_PM */
2438 * Free a 0-order page
2439 * cold == true ? free a cold page : free a hot page
2441 void free_hot_cold_page(struct page *page, bool cold)
2443 struct zone *zone = page_zone(page);
2444 struct per_cpu_pages *pcp;
2445 unsigned long flags;
2446 unsigned long pfn = page_to_pfn(page);
2449 if (!free_pcp_prepare(page))
2452 migratetype = get_pfnblock_migratetype(page, pfn);
2453 set_pcppage_migratetype(page, migratetype);
2454 local_irq_save(flags);
2455 __count_vm_event(PGFREE);
2458 * We only track unmovable, reclaimable and movable on pcp lists.
2459 * Free ISOLATE pages back to the allocator because they are being
2460 * offlined but treat RESERVE as movable pages so we can get those
2461 * areas back if necessary. Otherwise, we may have to free
2462 * excessively into the page allocator
2464 if (migratetype >= MIGRATE_PCPTYPES) {
2465 if (unlikely(is_migrate_isolate(migratetype))) {
2466 free_one_page(zone, page, pfn, 0, migratetype);
2469 migratetype = MIGRATE_MOVABLE;
2472 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2474 list_add(&page->lru, &pcp->lists[migratetype]);
2476 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2478 if (pcp->count >= pcp->high) {
2479 unsigned long batch = READ_ONCE(pcp->batch);
2480 free_pcppages_bulk(zone, batch, pcp);
2481 pcp->count -= batch;
2485 local_irq_restore(flags);
2489 * Free a list of 0-order pages
2491 void free_hot_cold_page_list(struct list_head *list, bool cold)
2493 struct page *page, *next;
2495 list_for_each_entry_safe(page, next, list, lru) {
2496 trace_mm_page_free_batched(page, cold);
2497 free_hot_cold_page(page, cold);
2502 * split_page takes a non-compound higher-order page, and splits it into
2503 * n (1<<order) sub-pages: page[0..n]
2504 * Each sub-page must be freed individually.
2506 * Note: this is probably too low level an operation for use in drivers.
2507 * Please consult with lkml before using this in your driver.
2509 void split_page(struct page *page, unsigned int order)
2513 VM_BUG_ON_PAGE(PageCompound(page), page);
2514 VM_BUG_ON_PAGE(!page_count(page), page);
2516 #ifdef CONFIG_KMEMCHECK
2518 * Split shadow pages too, because free(page[0]) would
2519 * otherwise free the whole shadow.
2521 if (kmemcheck_page_is_tracked(page))
2522 split_page(virt_to_page(page[0].shadow), order);
2525 for (i = 1; i < (1 << order); i++)
2526 set_page_refcounted(page + i);
2527 split_page_owner(page, order);
2529 EXPORT_SYMBOL_GPL(split_page);
2531 int __isolate_free_page(struct page *page, unsigned int order)
2533 unsigned long watermark;
2537 BUG_ON(!PageBuddy(page));
2539 zone = page_zone(page);
2540 mt = get_pageblock_migratetype(page);
2542 if (!is_migrate_isolate(mt)) {
2544 * Obey watermarks as if the page was being allocated. We can
2545 * emulate a high-order watermark check with a raised order-0
2546 * watermark, because we already know our high-order page
2549 watermark = min_wmark_pages(zone) + (1UL << order);
2550 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2553 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2556 /* Remove page from free list */
2557 list_del(&page->lru);
2558 zone->free_area[order].nr_free--;
2559 rmv_page_order(page);
2562 * Set the pageblock if the isolated page is at least half of a
2565 if (order >= pageblock_order - 1) {
2566 struct page *endpage = page + (1 << order) - 1;
2567 for (; page < endpage; page += pageblock_nr_pages) {
2568 int mt = get_pageblock_migratetype(page);
2569 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2570 && mt != MIGRATE_HIGHATOMIC)
2571 set_pageblock_migratetype(page,
2577 return 1UL << order;
2581 * Update NUMA hit/miss statistics
2583 * Must be called with interrupts disabled.
2585 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2588 enum zone_stat_item local_stat = NUMA_LOCAL;
2590 if (z->node != numa_node_id())
2591 local_stat = NUMA_OTHER;
2593 if (z->node == preferred_zone->node)
2594 __inc_zone_state(z, NUMA_HIT);
2596 __inc_zone_state(z, NUMA_MISS);
2597 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2599 __inc_zone_state(z, local_stat);
2604 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2607 struct page *buffered_rmqueue(struct zone *preferred_zone,
2608 struct zone *zone, unsigned int order,
2609 gfp_t gfp_flags, unsigned int alloc_flags,
2612 unsigned long flags;
2614 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2616 if (likely(order == 0)) {
2617 struct per_cpu_pages *pcp;
2618 struct list_head *list;
2620 local_irq_save(flags);
2622 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2623 list = &pcp->lists[migratetype];
2624 if (list_empty(list)) {
2625 pcp->count += rmqueue_bulk(zone, 0,
2628 if (unlikely(list_empty(list)))
2633 page = list_last_entry(list, struct page, lru);
2635 page = list_first_entry(list, struct page, lru);
2637 list_del(&page->lru);
2640 } while (check_new_pcp(page));
2643 * We most definitely don't want callers attempting to
2644 * allocate greater than order-1 page units with __GFP_NOFAIL.
2646 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2647 spin_lock_irqsave(&zone->lock, flags);
2651 if (alloc_flags & ALLOC_HARDER) {
2652 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2654 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2657 page = __rmqueue(zone, order, migratetype);
2658 } while (page && check_new_pages(page, order));
2659 spin_unlock(&zone->lock);
2662 __mod_zone_freepage_state(zone, -(1 << order),
2663 get_pcppage_migratetype(page));
2666 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2667 zone_statistics(preferred_zone, zone);
2668 local_irq_restore(flags);
2670 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2674 local_irq_restore(flags);
2678 #ifdef CONFIG_FAIL_PAGE_ALLOC
2681 struct fault_attr attr;
2683 bool ignore_gfp_highmem;
2684 bool ignore_gfp_reclaim;
2686 } fail_page_alloc = {
2687 .attr = FAULT_ATTR_INITIALIZER,
2688 .ignore_gfp_reclaim = true,
2689 .ignore_gfp_highmem = true,
2693 static int __init setup_fail_page_alloc(char *str)
2695 return setup_fault_attr(&fail_page_alloc.attr, str);
2697 __setup("fail_page_alloc=", setup_fail_page_alloc);
2699 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2701 if (order < fail_page_alloc.min_order)
2703 if (gfp_mask & __GFP_NOFAIL)
2705 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2707 if (fail_page_alloc.ignore_gfp_reclaim &&
2708 (gfp_mask & __GFP_DIRECT_RECLAIM))
2711 return should_fail(&fail_page_alloc.attr, 1 << order);
2714 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2716 static int __init fail_page_alloc_debugfs(void)
2718 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2721 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2722 &fail_page_alloc.attr);
2724 return PTR_ERR(dir);
2726 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2727 &fail_page_alloc.ignore_gfp_reclaim))
2729 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2730 &fail_page_alloc.ignore_gfp_highmem))
2732 if (!debugfs_create_u32("min-order", mode, dir,
2733 &fail_page_alloc.min_order))
2738 debugfs_remove_recursive(dir);
2743 late_initcall(fail_page_alloc_debugfs);
2745 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2747 #else /* CONFIG_FAIL_PAGE_ALLOC */
2749 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2754 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2757 * Return true if free base pages are above 'mark'. For high-order checks it
2758 * will return true of the order-0 watermark is reached and there is at least
2759 * one free page of a suitable size. Checking now avoids taking the zone lock
2760 * to check in the allocation paths if no pages are free.
2762 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2763 int classzone_idx, unsigned int alloc_flags,
2768 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2770 /* free_pages may go negative - that's OK */
2771 free_pages -= (1 << order) - 1;
2773 if (alloc_flags & ALLOC_HIGH)
2777 * If the caller does not have rights to ALLOC_HARDER then subtract
2778 * the high-atomic reserves. This will over-estimate the size of the
2779 * atomic reserve but it avoids a search.
2781 if (likely(!alloc_harder))
2782 free_pages -= z->nr_reserved_highatomic;
2787 /* If allocation can't use CMA areas don't use free CMA pages */
2788 if (!(alloc_flags & ALLOC_CMA))
2789 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2793 * Check watermarks for an order-0 allocation request. If these
2794 * are not met, then a high-order request also cannot go ahead
2795 * even if a suitable page happened to be free.
2797 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2800 /* If this is an order-0 request then the watermark is fine */
2804 /* For a high-order request, check at least one suitable page is free */
2805 for (o = order; o < MAX_ORDER; o++) {
2806 struct free_area *area = &z->free_area[o];
2815 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2816 if (!list_empty(&area->free_list[mt]))
2821 if ((alloc_flags & ALLOC_CMA) &&
2822 !list_empty(&area->free_list[MIGRATE_CMA])) {
2830 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2831 int classzone_idx, unsigned int alloc_flags)
2833 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2834 zone_page_state(z, NR_FREE_PAGES));
2837 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2838 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2840 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2844 /* If allocation can't use CMA areas don't use free CMA pages */
2845 if (!(alloc_flags & ALLOC_CMA))
2846 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2850 * Fast check for order-0 only. If this fails then the reserves
2851 * need to be calculated. There is a corner case where the check
2852 * passes but only the high-order atomic reserve are free. If
2853 * the caller is !atomic then it'll uselessly search the free
2854 * list. That corner case is then slower but it is harmless.
2856 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2859 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2863 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2864 unsigned long mark, int classzone_idx)
2866 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2868 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2869 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2871 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2876 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2878 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2881 #else /* CONFIG_NUMA */
2882 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2886 #endif /* CONFIG_NUMA */
2889 * get_page_from_freelist goes through the zonelist trying to allocate
2892 static struct page *
2893 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2894 const struct alloc_context *ac)
2896 struct zoneref *z = ac->preferred_zoneref;
2898 struct pglist_data *last_pgdat_dirty_limit = NULL;
2901 * Scan zonelist, looking for a zone with enough free.
2902 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2904 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2909 if (cpusets_enabled() &&
2910 (alloc_flags & ALLOC_CPUSET) &&
2911 !__cpuset_zone_allowed(zone, gfp_mask))
2914 * When allocating a page cache page for writing, we
2915 * want to get it from a node that is within its dirty
2916 * limit, such that no single node holds more than its
2917 * proportional share of globally allowed dirty pages.
2918 * The dirty limits take into account the node's
2919 * lowmem reserves and high watermark so that kswapd
2920 * should be able to balance it without having to
2921 * write pages from its LRU list.
2923 * XXX: For now, allow allocations to potentially
2924 * exceed the per-node dirty limit in the slowpath
2925 * (spread_dirty_pages unset) before going into reclaim,
2926 * which is important when on a NUMA setup the allowed
2927 * nodes are together not big enough to reach the
2928 * global limit. The proper fix for these situations
2929 * will require awareness of nodes in the
2930 * dirty-throttling and the flusher threads.
2932 if (ac->spread_dirty_pages) {
2933 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2936 if (!node_dirty_ok(zone->zone_pgdat)) {
2937 last_pgdat_dirty_limit = zone->zone_pgdat;
2942 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2943 if (!zone_watermark_fast(zone, order, mark,
2944 ac_classzone_idx(ac), alloc_flags)) {
2947 /* Checked here to keep the fast path fast */
2948 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2949 if (alloc_flags & ALLOC_NO_WATERMARKS)
2952 if (node_reclaim_mode == 0 ||
2953 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2956 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2958 case NODE_RECLAIM_NOSCAN:
2961 case NODE_RECLAIM_FULL:
2962 /* scanned but unreclaimable */
2965 /* did we reclaim enough */
2966 if (zone_watermark_ok(zone, order, mark,
2967 ac_classzone_idx(ac), alloc_flags))
2975 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2976 gfp_mask, alloc_flags, ac->migratetype);
2978 prep_new_page(page, order, gfp_mask, alloc_flags);
2981 * If this is a high-order atomic allocation then check
2982 * if the pageblock should be reserved for the future
2984 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2985 reserve_highatomic_pageblock(page, zone, order);
2995 * Large machines with many possible nodes should not always dump per-node
2996 * meminfo in irq context.
2998 static inline bool should_suppress_show_mem(void)
3003 ret = in_interrupt();
3008 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3010 unsigned int filter = SHOW_MEM_FILTER_NODES;
3011 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3013 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3017 * This documents exceptions given to allocations in certain
3018 * contexts that are allowed to allocate outside current's set
3021 if (!(gfp_mask & __GFP_NOMEMALLOC))
3022 if (test_thread_flag(TIF_MEMDIE) ||
3023 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3024 filter &= ~SHOW_MEM_FILTER_NODES;
3025 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3026 filter &= ~SHOW_MEM_FILTER_NODES;
3028 show_mem(filter, nodemask);
3031 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3033 struct va_format vaf;
3035 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3036 DEFAULT_RATELIMIT_BURST);
3038 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3039 debug_guardpage_minorder() > 0)
3042 pr_warn("%s: ", current->comm);
3044 va_start(args, fmt);
3047 pr_cont("%pV", &vaf);
3050 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3052 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3054 pr_cont("(null)\n");
3056 cpuset_print_current_mems_allowed();
3059 warn_alloc_show_mem(gfp_mask, nodemask);
3062 static inline struct page *
3063 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3064 unsigned int alloc_flags,
3065 const struct alloc_context *ac)
3069 page = get_page_from_freelist(gfp_mask, order,
3070 alloc_flags|ALLOC_CPUSET, ac);
3072 * fallback to ignore cpuset restriction if our nodes
3076 page = get_page_from_freelist(gfp_mask, order,
3082 static inline struct page *
3083 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3084 const struct alloc_context *ac, unsigned long *did_some_progress)
3086 struct oom_control oc = {
3087 .zonelist = ac->zonelist,
3088 .nodemask = ac->nodemask,
3090 .gfp_mask = gfp_mask,
3095 *did_some_progress = 0;
3098 * Acquire the oom lock. If that fails, somebody else is
3099 * making progress for us.
3101 if (!mutex_trylock(&oom_lock)) {
3102 *did_some_progress = 1;
3103 schedule_timeout_uninterruptible(1);
3108 * Go through the zonelist yet one more time, keep very high watermark
3109 * here, this is only to catch a parallel oom killing, we must fail if
3110 * we're still under heavy pressure.
3112 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3113 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3117 /* Coredumps can quickly deplete all memory reserves */
3118 if (current->flags & PF_DUMPCORE)
3120 /* The OOM killer will not help higher order allocs */
3121 if (order > PAGE_ALLOC_COSTLY_ORDER)
3123 /* The OOM killer does not needlessly kill tasks for lowmem */
3124 if (ac->high_zoneidx < ZONE_NORMAL)
3126 if (pm_suspended_storage())
3129 * XXX: GFP_NOFS allocations should rather fail than rely on
3130 * other request to make a forward progress.
3131 * We are in an unfortunate situation where out_of_memory cannot
3132 * do much for this context but let's try it to at least get
3133 * access to memory reserved if the current task is killed (see
3134 * out_of_memory). Once filesystems are ready to handle allocation
3135 * failures more gracefully we should just bail out here.
3138 /* The OOM killer may not free memory on a specific node */
3139 if (gfp_mask & __GFP_THISNODE)
3142 /* Exhausted what can be done so it's blamo time */
3143 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3144 *did_some_progress = 1;
3147 * Help non-failing allocations by giving them access to memory
3150 if (gfp_mask & __GFP_NOFAIL)
3151 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3152 ALLOC_NO_WATERMARKS, ac);
3155 mutex_unlock(&oom_lock);
3160 * Maximum number of compaction retries wit a progress before OOM
3161 * killer is consider as the only way to move forward.
3163 #define MAX_COMPACT_RETRIES 16
3165 #ifdef CONFIG_COMPACTION
3166 /* Try memory compaction for high-order allocations before reclaim */
3167 static struct page *
3168 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3169 unsigned int alloc_flags, const struct alloc_context *ac,
3170 enum compact_priority prio, enum compact_result *compact_result)
3177 current->flags |= PF_MEMALLOC;
3178 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3180 current->flags &= ~PF_MEMALLOC;
3182 if (*compact_result <= COMPACT_INACTIVE)
3186 * At least in one zone compaction wasn't deferred or skipped, so let's
3187 * count a compaction stall
3189 count_vm_event(COMPACTSTALL);
3191 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3194 struct zone *zone = page_zone(page);
3196 zone->compact_blockskip_flush = false;
3197 compaction_defer_reset(zone, order, true);
3198 count_vm_event(COMPACTSUCCESS);
3203 * It's bad if compaction run occurs and fails. The most likely reason
3204 * is that pages exist, but not enough to satisfy watermarks.
3206 count_vm_event(COMPACTFAIL);
3214 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3215 enum compact_result compact_result,
3216 enum compact_priority *compact_priority,
3217 int *compaction_retries)
3219 int max_retries = MAX_COMPACT_RETRIES;
3222 int retries = *compaction_retries;
3223 enum compact_priority priority = *compact_priority;
3228 if (compaction_made_progress(compact_result))
3229 (*compaction_retries)++;
3232 * compaction considers all the zone as desperately out of memory
3233 * so it doesn't really make much sense to retry except when the
3234 * failure could be caused by insufficient priority
3236 if (compaction_failed(compact_result))
3237 goto check_priority;
3240 * make sure the compaction wasn't deferred or didn't bail out early
3241 * due to locks contention before we declare that we should give up.
3242 * But do not retry if the given zonelist is not suitable for
3245 if (compaction_withdrawn(compact_result)) {
3246 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3251 * !costly requests are much more important than __GFP_REPEAT
3252 * costly ones because they are de facto nofail and invoke OOM
3253 * killer to move on while costly can fail and users are ready
3254 * to cope with that. 1/4 retries is rather arbitrary but we
3255 * would need much more detailed feedback from compaction to
3256 * make a better decision.
3258 if (order > PAGE_ALLOC_COSTLY_ORDER)
3260 if (*compaction_retries <= max_retries) {
3266 * Make sure there are attempts at the highest priority if we exhausted
3267 * all retries or failed at the lower priorities.
3270 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3271 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3273 if (*compact_priority > min_priority) {
3274 (*compact_priority)--;
3275 *compaction_retries = 0;
3279 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3283 static inline struct page *
3284 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3285 unsigned int alloc_flags, const struct alloc_context *ac,
3286 enum compact_priority prio, enum compact_result *compact_result)
3288 *compact_result = COMPACT_SKIPPED;
3293 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3294 enum compact_result compact_result,
3295 enum compact_priority *compact_priority,
3296 int *compaction_retries)
3301 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3305 * There are setups with compaction disabled which would prefer to loop
3306 * inside the allocator rather than hit the oom killer prematurely.
3307 * Let's give them a good hope and keep retrying while the order-0
3308 * watermarks are OK.
3310 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3312 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3313 ac_classzone_idx(ac), alloc_flags))
3318 #endif /* CONFIG_COMPACTION */
3320 /* Perform direct synchronous page reclaim */
3322 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3323 const struct alloc_context *ac)
3325 struct reclaim_state reclaim_state;
3330 /* We now go into synchronous reclaim */
3331 cpuset_memory_pressure_bump();
3332 current->flags |= PF_MEMALLOC;
3333 lockdep_set_current_reclaim_state(gfp_mask);
3334 reclaim_state.reclaimed_slab = 0;
3335 current->reclaim_state = &reclaim_state;
3337 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3340 current->reclaim_state = NULL;
3341 lockdep_clear_current_reclaim_state();
3342 current->flags &= ~PF_MEMALLOC;
3349 /* The really slow allocator path where we enter direct reclaim */
3350 static inline struct page *
3351 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3352 unsigned int alloc_flags, const struct alloc_context *ac,
3353 unsigned long *did_some_progress)
3355 struct page *page = NULL;
3356 bool drained = false;
3358 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3359 if (unlikely(!(*did_some_progress)))
3363 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3366 * If an allocation failed after direct reclaim, it could be because
3367 * pages are pinned on the per-cpu lists or in high alloc reserves.
3368 * Shrink them them and try again
3370 if (!page && !drained) {
3371 unreserve_highatomic_pageblock(ac, false);
3372 drain_all_pages(NULL);
3380 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3384 pg_data_t *last_pgdat = NULL;
3386 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3387 ac->high_zoneidx, ac->nodemask) {
3388 if (last_pgdat != zone->zone_pgdat)
3389 wakeup_kswapd(zone, order, ac->high_zoneidx);
3390 last_pgdat = zone->zone_pgdat;
3394 static inline unsigned int
3395 gfp_to_alloc_flags(gfp_t gfp_mask)
3397 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3399 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3400 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3403 * The caller may dip into page reserves a bit more if the caller
3404 * cannot run direct reclaim, or if the caller has realtime scheduling
3405 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3406 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3408 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3410 if (gfp_mask & __GFP_ATOMIC) {
3412 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3413 * if it can't schedule.
3415 if (!(gfp_mask & __GFP_NOMEMALLOC))
3416 alloc_flags |= ALLOC_HARDER;
3418 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3419 * comment for __cpuset_node_allowed().
3421 alloc_flags &= ~ALLOC_CPUSET;
3422 } else if (unlikely(rt_task(current)) && !in_interrupt())
3423 alloc_flags |= ALLOC_HARDER;
3426 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3427 alloc_flags |= ALLOC_CMA;
3432 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3434 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3437 if (gfp_mask & __GFP_MEMALLOC)
3439 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3441 if (!in_interrupt() &&
3442 ((current->flags & PF_MEMALLOC) ||
3443 unlikely(test_thread_flag(TIF_MEMDIE))))
3450 * Maximum number of reclaim retries without any progress before OOM killer
3451 * is consider as the only way to move forward.
3453 #define MAX_RECLAIM_RETRIES 16
3456 * Checks whether it makes sense to retry the reclaim to make a forward progress
3457 * for the given allocation request.
3458 * The reclaim feedback represented by did_some_progress (any progress during
3459 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3460 * any progress in a row) is considered as well as the reclaimable pages on the
3461 * applicable zone list (with a backoff mechanism which is a function of
3462 * no_progress_loops).
3464 * Returns true if a retry is viable or false to enter the oom path.
3467 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3468 struct alloc_context *ac, int alloc_flags,
3469 bool did_some_progress, int *no_progress_loops)
3475 * Costly allocations might have made a progress but this doesn't mean
3476 * their order will become available due to high fragmentation so
3477 * always increment the no progress counter for them
3479 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3480 *no_progress_loops = 0;
3482 (*no_progress_loops)++;
3485 * Make sure we converge to OOM if we cannot make any progress
3486 * several times in the row.
3488 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3489 /* Before OOM, exhaust highatomic_reserve */
3490 return unreserve_highatomic_pageblock(ac, true);
3494 * Keep reclaiming pages while there is a chance this will lead
3495 * somewhere. If none of the target zones can satisfy our allocation
3496 * request even if all reclaimable pages are considered then we are
3497 * screwed and have to go OOM.
3499 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3501 unsigned long available;
3502 unsigned long reclaimable;
3503 unsigned long min_wmark = min_wmark_pages(zone);
3506 available = reclaimable = zone_reclaimable_pages(zone);
3507 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3508 MAX_RECLAIM_RETRIES);
3509 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3512 * Would the allocation succeed if we reclaimed the whole
3515 wmark = __zone_watermark_ok(zone, order, min_wmark,
3516 ac_classzone_idx(ac), alloc_flags, available);
3517 trace_reclaim_retry_zone(z, order, reclaimable,
3518 available, min_wmark, *no_progress_loops, wmark);
3521 * If we didn't make any progress and have a lot of
3522 * dirty + writeback pages then we should wait for
3523 * an IO to complete to slow down the reclaim and
3524 * prevent from pre mature OOM
3526 if (!did_some_progress) {
3527 unsigned long write_pending;
3529 write_pending = zone_page_state_snapshot(zone,
3530 NR_ZONE_WRITE_PENDING);
3532 if (2 * write_pending > reclaimable) {
3533 congestion_wait(BLK_RW_ASYNC, HZ/10);
3539 * Memory allocation/reclaim might be called from a WQ
3540 * context and the current implementation of the WQ
3541 * concurrency control doesn't recognize that
3542 * a particular WQ is congested if the worker thread is
3543 * looping without ever sleeping. Therefore we have to
3544 * do a short sleep here rather than calling
3547 if (current->flags & PF_WQ_WORKER)
3548 schedule_timeout_uninterruptible(1);
3559 static inline struct page *
3560 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3561 struct alloc_context *ac)
3563 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3564 struct page *page = NULL;
3565 unsigned int alloc_flags;
3566 unsigned long did_some_progress;
3567 enum compact_priority compact_priority;
3568 enum compact_result compact_result;
3569 int compaction_retries;
3570 int no_progress_loops;
3571 unsigned long alloc_start = jiffies;
3572 unsigned int stall_timeout = 10 * HZ;
3573 unsigned int cpuset_mems_cookie;
3576 * In the slowpath, we sanity check order to avoid ever trying to
3577 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3578 * be using allocators in order of preference for an area that is
3581 if (order >= MAX_ORDER) {
3582 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3587 * We also sanity check to catch abuse of atomic reserves being used by
3588 * callers that are not in atomic context.
3590 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3591 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3592 gfp_mask &= ~__GFP_ATOMIC;
3595 compaction_retries = 0;
3596 no_progress_loops = 0;
3597 compact_priority = DEF_COMPACT_PRIORITY;
3598 cpuset_mems_cookie = read_mems_allowed_begin();
3601 * The fast path uses conservative alloc_flags to succeed only until
3602 * kswapd needs to be woken up, and to avoid the cost of setting up
3603 * alloc_flags precisely. So we do that now.
3605 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3608 * We need to recalculate the starting point for the zonelist iterator
3609 * because we might have used different nodemask in the fast path, or
3610 * there was a cpuset modification and we are retrying - otherwise we
3611 * could end up iterating over non-eligible zones endlessly.
3613 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3614 ac->high_zoneidx, ac->nodemask);
3615 if (!ac->preferred_zoneref->zone)
3618 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3619 wake_all_kswapds(order, ac);
3622 * The adjusted alloc_flags might result in immediate success, so try
3625 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3630 * For costly allocations, try direct compaction first, as it's likely
3631 * that we have enough base pages and don't need to reclaim. Don't try
3632 * that for allocations that are allowed to ignore watermarks, as the
3633 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3635 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3636 !gfp_pfmemalloc_allowed(gfp_mask)) {
3637 page = __alloc_pages_direct_compact(gfp_mask, order,
3639 INIT_COMPACT_PRIORITY,
3645 * Checks for costly allocations with __GFP_NORETRY, which
3646 * includes THP page fault allocations
3648 if (gfp_mask & __GFP_NORETRY) {
3650 * If compaction is deferred for high-order allocations,
3651 * it is because sync compaction recently failed. If
3652 * this is the case and the caller requested a THP
3653 * allocation, we do not want to heavily disrupt the
3654 * system, so we fail the allocation instead of entering
3657 if (compact_result == COMPACT_DEFERRED)
3661 * Looks like reclaim/compaction is worth trying, but
3662 * sync compaction could be very expensive, so keep
3663 * using async compaction.
3665 compact_priority = INIT_COMPACT_PRIORITY;
3670 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3671 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3672 wake_all_kswapds(order, ac);
3674 if (gfp_pfmemalloc_allowed(gfp_mask))
3675 alloc_flags = ALLOC_NO_WATERMARKS;
3678 * Reset the zonelist iterators if memory policies can be ignored.
3679 * These allocations are high priority and system rather than user
3682 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3683 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3684 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3685 ac->high_zoneidx, ac->nodemask);
3688 /* Attempt with potentially adjusted zonelist and alloc_flags */
3689 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3693 /* Caller is not willing to reclaim, we can't balance anything */
3694 if (!can_direct_reclaim)
3697 /* Make sure we know about allocations which stall for too long */
3698 if (time_after(jiffies, alloc_start + stall_timeout)) {
3699 warn_alloc(gfp_mask, ac->nodemask,
3700 "page allocation stalls for %ums, order:%u",
3701 jiffies_to_msecs(jiffies-alloc_start), order);
3702 stall_timeout += 10 * HZ;
3705 /* Avoid recursion of direct reclaim */
3706 if (current->flags & PF_MEMALLOC)
3709 /* Try direct reclaim and then allocating */
3710 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3711 &did_some_progress);
3715 /* Try direct compaction and then allocating */
3716 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3717 compact_priority, &compact_result);
3721 /* Do not loop if specifically requested */
3722 if (gfp_mask & __GFP_NORETRY)
3726 * Do not retry costly high order allocations unless they are
3729 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3732 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3733 did_some_progress > 0, &no_progress_loops))
3737 * It doesn't make any sense to retry for the compaction if the order-0
3738 * reclaim is not able to make any progress because the current
3739 * implementation of the compaction depends on the sufficient amount
3740 * of free memory (see __compaction_suitable)
3742 if (did_some_progress > 0 &&
3743 should_compact_retry(ac, order, alloc_flags,
3744 compact_result, &compact_priority,
3745 &compaction_retries))
3749 * It's possible we raced with cpuset update so the OOM would be
3750 * premature (see below the nopage: label for full explanation).
3752 if (read_mems_allowed_retry(cpuset_mems_cookie))
3755 /* Reclaim has failed us, start killing things */
3756 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3760 /* Avoid allocations with no watermarks from looping endlessly */
3761 if (test_thread_flag(TIF_MEMDIE))
3764 /* Retry as long as the OOM killer is making progress */
3765 if (did_some_progress) {
3766 no_progress_loops = 0;
3772 * When updating a task's mems_allowed or mempolicy nodemask, it is
3773 * possible to race with parallel threads in such a way that our
3774 * allocation can fail while the mask is being updated. If we are about
3775 * to fail, check if the cpuset changed during allocation and if so,
3778 if (read_mems_allowed_retry(cpuset_mems_cookie))
3782 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3785 if (gfp_mask & __GFP_NOFAIL) {
3787 * All existing users of the __GFP_NOFAIL are blockable, so warn
3788 * of any new users that actually require GFP_NOWAIT
3790 if (WARN_ON_ONCE(!can_direct_reclaim))
3794 * PF_MEMALLOC request from this context is rather bizarre
3795 * because we cannot reclaim anything and only can loop waiting
3796 * for somebody to do a work for us
3798 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
3801 * non failing costly orders are a hard requirement which we
3802 * are not prepared for much so let's warn about these users
3803 * so that we can identify them and convert them to something
3806 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
3809 * Help non-failing allocations by giving them access to memory
3810 * reserves but do not use ALLOC_NO_WATERMARKS because this
3811 * could deplete whole memory reserves which would just make
3812 * the situation worse
3814 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
3822 warn_alloc(gfp_mask, ac->nodemask,
3823 "page allocation failure: order:%u", order);
3829 * This is the 'heart' of the zoned buddy allocator.
3832 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3833 struct zonelist *zonelist, nodemask_t *nodemask)
3836 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3837 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3838 struct alloc_context ac = {
3839 .high_zoneidx = gfp_zone(gfp_mask),
3840 .zonelist = zonelist,
3841 .nodemask = nodemask,
3842 .migratetype = gfpflags_to_migratetype(gfp_mask),
3845 if (cpusets_enabled()) {
3846 alloc_mask |= __GFP_HARDWALL;
3847 alloc_flags |= ALLOC_CPUSET;
3849 ac.nodemask = &cpuset_current_mems_allowed;
3852 gfp_mask &= gfp_allowed_mask;
3854 lockdep_trace_alloc(gfp_mask);
3856 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3858 if (should_fail_alloc_page(gfp_mask, order))
3862 * Check the zones suitable for the gfp_mask contain at least one
3863 * valid zone. It's possible to have an empty zonelist as a result
3864 * of __GFP_THISNODE and a memoryless node
3866 if (unlikely(!zonelist->_zonerefs->zone))
3869 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3870 alloc_flags |= ALLOC_CMA;
3872 /* Dirty zone balancing only done in the fast path */
3873 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3876 * The preferred zone is used for statistics but crucially it is
3877 * also used as the starting point for the zonelist iterator. It
3878 * may get reset for allocations that ignore memory policies.
3880 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3881 ac.high_zoneidx, ac.nodemask);
3882 if (!ac.preferred_zoneref->zone) {
3885 * This might be due to race with cpuset_current_mems_allowed
3886 * update, so make sure we retry with original nodemask in the
3892 /* First allocation attempt */
3893 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3899 * Runtime PM, block IO and its error handling path can deadlock
3900 * because I/O on the device might not complete.
3902 alloc_mask = memalloc_noio_flags(gfp_mask);
3903 ac.spread_dirty_pages = false;
3906 * Restore the original nodemask if it was potentially replaced with
3907 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3909 if (unlikely(ac.nodemask != nodemask))
3910 ac.nodemask = nodemask;
3912 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3915 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3916 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3917 __free_pages(page, order);
3921 if (kmemcheck_enabled && page)
3922 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3924 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3928 EXPORT_SYMBOL(__alloc_pages_nodemask);
3931 * Common helper functions.
3933 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3938 * __get_free_pages() returns a 32-bit address, which cannot represent
3941 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3943 page = alloc_pages(gfp_mask, order);
3946 return (unsigned long) page_address(page);
3948 EXPORT_SYMBOL(__get_free_pages);
3950 unsigned long get_zeroed_page(gfp_t gfp_mask)
3952 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3954 EXPORT_SYMBOL(get_zeroed_page);
3956 void __free_pages(struct page *page, unsigned int order)
3958 if (put_page_testzero(page)) {
3960 free_hot_cold_page(page, false);
3962 __free_pages_ok(page, order);
3966 EXPORT_SYMBOL(__free_pages);
3968 void free_pages(unsigned long addr, unsigned int order)
3971 VM_BUG_ON(!virt_addr_valid((void *)addr));
3972 __free_pages(virt_to_page((void *)addr), order);
3976 EXPORT_SYMBOL(free_pages);
3980 * An arbitrary-length arbitrary-offset area of memory which resides
3981 * within a 0 or higher order page. Multiple fragments within that page
3982 * are individually refcounted, in the page's reference counter.
3984 * The page_frag functions below provide a simple allocation framework for
3985 * page fragments. This is used by the network stack and network device
3986 * drivers to provide a backing region of memory for use as either an
3987 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3989 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
3992 struct page *page = NULL;
3993 gfp_t gfp = gfp_mask;
3995 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3996 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3998 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3999 PAGE_FRAG_CACHE_MAX_ORDER);
4000 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4002 if (unlikely(!page))
4003 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4005 nc->va = page ? page_address(page) : NULL;
4010 void __page_frag_cache_drain(struct page *page, unsigned int count)
4012 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4014 if (page_ref_sub_and_test(page, count)) {
4015 unsigned int order = compound_order(page);
4018 free_hot_cold_page(page, false);
4020 __free_pages_ok(page, order);
4023 EXPORT_SYMBOL(__page_frag_cache_drain);
4025 void *page_frag_alloc(struct page_frag_cache *nc,
4026 unsigned int fragsz, gfp_t gfp_mask)
4028 unsigned int size = PAGE_SIZE;
4032 if (unlikely(!nc->va)) {
4034 page = __page_frag_cache_refill(nc, gfp_mask);
4038 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4039 /* if size can vary use size else just use PAGE_SIZE */
4042 /* Even if we own the page, we do not use atomic_set().
4043 * This would break get_page_unless_zero() users.
4045 page_ref_add(page, size - 1);
4047 /* reset page count bias and offset to start of new frag */
4048 nc->pfmemalloc = page_is_pfmemalloc(page);
4049 nc->pagecnt_bias = size;
4053 offset = nc->offset - fragsz;
4054 if (unlikely(offset < 0)) {
4055 page = virt_to_page(nc->va);
4057 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4060 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4061 /* if size can vary use size else just use PAGE_SIZE */
4064 /* OK, page count is 0, we can safely set it */
4065 set_page_count(page, size);
4067 /* reset page count bias and offset to start of new frag */
4068 nc->pagecnt_bias = size;
4069 offset = size - fragsz;
4073 nc->offset = offset;
4075 return nc->va + offset;
4077 EXPORT_SYMBOL(page_frag_alloc);
4080 * Frees a page fragment allocated out of either a compound or order 0 page.
4082 void page_frag_free(void *addr)
4084 struct page *page = virt_to_head_page(addr);
4086 if (unlikely(put_page_testzero(page)))
4087 __free_pages_ok(page, compound_order(page));
4089 EXPORT_SYMBOL(page_frag_free);
4091 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4095 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4096 unsigned long used = addr + PAGE_ALIGN(size);
4098 split_page(virt_to_page((void *)addr), order);
4099 while (used < alloc_end) {
4104 return (void *)addr;
4108 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4109 * @size: the number of bytes to allocate
4110 * @gfp_mask: GFP flags for the allocation
4112 * This function is similar to alloc_pages(), except that it allocates the
4113 * minimum number of pages to satisfy the request. alloc_pages() can only
4114 * allocate memory in power-of-two pages.
4116 * This function is also limited by MAX_ORDER.
4118 * Memory allocated by this function must be released by free_pages_exact().
4120 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4122 unsigned int order = get_order(size);
4125 addr = __get_free_pages(gfp_mask, order);
4126 return make_alloc_exact(addr, order, size);
4128 EXPORT_SYMBOL(alloc_pages_exact);
4131 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4133 * @nid: the preferred node ID where memory should be allocated
4134 * @size: the number of bytes to allocate
4135 * @gfp_mask: GFP flags for the allocation
4137 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4140 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4142 unsigned int order = get_order(size);
4143 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4146 return make_alloc_exact((unsigned long)page_address(p), order, size);
4150 * free_pages_exact - release memory allocated via alloc_pages_exact()
4151 * @virt: the value returned by alloc_pages_exact.
4152 * @size: size of allocation, same value as passed to alloc_pages_exact().
4154 * Release the memory allocated by a previous call to alloc_pages_exact.
4156 void free_pages_exact(void *virt, size_t size)
4158 unsigned long addr = (unsigned long)virt;
4159 unsigned long end = addr + PAGE_ALIGN(size);
4161 while (addr < end) {
4166 EXPORT_SYMBOL(free_pages_exact);
4169 * nr_free_zone_pages - count number of pages beyond high watermark
4170 * @offset: The zone index of the highest zone
4172 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4173 * high watermark within all zones at or below a given zone index. For each
4174 * zone, the number of pages is calculated as:
4175 * managed_pages - high_pages
4177 static unsigned long nr_free_zone_pages(int offset)
4182 /* Just pick one node, since fallback list is circular */
4183 unsigned long sum = 0;
4185 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4187 for_each_zone_zonelist(zone, z, zonelist, offset) {
4188 unsigned long size = zone->managed_pages;
4189 unsigned long high = high_wmark_pages(zone);
4198 * nr_free_buffer_pages - count number of pages beyond high watermark
4200 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4201 * watermark within ZONE_DMA and ZONE_NORMAL.
4203 unsigned long nr_free_buffer_pages(void)
4205 return nr_free_zone_pages(gfp_zone(GFP_USER));
4207 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4210 * nr_free_pagecache_pages - count number of pages beyond high watermark
4212 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4213 * high watermark within all zones.
4215 unsigned long nr_free_pagecache_pages(void)
4217 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4220 static inline void show_node(struct zone *zone)
4222 if (IS_ENABLED(CONFIG_NUMA))
4223 printk("Node %d ", zone_to_nid(zone));
4226 long si_mem_available(void)
4229 unsigned long pagecache;
4230 unsigned long wmark_low = 0;
4231 unsigned long pages[NR_LRU_LISTS];
4235 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4236 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4239 wmark_low += zone->watermark[WMARK_LOW];
4242 * Estimate the amount of memory available for userspace allocations,
4243 * without causing swapping.
4245 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4248 * Not all the page cache can be freed, otherwise the system will
4249 * start swapping. Assume at least half of the page cache, or the
4250 * low watermark worth of cache, needs to stay.
4252 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4253 pagecache -= min(pagecache / 2, wmark_low);
4254 available += pagecache;
4257 * Part of the reclaimable slab consists of items that are in use,
4258 * and cannot be freed. Cap this estimate at the low watermark.
4260 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4261 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4267 EXPORT_SYMBOL_GPL(si_mem_available);
4269 void si_meminfo(struct sysinfo *val)
4271 val->totalram = totalram_pages;
4272 val->sharedram = global_node_page_state(NR_SHMEM);
4273 val->freeram = global_page_state(NR_FREE_PAGES);
4274 val->bufferram = nr_blockdev_pages();
4275 val->totalhigh = totalhigh_pages;
4276 val->freehigh = nr_free_highpages();
4277 val->mem_unit = PAGE_SIZE;
4280 EXPORT_SYMBOL(si_meminfo);
4283 void si_meminfo_node(struct sysinfo *val, int nid)
4285 int zone_type; /* needs to be signed */
4286 unsigned long managed_pages = 0;
4287 unsigned long managed_highpages = 0;
4288 unsigned long free_highpages = 0;
4289 pg_data_t *pgdat = NODE_DATA(nid);
4291 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4292 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4293 val->totalram = managed_pages;
4294 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4295 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4296 #ifdef CONFIG_HIGHMEM
4297 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4298 struct zone *zone = &pgdat->node_zones[zone_type];
4300 if (is_highmem(zone)) {
4301 managed_highpages += zone->managed_pages;
4302 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4305 val->totalhigh = managed_highpages;
4306 val->freehigh = free_highpages;
4308 val->totalhigh = managed_highpages;
4309 val->freehigh = free_highpages;
4311 val->mem_unit = PAGE_SIZE;
4316 * Determine whether the node should be displayed or not, depending on whether
4317 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4319 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4321 if (!(flags & SHOW_MEM_FILTER_NODES))
4325 * no node mask - aka implicit memory numa policy. Do not bother with
4326 * the synchronization - read_mems_allowed_begin - because we do not
4327 * have to be precise here.
4330 nodemask = &cpuset_current_mems_allowed;
4332 return !node_isset(nid, *nodemask);
4335 #define K(x) ((x) << (PAGE_SHIFT-10))
4337 static void show_migration_types(unsigned char type)
4339 static const char types[MIGRATE_TYPES] = {
4340 [MIGRATE_UNMOVABLE] = 'U',
4341 [MIGRATE_MOVABLE] = 'M',
4342 [MIGRATE_RECLAIMABLE] = 'E',
4343 [MIGRATE_HIGHATOMIC] = 'H',
4345 [MIGRATE_CMA] = 'C',
4347 #ifdef CONFIG_MEMORY_ISOLATION
4348 [MIGRATE_ISOLATE] = 'I',
4351 char tmp[MIGRATE_TYPES + 1];
4355 for (i = 0; i < MIGRATE_TYPES; i++) {
4356 if (type & (1 << i))
4361 printk(KERN_CONT "(%s) ", tmp);
4365 * Show free area list (used inside shift_scroll-lock stuff)
4366 * We also calculate the percentage fragmentation. We do this by counting the
4367 * memory on each free list with the exception of the first item on the list.
4370 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4373 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4375 unsigned long free_pcp = 0;
4380 for_each_populated_zone(zone) {
4381 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4384 for_each_online_cpu(cpu)
4385 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4388 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4389 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4390 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4391 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4392 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4393 " free:%lu free_pcp:%lu free_cma:%lu\n",
4394 global_node_page_state(NR_ACTIVE_ANON),
4395 global_node_page_state(NR_INACTIVE_ANON),
4396 global_node_page_state(NR_ISOLATED_ANON),
4397 global_node_page_state(NR_ACTIVE_FILE),
4398 global_node_page_state(NR_INACTIVE_FILE),
4399 global_node_page_state(NR_ISOLATED_FILE),
4400 global_node_page_state(NR_UNEVICTABLE),
4401 global_node_page_state(NR_FILE_DIRTY),
4402 global_node_page_state(NR_WRITEBACK),
4403 global_node_page_state(NR_UNSTABLE_NFS),
4404 global_page_state(NR_SLAB_RECLAIMABLE),
4405 global_page_state(NR_SLAB_UNRECLAIMABLE),
4406 global_node_page_state(NR_FILE_MAPPED),
4407 global_node_page_state(NR_SHMEM),
4408 global_page_state(NR_PAGETABLE),
4409 global_page_state(NR_BOUNCE),
4410 global_page_state(NR_FREE_PAGES),
4412 global_page_state(NR_FREE_CMA_PAGES));
4414 for_each_online_pgdat(pgdat) {
4415 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4419 " active_anon:%lukB"
4420 " inactive_anon:%lukB"
4421 " active_file:%lukB"
4422 " inactive_file:%lukB"
4423 " unevictable:%lukB"
4424 " isolated(anon):%lukB"
4425 " isolated(file):%lukB"
4430 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4432 " shmem_pmdmapped: %lukB"
4435 " writeback_tmp:%lukB"
4437 " pages_scanned:%lu"
4438 " all_unreclaimable? %s"
4441 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4442 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4443 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4444 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4445 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4446 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4447 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4448 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4449 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4450 K(node_page_state(pgdat, NR_WRITEBACK)),
4451 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4452 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4453 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4455 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4457 K(node_page_state(pgdat, NR_SHMEM)),
4458 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4459 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4460 node_page_state(pgdat, NR_PAGES_SCANNED),
4461 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4464 for_each_populated_zone(zone) {
4467 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4471 for_each_online_cpu(cpu)
4472 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4481 " active_anon:%lukB"
4482 " inactive_anon:%lukB"
4483 " active_file:%lukB"
4484 " inactive_file:%lukB"
4485 " unevictable:%lukB"
4486 " writepending:%lukB"
4490 " slab_reclaimable:%lukB"
4491 " slab_unreclaimable:%lukB"
4492 " kernel_stack:%lukB"
4500 K(zone_page_state(zone, NR_FREE_PAGES)),
4501 K(min_wmark_pages(zone)),
4502 K(low_wmark_pages(zone)),
4503 K(high_wmark_pages(zone)),
4504 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4505 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4506 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4507 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4508 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4509 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4510 K(zone->present_pages),
4511 K(zone->managed_pages),
4512 K(zone_page_state(zone, NR_MLOCK)),
4513 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4514 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4515 zone_page_state(zone, NR_KERNEL_STACK_KB),
4516 K(zone_page_state(zone, NR_PAGETABLE)),
4517 K(zone_page_state(zone, NR_BOUNCE)),
4519 K(this_cpu_read(zone->pageset->pcp.count)),
4520 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4521 printk("lowmem_reserve[]:");
4522 for (i = 0; i < MAX_NR_ZONES; i++)
4523 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4524 printk(KERN_CONT "\n");
4527 for_each_populated_zone(zone) {
4529 unsigned long nr[MAX_ORDER], flags, total = 0;
4530 unsigned char types[MAX_ORDER];
4532 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4535 printk(KERN_CONT "%s: ", zone->name);
4537 spin_lock_irqsave(&zone->lock, flags);
4538 for (order = 0; order < MAX_ORDER; order++) {
4539 struct free_area *area = &zone->free_area[order];
4542 nr[order] = area->nr_free;
4543 total += nr[order] << order;
4546 for (type = 0; type < MIGRATE_TYPES; type++) {
4547 if (!list_empty(&area->free_list[type]))
4548 types[order] |= 1 << type;
4551 spin_unlock_irqrestore(&zone->lock, flags);
4552 for (order = 0; order < MAX_ORDER; order++) {
4553 printk(KERN_CONT "%lu*%lukB ",
4554 nr[order], K(1UL) << order);
4556 show_migration_types(types[order]);
4558 printk(KERN_CONT "= %lukB\n", K(total));
4561 hugetlb_show_meminfo();
4563 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4565 show_swap_cache_info();
4568 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4570 zoneref->zone = zone;
4571 zoneref->zone_idx = zone_idx(zone);
4575 * Builds allocation fallback zone lists.
4577 * Add all populated zones of a node to the zonelist.
4579 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4583 enum zone_type zone_type = MAX_NR_ZONES;
4587 zone = pgdat->node_zones + zone_type;
4588 if (managed_zone(zone)) {
4589 zoneref_set_zone(zone,
4590 &zonelist->_zonerefs[nr_zones++]);
4591 check_highest_zone(zone_type);
4593 } while (zone_type);
4601 * 0 = automatic detection of better ordering.
4602 * 1 = order by ([node] distance, -zonetype)
4603 * 2 = order by (-zonetype, [node] distance)
4605 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4606 * the same zonelist. So only NUMA can configure this param.
4608 #define ZONELIST_ORDER_DEFAULT 0
4609 #define ZONELIST_ORDER_NODE 1
4610 #define ZONELIST_ORDER_ZONE 2
4612 /* zonelist order in the kernel.
4613 * set_zonelist_order() will set this to NODE or ZONE.
4615 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4616 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4620 /* The value user specified ....changed by config */
4621 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4622 /* string for sysctl */
4623 #define NUMA_ZONELIST_ORDER_LEN 16
4624 char numa_zonelist_order[16] = "default";
4627 * interface for configure zonelist ordering.
4628 * command line option "numa_zonelist_order"
4629 * = "[dD]efault - default, automatic configuration.
4630 * = "[nN]ode - order by node locality, then by zone within node
4631 * = "[zZ]one - order by zone, then by locality within zone
4634 static int __parse_numa_zonelist_order(char *s)
4636 if (*s == 'd' || *s == 'D') {
4637 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4638 } else if (*s == 'n' || *s == 'N') {
4639 user_zonelist_order = ZONELIST_ORDER_NODE;
4640 } else if (*s == 'z' || *s == 'Z') {
4641 user_zonelist_order = ZONELIST_ORDER_ZONE;
4643 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4649 static __init int setup_numa_zonelist_order(char *s)
4656 ret = __parse_numa_zonelist_order(s);
4658 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4662 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4665 * sysctl handler for numa_zonelist_order
4667 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4668 void __user *buffer, size_t *length,
4671 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4673 static DEFINE_MUTEX(zl_order_mutex);
4675 mutex_lock(&zl_order_mutex);
4677 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4681 strcpy(saved_string, (char *)table->data);
4683 ret = proc_dostring(table, write, buffer, length, ppos);
4687 int oldval = user_zonelist_order;
4689 ret = __parse_numa_zonelist_order((char *)table->data);
4692 * bogus value. restore saved string
4694 strncpy((char *)table->data, saved_string,
4695 NUMA_ZONELIST_ORDER_LEN);
4696 user_zonelist_order = oldval;
4697 } else if (oldval != user_zonelist_order) {
4698 mutex_lock(&zonelists_mutex);
4699 build_all_zonelists(NULL, NULL);
4700 mutex_unlock(&zonelists_mutex);
4704 mutex_unlock(&zl_order_mutex);
4709 #define MAX_NODE_LOAD (nr_online_nodes)
4710 static int node_load[MAX_NUMNODES];
4713 * find_next_best_node - find the next node that should appear in a given node's fallback list
4714 * @node: node whose fallback list we're appending
4715 * @used_node_mask: nodemask_t of already used nodes
4717 * We use a number of factors to determine which is the next node that should
4718 * appear on a given node's fallback list. The node should not have appeared
4719 * already in @node's fallback list, and it should be the next closest node
4720 * according to the distance array (which contains arbitrary distance values
4721 * from each node to each node in the system), and should also prefer nodes
4722 * with no CPUs, since presumably they'll have very little allocation pressure
4723 * on them otherwise.
4724 * It returns -1 if no node is found.
4726 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4729 int min_val = INT_MAX;
4730 int best_node = NUMA_NO_NODE;
4731 const struct cpumask *tmp = cpumask_of_node(0);
4733 /* Use the local node if we haven't already */
4734 if (!node_isset(node, *used_node_mask)) {
4735 node_set(node, *used_node_mask);
4739 for_each_node_state(n, N_MEMORY) {
4741 /* Don't want a node to appear more than once */
4742 if (node_isset(n, *used_node_mask))
4745 /* Use the distance array to find the distance */
4746 val = node_distance(node, n);
4748 /* Penalize nodes under us ("prefer the next node") */
4751 /* Give preference to headless and unused nodes */
4752 tmp = cpumask_of_node(n);
4753 if (!cpumask_empty(tmp))
4754 val += PENALTY_FOR_NODE_WITH_CPUS;
4756 /* Slight preference for less loaded node */
4757 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4758 val += node_load[n];
4760 if (val < min_val) {
4767 node_set(best_node, *used_node_mask);
4774 * Build zonelists ordered by node and zones within node.
4775 * This results in maximum locality--normal zone overflows into local
4776 * DMA zone, if any--but risks exhausting DMA zone.
4778 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4781 struct zonelist *zonelist;
4783 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4784 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4786 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4787 zonelist->_zonerefs[j].zone = NULL;
4788 zonelist->_zonerefs[j].zone_idx = 0;
4792 * Build gfp_thisnode zonelists
4794 static void build_thisnode_zonelists(pg_data_t *pgdat)
4797 struct zonelist *zonelist;
4799 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4800 j = build_zonelists_node(pgdat, zonelist, 0);
4801 zonelist->_zonerefs[j].zone = NULL;
4802 zonelist->_zonerefs[j].zone_idx = 0;
4806 * Build zonelists ordered by zone and nodes within zones.
4807 * This results in conserving DMA zone[s] until all Normal memory is
4808 * exhausted, but results in overflowing to remote node while memory
4809 * may still exist in local DMA zone.
4811 static int node_order[MAX_NUMNODES];
4813 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4816 int zone_type; /* needs to be signed */
4818 struct zonelist *zonelist;
4820 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4822 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4823 for (j = 0; j < nr_nodes; j++) {
4824 node = node_order[j];
4825 z = &NODE_DATA(node)->node_zones[zone_type];
4826 if (managed_zone(z)) {
4828 &zonelist->_zonerefs[pos++]);
4829 check_highest_zone(zone_type);
4833 zonelist->_zonerefs[pos].zone = NULL;
4834 zonelist->_zonerefs[pos].zone_idx = 0;
4837 #if defined(CONFIG_64BIT)
4839 * Devices that require DMA32/DMA are relatively rare and do not justify a
4840 * penalty to every machine in case the specialised case applies. Default
4841 * to Node-ordering on 64-bit NUMA machines
4843 static int default_zonelist_order(void)
4845 return ZONELIST_ORDER_NODE;
4849 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4850 * by the kernel. If processes running on node 0 deplete the low memory zone
4851 * then reclaim will occur more frequency increasing stalls and potentially
4852 * be easier to OOM if a large percentage of the zone is under writeback or
4853 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4854 * Hence, default to zone ordering on 32-bit.
4856 static int default_zonelist_order(void)
4858 return ZONELIST_ORDER_ZONE;
4860 #endif /* CONFIG_64BIT */
4862 static void set_zonelist_order(void)
4864 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4865 current_zonelist_order = default_zonelist_order();
4867 current_zonelist_order = user_zonelist_order;
4870 static void build_zonelists(pg_data_t *pgdat)
4873 nodemask_t used_mask;
4874 int local_node, prev_node;
4875 struct zonelist *zonelist;
4876 unsigned int order = current_zonelist_order;
4878 /* initialize zonelists */
4879 for (i = 0; i < MAX_ZONELISTS; i++) {
4880 zonelist = pgdat->node_zonelists + i;
4881 zonelist->_zonerefs[0].zone = NULL;
4882 zonelist->_zonerefs[0].zone_idx = 0;
4885 /* NUMA-aware ordering of nodes */
4886 local_node = pgdat->node_id;
4887 load = nr_online_nodes;
4888 prev_node = local_node;
4889 nodes_clear(used_mask);
4891 memset(node_order, 0, sizeof(node_order));
4894 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4896 * We don't want to pressure a particular node.
4897 * So adding penalty to the first node in same
4898 * distance group to make it round-robin.
4900 if (node_distance(local_node, node) !=
4901 node_distance(local_node, prev_node))
4902 node_load[node] = load;
4906 if (order == ZONELIST_ORDER_NODE)
4907 build_zonelists_in_node_order(pgdat, node);
4909 node_order[i++] = node; /* remember order */
4912 if (order == ZONELIST_ORDER_ZONE) {
4913 /* calculate node order -- i.e., DMA last! */
4914 build_zonelists_in_zone_order(pgdat, i);
4917 build_thisnode_zonelists(pgdat);
4920 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4922 * Return node id of node used for "local" allocations.
4923 * I.e., first node id of first zone in arg node's generic zonelist.
4924 * Used for initializing percpu 'numa_mem', which is used primarily
4925 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4927 int local_memory_node(int node)
4931 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4932 gfp_zone(GFP_KERNEL),
4934 return z->zone->node;
4938 static void setup_min_unmapped_ratio(void);
4939 static void setup_min_slab_ratio(void);
4940 #else /* CONFIG_NUMA */
4942 static void set_zonelist_order(void)
4944 current_zonelist_order = ZONELIST_ORDER_ZONE;
4947 static void build_zonelists(pg_data_t *pgdat)
4949 int node, local_node;
4951 struct zonelist *zonelist;
4953 local_node = pgdat->node_id;
4955 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4956 j = build_zonelists_node(pgdat, zonelist, 0);
4959 * Now we build the zonelist so that it contains the zones
4960 * of all the other nodes.
4961 * We don't want to pressure a particular node, so when
4962 * building the zones for node N, we make sure that the
4963 * zones coming right after the local ones are those from
4964 * node N+1 (modulo N)
4966 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4967 if (!node_online(node))
4969 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4971 for (node = 0; node < local_node; node++) {
4972 if (!node_online(node))
4974 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4977 zonelist->_zonerefs[j].zone = NULL;
4978 zonelist->_zonerefs[j].zone_idx = 0;
4981 #endif /* CONFIG_NUMA */
4984 * Boot pageset table. One per cpu which is going to be used for all
4985 * zones and all nodes. The parameters will be set in such a way
4986 * that an item put on a list will immediately be handed over to
4987 * the buddy list. This is safe since pageset manipulation is done
4988 * with interrupts disabled.
4990 * The boot_pagesets must be kept even after bootup is complete for
4991 * unused processors and/or zones. They do play a role for bootstrapping
4992 * hotplugged processors.
4994 * zoneinfo_show() and maybe other functions do
4995 * not check if the processor is online before following the pageset pointer.
4996 * Other parts of the kernel may not check if the zone is available.
4998 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4999 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5000 static void setup_zone_pageset(struct zone *zone);
5003 * Global mutex to protect against size modification of zonelists
5004 * as well as to serialize pageset setup for the new populated zone.
5006 DEFINE_MUTEX(zonelists_mutex);
5008 /* return values int ....just for stop_machine() */
5009 static int __build_all_zonelists(void *data)
5013 pg_data_t *self = data;
5016 memset(node_load, 0, sizeof(node_load));
5019 if (self && !node_online(self->node_id)) {
5020 build_zonelists(self);
5023 for_each_online_node(nid) {
5024 pg_data_t *pgdat = NODE_DATA(nid);
5026 build_zonelists(pgdat);
5030 * Initialize the boot_pagesets that are going to be used
5031 * for bootstrapping processors. The real pagesets for
5032 * each zone will be allocated later when the per cpu
5033 * allocator is available.
5035 * boot_pagesets are used also for bootstrapping offline
5036 * cpus if the system is already booted because the pagesets
5037 * are needed to initialize allocators on a specific cpu too.
5038 * F.e. the percpu allocator needs the page allocator which
5039 * needs the percpu allocator in order to allocate its pagesets
5040 * (a chicken-egg dilemma).
5042 for_each_possible_cpu(cpu) {
5043 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5045 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5047 * We now know the "local memory node" for each node--
5048 * i.e., the node of the first zone in the generic zonelist.
5049 * Set up numa_mem percpu variable for on-line cpus. During
5050 * boot, only the boot cpu should be on-line; we'll init the
5051 * secondary cpus' numa_mem as they come on-line. During
5052 * node/memory hotplug, we'll fixup all on-line cpus.
5054 if (cpu_online(cpu))
5055 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5062 static noinline void __init
5063 build_all_zonelists_init(void)
5065 __build_all_zonelists(NULL);
5066 mminit_verify_zonelist();
5067 cpuset_init_current_mems_allowed();
5071 * Called with zonelists_mutex held always
5072 * unless system_state == SYSTEM_BOOTING.
5074 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5075 * [we're only called with non-NULL zone through __meminit paths] and
5076 * (2) call of __init annotated helper build_all_zonelists_init
5077 * [protected by SYSTEM_BOOTING].
5079 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5081 set_zonelist_order();
5083 if (system_state == SYSTEM_BOOTING) {
5084 build_all_zonelists_init();
5086 #ifdef CONFIG_MEMORY_HOTPLUG
5088 setup_zone_pageset(zone);
5090 /* we have to stop all cpus to guarantee there is no user
5092 stop_machine(__build_all_zonelists, pgdat, NULL);
5093 /* cpuset refresh routine should be here */
5095 vm_total_pages = nr_free_pagecache_pages();
5097 * Disable grouping by mobility if the number of pages in the
5098 * system is too low to allow the mechanism to work. It would be
5099 * more accurate, but expensive to check per-zone. This check is
5100 * made on memory-hotadd so a system can start with mobility
5101 * disabled and enable it later
5103 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5104 page_group_by_mobility_disabled = 1;
5106 page_group_by_mobility_disabled = 0;
5108 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5110 zonelist_order_name[current_zonelist_order],
5111 page_group_by_mobility_disabled ? "off" : "on",
5114 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5119 * Initially all pages are reserved - free ones are freed
5120 * up by free_all_bootmem() once the early boot process is
5121 * done. Non-atomic initialization, single-pass.
5123 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5124 unsigned long start_pfn, enum memmap_context context)
5126 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5127 unsigned long end_pfn = start_pfn + size;
5128 pg_data_t *pgdat = NODE_DATA(nid);
5130 unsigned long nr_initialised = 0;
5131 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5132 struct memblock_region *r = NULL, *tmp;
5135 if (highest_memmap_pfn < end_pfn - 1)
5136 highest_memmap_pfn = end_pfn - 1;
5139 * Honor reservation requested by the driver for this ZONE_DEVICE
5142 if (altmap && start_pfn == altmap->base_pfn)
5143 start_pfn += altmap->reserve;
5145 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5147 * There can be holes in boot-time mem_map[]s handed to this
5148 * function. They do not exist on hotplugged memory.
5150 if (context != MEMMAP_EARLY)
5153 if (!early_pfn_valid(pfn)) {
5154 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5156 * Skip to the pfn preceding the next valid one (or
5157 * end_pfn), such that we hit a valid pfn (or end_pfn)
5158 * on our next iteration of the loop.
5160 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5164 if (!early_pfn_in_nid(pfn, nid))
5166 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5169 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5171 * Check given memblock attribute by firmware which can affect
5172 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5173 * mirrored, it's an overlapped memmap init. skip it.
5175 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5176 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5177 for_each_memblock(memory, tmp)
5178 if (pfn < memblock_region_memory_end_pfn(tmp))
5182 if (pfn >= memblock_region_memory_base_pfn(r) &&
5183 memblock_is_mirror(r)) {
5184 /* already initialized as NORMAL */
5185 pfn = memblock_region_memory_end_pfn(r);
5193 * Mark the block movable so that blocks are reserved for
5194 * movable at startup. This will force kernel allocations
5195 * to reserve their blocks rather than leaking throughout
5196 * the address space during boot when many long-lived
5197 * kernel allocations are made.
5199 * bitmap is created for zone's valid pfn range. but memmap
5200 * can be created for invalid pages (for alignment)
5201 * check here not to call set_pageblock_migratetype() against
5204 if (!(pfn & (pageblock_nr_pages - 1))) {
5205 struct page *page = pfn_to_page(pfn);
5207 __init_single_page(page, pfn, zone, nid);
5208 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5210 __init_single_pfn(pfn, zone, nid);
5215 static void __meminit zone_init_free_lists(struct zone *zone)
5217 unsigned int order, t;
5218 for_each_migratetype_order(order, t) {
5219 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5220 zone->free_area[order].nr_free = 0;
5224 #ifndef __HAVE_ARCH_MEMMAP_INIT
5225 #define memmap_init(size, nid, zone, start_pfn) \
5226 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5229 static int zone_batchsize(struct zone *zone)
5235 * The per-cpu-pages pools are set to around 1000th of the
5236 * size of the zone. But no more than 1/2 of a meg.
5238 * OK, so we don't know how big the cache is. So guess.
5240 batch = zone->managed_pages / 1024;
5241 if (batch * PAGE_SIZE > 512 * 1024)
5242 batch = (512 * 1024) / PAGE_SIZE;
5243 batch /= 4; /* We effectively *= 4 below */
5248 * Clamp the batch to a 2^n - 1 value. Having a power
5249 * of 2 value was found to be more likely to have
5250 * suboptimal cache aliasing properties in some cases.
5252 * For example if 2 tasks are alternately allocating
5253 * batches of pages, one task can end up with a lot
5254 * of pages of one half of the possible page colors
5255 * and the other with pages of the other colors.
5257 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5262 /* The deferral and batching of frees should be suppressed under NOMMU
5265 * The problem is that NOMMU needs to be able to allocate large chunks
5266 * of contiguous memory as there's no hardware page translation to
5267 * assemble apparent contiguous memory from discontiguous pages.
5269 * Queueing large contiguous runs of pages for batching, however,
5270 * causes the pages to actually be freed in smaller chunks. As there
5271 * can be a significant delay between the individual batches being
5272 * recycled, this leads to the once large chunks of space being
5273 * fragmented and becoming unavailable for high-order allocations.
5280 * pcp->high and pcp->batch values are related and dependent on one another:
5281 * ->batch must never be higher then ->high.
5282 * The following function updates them in a safe manner without read side
5285 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5286 * those fields changing asynchronously (acording the the above rule).
5288 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5289 * outside of boot time (or some other assurance that no concurrent updaters
5292 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5293 unsigned long batch)
5295 /* start with a fail safe value for batch */
5299 /* Update high, then batch, in order */
5306 /* a companion to pageset_set_high() */
5307 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5309 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5312 static void pageset_init(struct per_cpu_pageset *p)
5314 struct per_cpu_pages *pcp;
5317 memset(p, 0, sizeof(*p));
5321 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5322 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5325 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5328 pageset_set_batch(p, batch);
5332 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5333 * to the value high for the pageset p.
5335 static void pageset_set_high(struct per_cpu_pageset *p,
5338 unsigned long batch = max(1UL, high / 4);
5339 if ((high / 4) > (PAGE_SHIFT * 8))
5340 batch = PAGE_SHIFT * 8;
5342 pageset_update(&p->pcp, high, batch);
5345 static void pageset_set_high_and_batch(struct zone *zone,
5346 struct per_cpu_pageset *pcp)
5348 if (percpu_pagelist_fraction)
5349 pageset_set_high(pcp,
5350 (zone->managed_pages /
5351 percpu_pagelist_fraction));
5353 pageset_set_batch(pcp, zone_batchsize(zone));
5356 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5358 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5361 pageset_set_high_and_batch(zone, pcp);
5364 static void __meminit setup_zone_pageset(struct zone *zone)
5367 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5368 for_each_possible_cpu(cpu)
5369 zone_pageset_init(zone, cpu);
5373 * Allocate per cpu pagesets and initialize them.
5374 * Before this call only boot pagesets were available.
5376 void __init setup_per_cpu_pageset(void)
5378 struct pglist_data *pgdat;
5381 for_each_populated_zone(zone)
5382 setup_zone_pageset(zone);
5384 for_each_online_pgdat(pgdat)
5385 pgdat->per_cpu_nodestats =
5386 alloc_percpu(struct per_cpu_nodestat);
5389 static __meminit void zone_pcp_init(struct zone *zone)
5392 * per cpu subsystem is not up at this point. The following code
5393 * relies on the ability of the linker to provide the
5394 * offset of a (static) per cpu variable into the per cpu area.
5396 zone->pageset = &boot_pageset;
5398 if (populated_zone(zone))
5399 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5400 zone->name, zone->present_pages,
5401 zone_batchsize(zone));
5404 int __meminit init_currently_empty_zone(struct zone *zone,
5405 unsigned long zone_start_pfn,
5408 struct pglist_data *pgdat = zone->zone_pgdat;
5410 pgdat->nr_zones = zone_idx(zone) + 1;
5412 zone->zone_start_pfn = zone_start_pfn;
5414 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5415 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5417 (unsigned long)zone_idx(zone),
5418 zone_start_pfn, (zone_start_pfn + size));
5420 zone_init_free_lists(zone);
5421 zone->initialized = 1;
5426 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5427 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5430 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5432 int __meminit __early_pfn_to_nid(unsigned long pfn,
5433 struct mminit_pfnnid_cache *state)
5435 unsigned long start_pfn, end_pfn;
5438 if (state->last_start <= pfn && pfn < state->last_end)
5439 return state->last_nid;
5441 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5443 state->last_start = start_pfn;
5444 state->last_end = end_pfn;
5445 state->last_nid = nid;
5450 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5453 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5454 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5455 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5457 * If an architecture guarantees that all ranges registered contain no holes
5458 * and may be freed, this this function may be used instead of calling
5459 * memblock_free_early_nid() manually.
5461 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5463 unsigned long start_pfn, end_pfn;
5466 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5467 start_pfn = min(start_pfn, max_low_pfn);
5468 end_pfn = min(end_pfn, max_low_pfn);
5470 if (start_pfn < end_pfn)
5471 memblock_free_early_nid(PFN_PHYS(start_pfn),
5472 (end_pfn - start_pfn) << PAGE_SHIFT,
5478 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5479 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5481 * If an architecture guarantees that all ranges registered contain no holes and may
5482 * be freed, this function may be used instead of calling memory_present() manually.
5484 void __init sparse_memory_present_with_active_regions(int nid)
5486 unsigned long start_pfn, end_pfn;
5489 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5490 memory_present(this_nid, start_pfn, end_pfn);
5494 * get_pfn_range_for_nid - Return the start and end page frames for a node
5495 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5496 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5497 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5499 * It returns the start and end page frame of a node based on information
5500 * provided by memblock_set_node(). If called for a node
5501 * with no available memory, a warning is printed and the start and end
5504 void __meminit get_pfn_range_for_nid(unsigned int nid,
5505 unsigned long *start_pfn, unsigned long *end_pfn)
5507 unsigned long this_start_pfn, this_end_pfn;
5513 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5514 *start_pfn = min(*start_pfn, this_start_pfn);
5515 *end_pfn = max(*end_pfn, this_end_pfn);
5518 if (*start_pfn == -1UL)
5523 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5524 * assumption is made that zones within a node are ordered in monotonic
5525 * increasing memory addresses so that the "highest" populated zone is used
5527 static void __init find_usable_zone_for_movable(void)
5530 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5531 if (zone_index == ZONE_MOVABLE)
5534 if (arch_zone_highest_possible_pfn[zone_index] >
5535 arch_zone_lowest_possible_pfn[zone_index])
5539 VM_BUG_ON(zone_index == -1);
5540 movable_zone = zone_index;
5544 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5545 * because it is sized independent of architecture. Unlike the other zones,
5546 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5547 * in each node depending on the size of each node and how evenly kernelcore
5548 * is distributed. This helper function adjusts the zone ranges
5549 * provided by the architecture for a given node by using the end of the
5550 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5551 * zones within a node are in order of monotonic increases memory addresses
5553 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5554 unsigned long zone_type,
5555 unsigned long node_start_pfn,
5556 unsigned long node_end_pfn,
5557 unsigned long *zone_start_pfn,
5558 unsigned long *zone_end_pfn)
5560 /* Only adjust if ZONE_MOVABLE is on this node */
5561 if (zone_movable_pfn[nid]) {
5562 /* Size ZONE_MOVABLE */
5563 if (zone_type == ZONE_MOVABLE) {
5564 *zone_start_pfn = zone_movable_pfn[nid];
5565 *zone_end_pfn = min(node_end_pfn,
5566 arch_zone_highest_possible_pfn[movable_zone]);
5568 /* Adjust for ZONE_MOVABLE starting within this range */
5569 } else if (!mirrored_kernelcore &&
5570 *zone_start_pfn < zone_movable_pfn[nid] &&
5571 *zone_end_pfn > zone_movable_pfn[nid]) {
5572 *zone_end_pfn = zone_movable_pfn[nid];
5574 /* Check if this whole range is within ZONE_MOVABLE */
5575 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5576 *zone_start_pfn = *zone_end_pfn;
5581 * Return the number of pages a zone spans in a node, including holes
5582 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5584 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5585 unsigned long zone_type,
5586 unsigned long node_start_pfn,
5587 unsigned long node_end_pfn,
5588 unsigned long *zone_start_pfn,
5589 unsigned long *zone_end_pfn,
5590 unsigned long *ignored)
5592 /* When hotadd a new node from cpu_up(), the node should be empty */
5593 if (!node_start_pfn && !node_end_pfn)
5596 /* Get the start and end of the zone */
5597 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5598 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5599 adjust_zone_range_for_zone_movable(nid, zone_type,
5600 node_start_pfn, node_end_pfn,
5601 zone_start_pfn, zone_end_pfn);
5603 /* Check that this node has pages within the zone's required range */
5604 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5607 /* Move the zone boundaries inside the node if necessary */
5608 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5609 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5611 /* Return the spanned pages */
5612 return *zone_end_pfn - *zone_start_pfn;
5616 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5617 * then all holes in the requested range will be accounted for.
5619 unsigned long __meminit __absent_pages_in_range(int nid,
5620 unsigned long range_start_pfn,
5621 unsigned long range_end_pfn)
5623 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5624 unsigned long start_pfn, end_pfn;
5627 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5628 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5629 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5630 nr_absent -= end_pfn - start_pfn;
5636 * absent_pages_in_range - Return number of page frames in holes within a range
5637 * @start_pfn: The start PFN to start searching for holes
5638 * @end_pfn: The end PFN to stop searching for holes
5640 * It returns the number of pages frames in memory holes within a range.
5642 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5643 unsigned long end_pfn)
5645 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5648 /* Return the number of page frames in holes in a zone on a node */
5649 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5650 unsigned long zone_type,
5651 unsigned long node_start_pfn,
5652 unsigned long node_end_pfn,
5653 unsigned long *ignored)
5655 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5656 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5657 unsigned long zone_start_pfn, zone_end_pfn;
5658 unsigned long nr_absent;
5660 /* When hotadd a new node from cpu_up(), the node should be empty */
5661 if (!node_start_pfn && !node_end_pfn)
5664 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5665 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5667 adjust_zone_range_for_zone_movable(nid, zone_type,
5668 node_start_pfn, node_end_pfn,
5669 &zone_start_pfn, &zone_end_pfn);
5670 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5673 * ZONE_MOVABLE handling.
5674 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5677 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5678 unsigned long start_pfn, end_pfn;
5679 struct memblock_region *r;
5681 for_each_memblock(memory, r) {
5682 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5683 zone_start_pfn, zone_end_pfn);
5684 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5685 zone_start_pfn, zone_end_pfn);
5687 if (zone_type == ZONE_MOVABLE &&
5688 memblock_is_mirror(r))
5689 nr_absent += end_pfn - start_pfn;
5691 if (zone_type == ZONE_NORMAL &&
5692 !memblock_is_mirror(r))
5693 nr_absent += end_pfn - start_pfn;
5700 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5701 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5702 unsigned long zone_type,
5703 unsigned long node_start_pfn,
5704 unsigned long node_end_pfn,
5705 unsigned long *zone_start_pfn,
5706 unsigned long *zone_end_pfn,
5707 unsigned long *zones_size)
5711 *zone_start_pfn = node_start_pfn;
5712 for (zone = 0; zone < zone_type; zone++)
5713 *zone_start_pfn += zones_size[zone];
5715 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5717 return zones_size[zone_type];
5720 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5721 unsigned long zone_type,
5722 unsigned long node_start_pfn,
5723 unsigned long node_end_pfn,
5724 unsigned long *zholes_size)
5729 return zholes_size[zone_type];
5732 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5734 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5735 unsigned long node_start_pfn,
5736 unsigned long node_end_pfn,
5737 unsigned long *zones_size,
5738 unsigned long *zholes_size)
5740 unsigned long realtotalpages = 0, totalpages = 0;
5743 for (i = 0; i < MAX_NR_ZONES; i++) {
5744 struct zone *zone = pgdat->node_zones + i;
5745 unsigned long zone_start_pfn, zone_end_pfn;
5746 unsigned long size, real_size;
5748 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5754 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5755 node_start_pfn, node_end_pfn,
5758 zone->zone_start_pfn = zone_start_pfn;
5760 zone->zone_start_pfn = 0;
5761 zone->spanned_pages = size;
5762 zone->present_pages = real_size;
5765 realtotalpages += real_size;
5768 pgdat->node_spanned_pages = totalpages;
5769 pgdat->node_present_pages = realtotalpages;
5770 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5774 #ifndef CONFIG_SPARSEMEM
5776 * Calculate the size of the zone->blockflags rounded to an unsigned long
5777 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5778 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5779 * round what is now in bits to nearest long in bits, then return it in
5782 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5784 unsigned long usemapsize;
5786 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5787 usemapsize = roundup(zonesize, pageblock_nr_pages);
5788 usemapsize = usemapsize >> pageblock_order;
5789 usemapsize *= NR_PAGEBLOCK_BITS;
5790 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5792 return usemapsize / 8;
5795 static void __init setup_usemap(struct pglist_data *pgdat,
5797 unsigned long zone_start_pfn,
5798 unsigned long zonesize)
5800 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5801 zone->pageblock_flags = NULL;
5803 zone->pageblock_flags =
5804 memblock_virt_alloc_node_nopanic(usemapsize,
5808 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5809 unsigned long zone_start_pfn, unsigned long zonesize) {}
5810 #endif /* CONFIG_SPARSEMEM */
5812 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5814 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5815 void __paginginit set_pageblock_order(void)
5819 /* Check that pageblock_nr_pages has not already been setup */
5820 if (pageblock_order)
5823 if (HPAGE_SHIFT > PAGE_SHIFT)
5824 order = HUGETLB_PAGE_ORDER;
5826 order = MAX_ORDER - 1;
5829 * Assume the largest contiguous order of interest is a huge page.
5830 * This value may be variable depending on boot parameters on IA64 and
5833 pageblock_order = order;
5835 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5838 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5839 * is unused as pageblock_order is set at compile-time. See
5840 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5843 void __paginginit set_pageblock_order(void)
5847 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5849 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5850 unsigned long present_pages)
5852 unsigned long pages = spanned_pages;
5855 * Provide a more accurate estimation if there are holes within
5856 * the zone and SPARSEMEM is in use. If there are holes within the
5857 * zone, each populated memory region may cost us one or two extra
5858 * memmap pages due to alignment because memmap pages for each
5859 * populated regions may not naturally algined on page boundary.
5860 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5862 if (spanned_pages > present_pages + (present_pages >> 4) &&
5863 IS_ENABLED(CONFIG_SPARSEMEM))
5864 pages = present_pages;
5866 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5870 * Set up the zone data structures:
5871 * - mark all pages reserved
5872 * - mark all memory queues empty
5873 * - clear the memory bitmaps
5875 * NOTE: pgdat should get zeroed by caller.
5877 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5880 int nid = pgdat->node_id;
5883 pgdat_resize_init(pgdat);
5884 #ifdef CONFIG_NUMA_BALANCING
5885 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5886 pgdat->numabalancing_migrate_nr_pages = 0;
5887 pgdat->numabalancing_migrate_next_window = jiffies;
5889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5890 spin_lock_init(&pgdat->split_queue_lock);
5891 INIT_LIST_HEAD(&pgdat->split_queue);
5892 pgdat->split_queue_len = 0;
5894 init_waitqueue_head(&pgdat->kswapd_wait);
5895 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5896 #ifdef CONFIG_COMPACTION
5897 init_waitqueue_head(&pgdat->kcompactd_wait);
5899 pgdat_page_ext_init(pgdat);
5900 spin_lock_init(&pgdat->lru_lock);
5901 lruvec_init(node_lruvec(pgdat));
5903 for (j = 0; j < MAX_NR_ZONES; j++) {
5904 struct zone *zone = pgdat->node_zones + j;
5905 unsigned long size, realsize, freesize, memmap_pages;
5906 unsigned long zone_start_pfn = zone->zone_start_pfn;
5908 size = zone->spanned_pages;
5909 realsize = freesize = zone->present_pages;
5912 * Adjust freesize so that it accounts for how much memory
5913 * is used by this zone for memmap. This affects the watermark
5914 * and per-cpu initialisations
5916 memmap_pages = calc_memmap_size(size, realsize);
5917 if (!is_highmem_idx(j)) {
5918 if (freesize >= memmap_pages) {
5919 freesize -= memmap_pages;
5922 " %s zone: %lu pages used for memmap\n",
5923 zone_names[j], memmap_pages);
5925 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5926 zone_names[j], memmap_pages, freesize);
5929 /* Account for reserved pages */
5930 if (j == 0 && freesize > dma_reserve) {
5931 freesize -= dma_reserve;
5932 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5933 zone_names[0], dma_reserve);
5936 if (!is_highmem_idx(j))
5937 nr_kernel_pages += freesize;
5938 /* Charge for highmem memmap if there are enough kernel pages */
5939 else if (nr_kernel_pages > memmap_pages * 2)
5940 nr_kernel_pages -= memmap_pages;
5941 nr_all_pages += freesize;
5944 * Set an approximate value for lowmem here, it will be adjusted
5945 * when the bootmem allocator frees pages into the buddy system.
5946 * And all highmem pages will be managed by the buddy system.
5948 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5952 zone->name = zone_names[j];
5953 zone->zone_pgdat = pgdat;
5954 spin_lock_init(&zone->lock);
5955 zone_seqlock_init(zone);
5956 zone_pcp_init(zone);
5961 set_pageblock_order();
5962 setup_usemap(pgdat, zone, zone_start_pfn, size);
5963 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5965 memmap_init(size, nid, j, zone_start_pfn);
5969 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5971 unsigned long __maybe_unused start = 0;
5972 unsigned long __maybe_unused offset = 0;
5974 /* Skip empty nodes */
5975 if (!pgdat->node_spanned_pages)
5978 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5979 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5980 offset = pgdat->node_start_pfn - start;
5981 /* ia64 gets its own node_mem_map, before this, without bootmem */
5982 if (!pgdat->node_mem_map) {
5983 unsigned long size, end;
5987 * The zone's endpoints aren't required to be MAX_ORDER
5988 * aligned but the node_mem_map endpoints must be in order
5989 * for the buddy allocator to function correctly.
5991 end = pgdat_end_pfn(pgdat);
5992 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5993 size = (end - start) * sizeof(struct page);
5994 map = alloc_remap(pgdat->node_id, size);
5996 map = memblock_virt_alloc_node_nopanic(size,
5998 pgdat->node_mem_map = map + offset;
6000 #ifndef CONFIG_NEED_MULTIPLE_NODES
6002 * With no DISCONTIG, the global mem_map is just set as node 0's
6004 if (pgdat == NODE_DATA(0)) {
6005 mem_map = NODE_DATA(0)->node_mem_map;
6006 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6007 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6009 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6012 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6015 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6016 unsigned long node_start_pfn, unsigned long *zholes_size)
6018 pg_data_t *pgdat = NODE_DATA(nid);
6019 unsigned long start_pfn = 0;
6020 unsigned long end_pfn = 0;
6022 /* pg_data_t should be reset to zero when it's allocated */
6023 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6025 reset_deferred_meminit(pgdat);
6026 pgdat->node_id = nid;
6027 pgdat->node_start_pfn = node_start_pfn;
6028 pgdat->per_cpu_nodestats = NULL;
6029 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6030 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6031 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6032 (u64)start_pfn << PAGE_SHIFT,
6033 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6035 start_pfn = node_start_pfn;
6037 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6038 zones_size, zholes_size);
6040 alloc_node_mem_map(pgdat);
6041 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6042 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6043 nid, (unsigned long)pgdat,
6044 (unsigned long)pgdat->node_mem_map);
6047 free_area_init_core(pgdat);
6050 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6052 #if MAX_NUMNODES > 1
6054 * Figure out the number of possible node ids.
6056 void __init setup_nr_node_ids(void)
6058 unsigned int highest;
6060 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6061 nr_node_ids = highest + 1;
6066 * node_map_pfn_alignment - determine the maximum internode alignment
6068 * This function should be called after node map is populated and sorted.
6069 * It calculates the maximum power of two alignment which can distinguish
6072 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6073 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6074 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6075 * shifted, 1GiB is enough and this function will indicate so.
6077 * This is used to test whether pfn -> nid mapping of the chosen memory
6078 * model has fine enough granularity to avoid incorrect mapping for the
6079 * populated node map.
6081 * Returns the determined alignment in pfn's. 0 if there is no alignment
6082 * requirement (single node).
6084 unsigned long __init node_map_pfn_alignment(void)
6086 unsigned long accl_mask = 0, last_end = 0;
6087 unsigned long start, end, mask;
6091 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6092 if (!start || last_nid < 0 || last_nid == nid) {
6099 * Start with a mask granular enough to pin-point to the
6100 * start pfn and tick off bits one-by-one until it becomes
6101 * too coarse to separate the current node from the last.
6103 mask = ~((1 << __ffs(start)) - 1);
6104 while (mask && last_end <= (start & (mask << 1)))
6107 /* accumulate all internode masks */
6111 /* convert mask to number of pages */
6112 return ~accl_mask + 1;
6115 /* Find the lowest pfn for a node */
6116 static unsigned long __init find_min_pfn_for_node(int nid)
6118 unsigned long min_pfn = ULONG_MAX;
6119 unsigned long start_pfn;
6122 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6123 min_pfn = min(min_pfn, start_pfn);
6125 if (min_pfn == ULONG_MAX) {
6126 pr_warn("Could not find start_pfn for node %d\n", nid);
6134 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6136 * It returns the minimum PFN based on information provided via
6137 * memblock_set_node().
6139 unsigned long __init find_min_pfn_with_active_regions(void)
6141 return find_min_pfn_for_node(MAX_NUMNODES);
6145 * early_calculate_totalpages()
6146 * Sum pages in active regions for movable zone.
6147 * Populate N_MEMORY for calculating usable_nodes.
6149 static unsigned long __init early_calculate_totalpages(void)
6151 unsigned long totalpages = 0;
6152 unsigned long start_pfn, end_pfn;
6155 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6156 unsigned long pages = end_pfn - start_pfn;
6158 totalpages += pages;
6160 node_set_state(nid, N_MEMORY);
6166 * Find the PFN the Movable zone begins in each node. Kernel memory
6167 * is spread evenly between nodes as long as the nodes have enough
6168 * memory. When they don't, some nodes will have more kernelcore than
6171 static void __init find_zone_movable_pfns_for_nodes(void)
6174 unsigned long usable_startpfn;
6175 unsigned long kernelcore_node, kernelcore_remaining;
6176 /* save the state before borrow the nodemask */
6177 nodemask_t saved_node_state = node_states[N_MEMORY];
6178 unsigned long totalpages = early_calculate_totalpages();
6179 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6180 struct memblock_region *r;
6182 /* Need to find movable_zone earlier when movable_node is specified. */
6183 find_usable_zone_for_movable();
6186 * If movable_node is specified, ignore kernelcore and movablecore
6189 if (movable_node_is_enabled()) {
6190 for_each_memblock(memory, r) {
6191 if (!memblock_is_hotpluggable(r))
6196 usable_startpfn = PFN_DOWN(r->base);
6197 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6198 min(usable_startpfn, zone_movable_pfn[nid]) :
6206 * If kernelcore=mirror is specified, ignore movablecore option
6208 if (mirrored_kernelcore) {
6209 bool mem_below_4gb_not_mirrored = false;
6211 for_each_memblock(memory, r) {
6212 if (memblock_is_mirror(r))
6217 usable_startpfn = memblock_region_memory_base_pfn(r);
6219 if (usable_startpfn < 0x100000) {
6220 mem_below_4gb_not_mirrored = true;
6224 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6225 min(usable_startpfn, zone_movable_pfn[nid]) :
6229 if (mem_below_4gb_not_mirrored)
6230 pr_warn("This configuration results in unmirrored kernel memory.");
6236 * If movablecore=nn[KMG] was specified, calculate what size of
6237 * kernelcore that corresponds so that memory usable for
6238 * any allocation type is evenly spread. If both kernelcore
6239 * and movablecore are specified, then the value of kernelcore
6240 * will be used for required_kernelcore if it's greater than
6241 * what movablecore would have allowed.
6243 if (required_movablecore) {
6244 unsigned long corepages;
6247 * Round-up so that ZONE_MOVABLE is at least as large as what
6248 * was requested by the user
6250 required_movablecore =
6251 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6252 required_movablecore = min(totalpages, required_movablecore);
6253 corepages = totalpages - required_movablecore;
6255 required_kernelcore = max(required_kernelcore, corepages);
6259 * If kernelcore was not specified or kernelcore size is larger
6260 * than totalpages, there is no ZONE_MOVABLE.
6262 if (!required_kernelcore || required_kernelcore >= totalpages)
6265 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6266 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6269 /* Spread kernelcore memory as evenly as possible throughout nodes */
6270 kernelcore_node = required_kernelcore / usable_nodes;
6271 for_each_node_state(nid, N_MEMORY) {
6272 unsigned long start_pfn, end_pfn;
6275 * Recalculate kernelcore_node if the division per node
6276 * now exceeds what is necessary to satisfy the requested
6277 * amount of memory for the kernel
6279 if (required_kernelcore < kernelcore_node)
6280 kernelcore_node = required_kernelcore / usable_nodes;
6283 * As the map is walked, we track how much memory is usable
6284 * by the kernel using kernelcore_remaining. When it is
6285 * 0, the rest of the node is usable by ZONE_MOVABLE
6287 kernelcore_remaining = kernelcore_node;
6289 /* Go through each range of PFNs within this node */
6290 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6291 unsigned long size_pages;
6293 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6294 if (start_pfn >= end_pfn)
6297 /* Account for what is only usable for kernelcore */
6298 if (start_pfn < usable_startpfn) {
6299 unsigned long kernel_pages;
6300 kernel_pages = min(end_pfn, usable_startpfn)
6303 kernelcore_remaining -= min(kernel_pages,
6304 kernelcore_remaining);
6305 required_kernelcore -= min(kernel_pages,
6306 required_kernelcore);
6308 /* Continue if range is now fully accounted */
6309 if (end_pfn <= usable_startpfn) {
6312 * Push zone_movable_pfn to the end so
6313 * that if we have to rebalance
6314 * kernelcore across nodes, we will
6315 * not double account here
6317 zone_movable_pfn[nid] = end_pfn;
6320 start_pfn = usable_startpfn;
6324 * The usable PFN range for ZONE_MOVABLE is from
6325 * start_pfn->end_pfn. Calculate size_pages as the
6326 * number of pages used as kernelcore
6328 size_pages = end_pfn - start_pfn;
6329 if (size_pages > kernelcore_remaining)
6330 size_pages = kernelcore_remaining;
6331 zone_movable_pfn[nid] = start_pfn + size_pages;
6334 * Some kernelcore has been met, update counts and
6335 * break if the kernelcore for this node has been
6338 required_kernelcore -= min(required_kernelcore,
6340 kernelcore_remaining -= size_pages;
6341 if (!kernelcore_remaining)
6347 * If there is still required_kernelcore, we do another pass with one
6348 * less node in the count. This will push zone_movable_pfn[nid] further
6349 * along on the nodes that still have memory until kernelcore is
6353 if (usable_nodes && required_kernelcore > usable_nodes)
6357 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6358 for (nid = 0; nid < MAX_NUMNODES; nid++)
6359 zone_movable_pfn[nid] =
6360 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6363 /* restore the node_state */
6364 node_states[N_MEMORY] = saved_node_state;
6367 /* Any regular or high memory on that node ? */
6368 static void check_for_memory(pg_data_t *pgdat, int nid)
6370 enum zone_type zone_type;
6372 if (N_MEMORY == N_NORMAL_MEMORY)
6375 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6376 struct zone *zone = &pgdat->node_zones[zone_type];
6377 if (populated_zone(zone)) {
6378 node_set_state(nid, N_HIGH_MEMORY);
6379 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6380 zone_type <= ZONE_NORMAL)
6381 node_set_state(nid, N_NORMAL_MEMORY);
6388 * free_area_init_nodes - Initialise all pg_data_t and zone data
6389 * @max_zone_pfn: an array of max PFNs for each zone
6391 * This will call free_area_init_node() for each active node in the system.
6392 * Using the page ranges provided by memblock_set_node(), the size of each
6393 * zone in each node and their holes is calculated. If the maximum PFN
6394 * between two adjacent zones match, it is assumed that the zone is empty.
6395 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6396 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6397 * starts where the previous one ended. For example, ZONE_DMA32 starts
6398 * at arch_max_dma_pfn.
6400 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6402 unsigned long start_pfn, end_pfn;
6405 /* Record where the zone boundaries are */
6406 memset(arch_zone_lowest_possible_pfn, 0,
6407 sizeof(arch_zone_lowest_possible_pfn));
6408 memset(arch_zone_highest_possible_pfn, 0,
6409 sizeof(arch_zone_highest_possible_pfn));
6411 start_pfn = find_min_pfn_with_active_regions();
6413 for (i = 0; i < MAX_NR_ZONES; i++) {
6414 if (i == ZONE_MOVABLE)
6417 end_pfn = max(max_zone_pfn[i], start_pfn);
6418 arch_zone_lowest_possible_pfn[i] = start_pfn;
6419 arch_zone_highest_possible_pfn[i] = end_pfn;
6421 start_pfn = end_pfn;
6423 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6424 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6426 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6427 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6428 find_zone_movable_pfns_for_nodes();
6430 /* Print out the zone ranges */
6431 pr_info("Zone ranges:\n");
6432 for (i = 0; i < MAX_NR_ZONES; i++) {
6433 if (i == ZONE_MOVABLE)
6435 pr_info(" %-8s ", zone_names[i]);
6436 if (arch_zone_lowest_possible_pfn[i] ==
6437 arch_zone_highest_possible_pfn[i])
6440 pr_cont("[mem %#018Lx-%#018Lx]\n",
6441 (u64)arch_zone_lowest_possible_pfn[i]
6443 ((u64)arch_zone_highest_possible_pfn[i]
6444 << PAGE_SHIFT) - 1);
6447 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6448 pr_info("Movable zone start for each node\n");
6449 for (i = 0; i < MAX_NUMNODES; i++) {
6450 if (zone_movable_pfn[i])
6451 pr_info(" Node %d: %#018Lx\n", i,
6452 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6455 /* Print out the early node map */
6456 pr_info("Early memory node ranges\n");
6457 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6458 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6459 (u64)start_pfn << PAGE_SHIFT,
6460 ((u64)end_pfn << PAGE_SHIFT) - 1);
6462 /* Initialise every node */
6463 mminit_verify_pageflags_layout();
6464 setup_nr_node_ids();
6465 for_each_online_node(nid) {
6466 pg_data_t *pgdat = NODE_DATA(nid);
6467 free_area_init_node(nid, NULL,
6468 find_min_pfn_for_node(nid), NULL);
6470 /* Any memory on that node */
6471 if (pgdat->node_present_pages)
6472 node_set_state(nid, N_MEMORY);
6473 check_for_memory(pgdat, nid);
6477 static int __init cmdline_parse_core(char *p, unsigned long *core)
6479 unsigned long long coremem;
6483 coremem = memparse(p, &p);
6484 *core = coremem >> PAGE_SHIFT;
6486 /* Paranoid check that UL is enough for the coremem value */
6487 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6493 * kernelcore=size sets the amount of memory for use for allocations that
6494 * cannot be reclaimed or migrated.
6496 static int __init cmdline_parse_kernelcore(char *p)
6498 /* parse kernelcore=mirror */
6499 if (parse_option_str(p, "mirror")) {
6500 mirrored_kernelcore = true;
6504 return cmdline_parse_core(p, &required_kernelcore);
6508 * movablecore=size sets the amount of memory for use for allocations that
6509 * can be reclaimed or migrated.
6511 static int __init cmdline_parse_movablecore(char *p)
6513 return cmdline_parse_core(p, &required_movablecore);
6516 early_param("kernelcore", cmdline_parse_kernelcore);
6517 early_param("movablecore", cmdline_parse_movablecore);
6519 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6521 void adjust_managed_page_count(struct page *page, long count)
6523 spin_lock(&managed_page_count_lock);
6524 page_zone(page)->managed_pages += count;
6525 totalram_pages += count;
6526 #ifdef CONFIG_HIGHMEM
6527 if (PageHighMem(page))
6528 totalhigh_pages += count;
6530 spin_unlock(&managed_page_count_lock);
6532 EXPORT_SYMBOL(adjust_managed_page_count);
6534 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6537 unsigned long pages = 0;
6539 start = (void *)PAGE_ALIGN((unsigned long)start);
6540 end = (void *)((unsigned long)end & PAGE_MASK);
6541 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6542 if ((unsigned int)poison <= 0xFF)
6543 memset(pos, poison, PAGE_SIZE);
6544 free_reserved_page(virt_to_page(pos));
6548 pr_info("Freeing %s memory: %ldK\n",
6549 s, pages << (PAGE_SHIFT - 10));
6553 EXPORT_SYMBOL(free_reserved_area);
6555 #ifdef CONFIG_HIGHMEM
6556 void free_highmem_page(struct page *page)
6558 __free_reserved_page(page);
6560 page_zone(page)->managed_pages++;
6566 void __init mem_init_print_info(const char *str)
6568 unsigned long physpages, codesize, datasize, rosize, bss_size;
6569 unsigned long init_code_size, init_data_size;
6571 physpages = get_num_physpages();
6572 codesize = _etext - _stext;
6573 datasize = _edata - _sdata;
6574 rosize = __end_rodata - __start_rodata;
6575 bss_size = __bss_stop - __bss_start;
6576 init_data_size = __init_end - __init_begin;
6577 init_code_size = _einittext - _sinittext;
6580 * Detect special cases and adjust section sizes accordingly:
6581 * 1) .init.* may be embedded into .data sections
6582 * 2) .init.text.* may be out of [__init_begin, __init_end],
6583 * please refer to arch/tile/kernel/vmlinux.lds.S.
6584 * 3) .rodata.* may be embedded into .text or .data sections.
6586 #define adj_init_size(start, end, size, pos, adj) \
6588 if (start <= pos && pos < end && size > adj) \
6592 adj_init_size(__init_begin, __init_end, init_data_size,
6593 _sinittext, init_code_size);
6594 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6595 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6596 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6597 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6599 #undef adj_init_size
6601 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6602 #ifdef CONFIG_HIGHMEM
6606 nr_free_pages() << (PAGE_SHIFT - 10),
6607 physpages << (PAGE_SHIFT - 10),
6608 codesize >> 10, datasize >> 10, rosize >> 10,
6609 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6610 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6611 totalcma_pages << (PAGE_SHIFT - 10),
6612 #ifdef CONFIG_HIGHMEM
6613 totalhigh_pages << (PAGE_SHIFT - 10),
6615 str ? ", " : "", str ? str : "");
6619 * set_dma_reserve - set the specified number of pages reserved in the first zone
6620 * @new_dma_reserve: The number of pages to mark reserved
6622 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6623 * In the DMA zone, a significant percentage may be consumed by kernel image
6624 * and other unfreeable allocations which can skew the watermarks badly. This
6625 * function may optionally be used to account for unfreeable pages in the
6626 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6627 * smaller per-cpu batchsize.
6629 void __init set_dma_reserve(unsigned long new_dma_reserve)
6631 dma_reserve = new_dma_reserve;
6634 void __init free_area_init(unsigned long *zones_size)
6636 free_area_init_node(0, zones_size,
6637 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6640 static int page_alloc_cpu_dead(unsigned int cpu)
6643 lru_add_drain_cpu(cpu);
6647 * Spill the event counters of the dead processor
6648 * into the current processors event counters.
6649 * This artificially elevates the count of the current
6652 vm_events_fold_cpu(cpu);
6655 * Zero the differential counters of the dead processor
6656 * so that the vm statistics are consistent.
6658 * This is only okay since the processor is dead and cannot
6659 * race with what we are doing.
6661 cpu_vm_stats_fold(cpu);
6665 void __init page_alloc_init(void)
6669 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6670 "mm/page_alloc:dead", NULL,
6671 page_alloc_cpu_dead);
6676 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6677 * or min_free_kbytes changes.
6679 static void calculate_totalreserve_pages(void)
6681 struct pglist_data *pgdat;
6682 unsigned long reserve_pages = 0;
6683 enum zone_type i, j;
6685 for_each_online_pgdat(pgdat) {
6687 pgdat->totalreserve_pages = 0;
6689 for (i = 0; i < MAX_NR_ZONES; i++) {
6690 struct zone *zone = pgdat->node_zones + i;
6693 /* Find valid and maximum lowmem_reserve in the zone */
6694 for (j = i; j < MAX_NR_ZONES; j++) {
6695 if (zone->lowmem_reserve[j] > max)
6696 max = zone->lowmem_reserve[j];
6699 /* we treat the high watermark as reserved pages. */
6700 max += high_wmark_pages(zone);
6702 if (max > zone->managed_pages)
6703 max = zone->managed_pages;
6705 pgdat->totalreserve_pages += max;
6707 reserve_pages += max;
6710 totalreserve_pages = reserve_pages;
6714 * setup_per_zone_lowmem_reserve - called whenever
6715 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6716 * has a correct pages reserved value, so an adequate number of
6717 * pages are left in the zone after a successful __alloc_pages().
6719 static void setup_per_zone_lowmem_reserve(void)
6721 struct pglist_data *pgdat;
6722 enum zone_type j, idx;
6724 for_each_online_pgdat(pgdat) {
6725 for (j = 0; j < MAX_NR_ZONES; j++) {
6726 struct zone *zone = pgdat->node_zones + j;
6727 unsigned long managed_pages = zone->managed_pages;
6729 zone->lowmem_reserve[j] = 0;
6733 struct zone *lower_zone;
6737 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6738 sysctl_lowmem_reserve_ratio[idx] = 1;
6740 lower_zone = pgdat->node_zones + idx;
6741 lower_zone->lowmem_reserve[j] = managed_pages /
6742 sysctl_lowmem_reserve_ratio[idx];
6743 managed_pages += lower_zone->managed_pages;
6748 /* update totalreserve_pages */
6749 calculate_totalreserve_pages();
6752 static void __setup_per_zone_wmarks(void)
6754 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6755 unsigned long lowmem_pages = 0;
6757 unsigned long flags;
6759 /* Calculate total number of !ZONE_HIGHMEM pages */
6760 for_each_zone(zone) {
6761 if (!is_highmem(zone))
6762 lowmem_pages += zone->managed_pages;
6765 for_each_zone(zone) {
6768 spin_lock_irqsave(&zone->lock, flags);
6769 tmp = (u64)pages_min * zone->managed_pages;
6770 do_div(tmp, lowmem_pages);
6771 if (is_highmem(zone)) {
6773 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6774 * need highmem pages, so cap pages_min to a small
6777 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6778 * deltas control asynch page reclaim, and so should
6779 * not be capped for highmem.
6781 unsigned long min_pages;
6783 min_pages = zone->managed_pages / 1024;
6784 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6785 zone->watermark[WMARK_MIN] = min_pages;
6788 * If it's a lowmem zone, reserve a number of pages
6789 * proportionate to the zone's size.
6791 zone->watermark[WMARK_MIN] = tmp;
6795 * Set the kswapd watermarks distance according to the
6796 * scale factor in proportion to available memory, but
6797 * ensure a minimum size on small systems.
6799 tmp = max_t(u64, tmp >> 2,
6800 mult_frac(zone->managed_pages,
6801 watermark_scale_factor, 10000));
6803 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6804 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6806 spin_unlock_irqrestore(&zone->lock, flags);
6809 /* update totalreserve_pages */
6810 calculate_totalreserve_pages();
6814 * setup_per_zone_wmarks - called when min_free_kbytes changes
6815 * or when memory is hot-{added|removed}
6817 * Ensures that the watermark[min,low,high] values for each zone are set
6818 * correctly with respect to min_free_kbytes.
6820 void setup_per_zone_wmarks(void)
6822 mutex_lock(&zonelists_mutex);
6823 __setup_per_zone_wmarks();
6824 mutex_unlock(&zonelists_mutex);
6828 * Initialise min_free_kbytes.
6830 * For small machines we want it small (128k min). For large machines
6831 * we want it large (64MB max). But it is not linear, because network
6832 * bandwidth does not increase linearly with machine size. We use
6834 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6835 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6851 int __meminit init_per_zone_wmark_min(void)
6853 unsigned long lowmem_kbytes;
6854 int new_min_free_kbytes;
6856 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6857 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6859 if (new_min_free_kbytes > user_min_free_kbytes) {
6860 min_free_kbytes = new_min_free_kbytes;
6861 if (min_free_kbytes < 128)
6862 min_free_kbytes = 128;
6863 if (min_free_kbytes > 65536)
6864 min_free_kbytes = 65536;
6866 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6867 new_min_free_kbytes, user_min_free_kbytes);
6869 setup_per_zone_wmarks();
6870 refresh_zone_stat_thresholds();
6871 setup_per_zone_lowmem_reserve();
6874 setup_min_unmapped_ratio();
6875 setup_min_slab_ratio();
6880 core_initcall(init_per_zone_wmark_min)
6883 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6884 * that we can call two helper functions whenever min_free_kbytes
6887 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6888 void __user *buffer, size_t *length, loff_t *ppos)
6892 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6897 user_min_free_kbytes = min_free_kbytes;
6898 setup_per_zone_wmarks();
6903 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6904 void __user *buffer, size_t *length, loff_t *ppos)
6908 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6913 setup_per_zone_wmarks();
6919 static void setup_min_unmapped_ratio(void)
6924 for_each_online_pgdat(pgdat)
6925 pgdat->min_unmapped_pages = 0;
6928 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6929 sysctl_min_unmapped_ratio) / 100;
6933 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6934 void __user *buffer, size_t *length, loff_t *ppos)
6938 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6942 setup_min_unmapped_ratio();
6947 static void setup_min_slab_ratio(void)
6952 for_each_online_pgdat(pgdat)
6953 pgdat->min_slab_pages = 0;
6956 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6957 sysctl_min_slab_ratio) / 100;
6960 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6961 void __user *buffer, size_t *length, loff_t *ppos)
6965 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6969 setup_min_slab_ratio();
6976 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6977 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6978 * whenever sysctl_lowmem_reserve_ratio changes.
6980 * The reserve ratio obviously has absolutely no relation with the
6981 * minimum watermarks. The lowmem reserve ratio can only make sense
6982 * if in function of the boot time zone sizes.
6984 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6985 void __user *buffer, size_t *length, loff_t *ppos)
6987 proc_dointvec_minmax(table, write, buffer, length, ppos);
6988 setup_per_zone_lowmem_reserve();
6993 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6994 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6995 * pagelist can have before it gets flushed back to buddy allocator.
6997 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6998 void __user *buffer, size_t *length, loff_t *ppos)
7001 int old_percpu_pagelist_fraction;
7004 mutex_lock(&pcp_batch_high_lock);
7005 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7007 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7008 if (!write || ret < 0)
7011 /* Sanity checking to avoid pcp imbalance */
7012 if (percpu_pagelist_fraction &&
7013 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7014 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7020 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7023 for_each_populated_zone(zone) {
7026 for_each_possible_cpu(cpu)
7027 pageset_set_high_and_batch(zone,
7028 per_cpu_ptr(zone->pageset, cpu));
7031 mutex_unlock(&pcp_batch_high_lock);
7036 int hashdist = HASHDIST_DEFAULT;
7038 static int __init set_hashdist(char *str)
7042 hashdist = simple_strtoul(str, &str, 0);
7045 __setup("hashdist=", set_hashdist);
7048 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7050 * Returns the number of pages that arch has reserved but
7051 * is not known to alloc_large_system_hash().
7053 static unsigned long __init arch_reserved_kernel_pages(void)
7060 * allocate a large system hash table from bootmem
7061 * - it is assumed that the hash table must contain an exact power-of-2
7062 * quantity of entries
7063 * - limit is the number of hash buckets, not the total allocation size
7065 void *__init alloc_large_system_hash(const char *tablename,
7066 unsigned long bucketsize,
7067 unsigned long numentries,
7070 unsigned int *_hash_shift,
7071 unsigned int *_hash_mask,
7072 unsigned long low_limit,
7073 unsigned long high_limit)
7075 unsigned long long max = high_limit;
7076 unsigned long log2qty, size;
7079 /* allow the kernel cmdline to have a say */
7081 /* round applicable memory size up to nearest megabyte */
7082 numentries = nr_kernel_pages;
7083 numentries -= arch_reserved_kernel_pages();
7085 /* It isn't necessary when PAGE_SIZE >= 1MB */
7086 if (PAGE_SHIFT < 20)
7087 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7089 /* limit to 1 bucket per 2^scale bytes of low memory */
7090 if (scale > PAGE_SHIFT)
7091 numentries >>= (scale - PAGE_SHIFT);
7093 numentries <<= (PAGE_SHIFT - scale);
7095 /* Make sure we've got at least a 0-order allocation.. */
7096 if (unlikely(flags & HASH_SMALL)) {
7097 /* Makes no sense without HASH_EARLY */
7098 WARN_ON(!(flags & HASH_EARLY));
7099 if (!(numentries >> *_hash_shift)) {
7100 numentries = 1UL << *_hash_shift;
7101 BUG_ON(!numentries);
7103 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7104 numentries = PAGE_SIZE / bucketsize;
7106 numentries = roundup_pow_of_two(numentries);
7108 /* limit allocation size to 1/16 total memory by default */
7110 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7111 do_div(max, bucketsize);
7113 max = min(max, 0x80000000ULL);
7115 if (numentries < low_limit)
7116 numentries = low_limit;
7117 if (numentries > max)
7120 log2qty = ilog2(numentries);
7123 size = bucketsize << log2qty;
7124 if (flags & HASH_EARLY)
7125 table = memblock_virt_alloc_nopanic(size, 0);
7127 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7130 * If bucketsize is not a power-of-two, we may free
7131 * some pages at the end of hash table which
7132 * alloc_pages_exact() automatically does
7134 if (get_order(size) < MAX_ORDER) {
7135 table = alloc_pages_exact(size, GFP_ATOMIC);
7136 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7139 } while (!table && size > PAGE_SIZE && --log2qty);
7142 panic("Failed to allocate %s hash table\n", tablename);
7144 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7145 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7148 *_hash_shift = log2qty;
7150 *_hash_mask = (1 << log2qty) - 1;
7156 * This function checks whether pageblock includes unmovable pages or not.
7157 * If @count is not zero, it is okay to include less @count unmovable pages
7159 * PageLRU check without isolation or lru_lock could race so that
7160 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7161 * expect this function should be exact.
7163 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7164 bool skip_hwpoisoned_pages)
7166 unsigned long pfn, iter, found;
7170 * For avoiding noise data, lru_add_drain_all() should be called
7171 * If ZONE_MOVABLE, the zone never contains unmovable pages
7173 if (zone_idx(zone) == ZONE_MOVABLE)
7175 mt = get_pageblock_migratetype(page);
7176 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7179 pfn = page_to_pfn(page);
7180 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7181 unsigned long check = pfn + iter;
7183 if (!pfn_valid_within(check))
7186 page = pfn_to_page(check);
7189 * Hugepages are not in LRU lists, but they're movable.
7190 * We need not scan over tail pages bacause we don't
7191 * handle each tail page individually in migration.
7193 if (PageHuge(page)) {
7194 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7199 * We can't use page_count without pin a page
7200 * because another CPU can free compound page.
7201 * This check already skips compound tails of THP
7202 * because their page->_refcount is zero at all time.
7204 if (!page_ref_count(page)) {
7205 if (PageBuddy(page))
7206 iter += (1 << page_order(page)) - 1;
7211 * The HWPoisoned page may be not in buddy system, and
7212 * page_count() is not 0.
7214 if (skip_hwpoisoned_pages && PageHWPoison(page))
7220 * If there are RECLAIMABLE pages, we need to check
7221 * it. But now, memory offline itself doesn't call
7222 * shrink_node_slabs() and it still to be fixed.
7225 * If the page is not RAM, page_count()should be 0.
7226 * we don't need more check. This is an _used_ not-movable page.
7228 * The problematic thing here is PG_reserved pages. PG_reserved
7229 * is set to both of a memory hole page and a _used_ kernel
7238 bool is_pageblock_removable_nolock(struct page *page)
7244 * We have to be careful here because we are iterating over memory
7245 * sections which are not zone aware so we might end up outside of
7246 * the zone but still within the section.
7247 * We have to take care about the node as well. If the node is offline
7248 * its NODE_DATA will be NULL - see page_zone.
7250 if (!node_online(page_to_nid(page)))
7253 zone = page_zone(page);
7254 pfn = page_to_pfn(page);
7255 if (!zone_spans_pfn(zone, pfn))
7258 return !has_unmovable_pages(zone, page, 0, true);
7261 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7263 static unsigned long pfn_max_align_down(unsigned long pfn)
7265 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7266 pageblock_nr_pages) - 1);
7269 static unsigned long pfn_max_align_up(unsigned long pfn)
7271 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7272 pageblock_nr_pages));
7275 /* [start, end) must belong to a single zone. */
7276 static int __alloc_contig_migrate_range(struct compact_control *cc,
7277 unsigned long start, unsigned long end)
7279 /* This function is based on compact_zone() from compaction.c. */
7280 unsigned long nr_reclaimed;
7281 unsigned long pfn = start;
7282 unsigned int tries = 0;
7287 while (pfn < end || !list_empty(&cc->migratepages)) {
7288 if (fatal_signal_pending(current)) {
7293 if (list_empty(&cc->migratepages)) {
7294 cc->nr_migratepages = 0;
7295 pfn = isolate_migratepages_range(cc, pfn, end);
7301 } else if (++tries == 5) {
7302 ret = ret < 0 ? ret : -EBUSY;
7306 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7308 cc->nr_migratepages -= nr_reclaimed;
7310 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7311 NULL, 0, cc->mode, MR_CMA);
7314 putback_movable_pages(&cc->migratepages);
7321 * alloc_contig_range() -- tries to allocate given range of pages
7322 * @start: start PFN to allocate
7323 * @end: one-past-the-last PFN to allocate
7324 * @migratetype: migratetype of the underlaying pageblocks (either
7325 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7326 * in range must have the same migratetype and it must
7327 * be either of the two.
7329 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7330 * aligned, however it's the caller's responsibility to guarantee that
7331 * we are the only thread that changes migrate type of pageblocks the
7334 * The PFN range must belong to a single zone.
7336 * Returns zero on success or negative error code. On success all
7337 * pages which PFN is in [start, end) are allocated for the caller and
7338 * need to be freed with free_contig_range().
7340 int alloc_contig_range(unsigned long start, unsigned long end,
7341 unsigned migratetype)
7343 unsigned long outer_start, outer_end;
7347 struct compact_control cc = {
7348 .nr_migratepages = 0,
7350 .zone = page_zone(pfn_to_page(start)),
7351 .mode = MIGRATE_SYNC,
7352 .ignore_skip_hint = true,
7353 .gfp_mask = GFP_KERNEL,
7355 INIT_LIST_HEAD(&cc.migratepages);
7358 * What we do here is we mark all pageblocks in range as
7359 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7360 * have different sizes, and due to the way page allocator
7361 * work, we align the range to biggest of the two pages so
7362 * that page allocator won't try to merge buddies from
7363 * different pageblocks and change MIGRATE_ISOLATE to some
7364 * other migration type.
7366 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7367 * migrate the pages from an unaligned range (ie. pages that
7368 * we are interested in). This will put all the pages in
7369 * range back to page allocator as MIGRATE_ISOLATE.
7371 * When this is done, we take the pages in range from page
7372 * allocator removing them from the buddy system. This way
7373 * page allocator will never consider using them.
7375 * This lets us mark the pageblocks back as
7376 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7377 * aligned range but not in the unaligned, original range are
7378 * put back to page allocator so that buddy can use them.
7381 ret = start_isolate_page_range(pfn_max_align_down(start),
7382 pfn_max_align_up(end), migratetype,
7388 * In case of -EBUSY, we'd like to know which page causes problem.
7389 * So, just fall through. We will check it in test_pages_isolated().
7391 ret = __alloc_contig_migrate_range(&cc, start, end);
7392 if (ret && ret != -EBUSY)
7396 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7397 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7398 * more, all pages in [start, end) are free in page allocator.
7399 * What we are going to do is to allocate all pages from
7400 * [start, end) (that is remove them from page allocator).
7402 * The only problem is that pages at the beginning and at the
7403 * end of interesting range may be not aligned with pages that
7404 * page allocator holds, ie. they can be part of higher order
7405 * pages. Because of this, we reserve the bigger range and
7406 * once this is done free the pages we are not interested in.
7408 * We don't have to hold zone->lock here because the pages are
7409 * isolated thus they won't get removed from buddy.
7412 lru_add_drain_all();
7413 drain_all_pages(cc.zone);
7416 outer_start = start;
7417 while (!PageBuddy(pfn_to_page(outer_start))) {
7418 if (++order >= MAX_ORDER) {
7419 outer_start = start;
7422 outer_start &= ~0UL << order;
7425 if (outer_start != start) {
7426 order = page_order(pfn_to_page(outer_start));
7429 * outer_start page could be small order buddy page and
7430 * it doesn't include start page. Adjust outer_start
7431 * in this case to report failed page properly
7432 * on tracepoint in test_pages_isolated()
7434 if (outer_start + (1UL << order) <= start)
7435 outer_start = start;
7438 /* Make sure the range is really isolated. */
7439 if (test_pages_isolated(outer_start, end, false)) {
7440 pr_info("%s: [%lx, %lx) PFNs busy\n",
7441 __func__, outer_start, end);
7446 /* Grab isolated pages from freelists. */
7447 outer_end = isolate_freepages_range(&cc, outer_start, end);
7453 /* Free head and tail (if any) */
7454 if (start != outer_start)
7455 free_contig_range(outer_start, start - outer_start);
7456 if (end != outer_end)
7457 free_contig_range(end, outer_end - end);
7460 undo_isolate_page_range(pfn_max_align_down(start),
7461 pfn_max_align_up(end), migratetype);
7465 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7467 unsigned int count = 0;
7469 for (; nr_pages--; pfn++) {
7470 struct page *page = pfn_to_page(pfn);
7472 count += page_count(page) != 1;
7475 WARN(count != 0, "%d pages are still in use!\n", count);
7479 #ifdef CONFIG_MEMORY_HOTPLUG
7481 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7482 * page high values need to be recalulated.
7484 void __meminit zone_pcp_update(struct zone *zone)
7487 mutex_lock(&pcp_batch_high_lock);
7488 for_each_possible_cpu(cpu)
7489 pageset_set_high_and_batch(zone,
7490 per_cpu_ptr(zone->pageset, cpu));
7491 mutex_unlock(&pcp_batch_high_lock);
7495 void zone_pcp_reset(struct zone *zone)
7497 unsigned long flags;
7499 struct per_cpu_pageset *pset;
7501 /* avoid races with drain_pages() */
7502 local_irq_save(flags);
7503 if (zone->pageset != &boot_pageset) {
7504 for_each_online_cpu(cpu) {
7505 pset = per_cpu_ptr(zone->pageset, cpu);
7506 drain_zonestat(zone, pset);
7508 free_percpu(zone->pageset);
7509 zone->pageset = &boot_pageset;
7511 local_irq_restore(flags);
7514 #ifdef CONFIG_MEMORY_HOTREMOVE
7516 * All pages in the range must be in a single zone and isolated
7517 * before calling this.
7520 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7524 unsigned int order, i;
7526 unsigned long flags;
7527 /* find the first valid pfn */
7528 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7533 zone = page_zone(pfn_to_page(pfn));
7534 spin_lock_irqsave(&zone->lock, flags);
7536 while (pfn < end_pfn) {
7537 if (!pfn_valid(pfn)) {
7541 page = pfn_to_page(pfn);
7543 * The HWPoisoned page may be not in buddy system, and
7544 * page_count() is not 0.
7546 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7548 SetPageReserved(page);
7552 BUG_ON(page_count(page));
7553 BUG_ON(!PageBuddy(page));
7554 order = page_order(page);
7555 #ifdef CONFIG_DEBUG_VM
7556 pr_info("remove from free list %lx %d %lx\n",
7557 pfn, 1 << order, end_pfn);
7559 list_del(&page->lru);
7560 rmv_page_order(page);
7561 zone->free_area[order].nr_free--;
7562 for (i = 0; i < (1 << order); i++)
7563 SetPageReserved((page+i));
7564 pfn += (1 << order);
7566 spin_unlock_irqrestore(&zone->lock, flags);
7570 bool is_free_buddy_page(struct page *page)
7572 struct zone *zone = page_zone(page);
7573 unsigned long pfn = page_to_pfn(page);
7574 unsigned long flags;
7577 spin_lock_irqsave(&zone->lock, flags);
7578 for (order = 0; order < MAX_ORDER; order++) {
7579 struct page *page_head = page - (pfn & ((1 << order) - 1));
7581 if (PageBuddy(page_head) && page_order(page_head) >= order)
7584 spin_unlock_irqrestore(&zone->lock, flags);
7586 return order < MAX_ORDER;