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/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 /* Protect totalram_pages and zone->managed_pages */
126 static DEFINE_SPINLOCK(managed_page_count_lock);
128 unsigned long totalram_pages __read_mostly;
129 unsigned long totalreserve_pages __read_mostly;
130 unsigned long totalcma_pages __read_mostly;
132 int percpu_pagelist_fraction;
133 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
143 static inline int get_pcppage_migratetype(struct page *page)
148 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
150 page->index = migratetype;
153 #ifdef CONFIG_PM_SLEEP
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with pm_mutex held (gfp_allowed_mask also should
159 * only be modified with pm_mutex held, unless the suspend/hibernate code is
160 * guaranteed not to run in parallel with that modification).
163 static gfp_t saved_gfp_mask;
165 void pm_restore_gfp_mask(void)
167 WARN_ON(!mutex_is_locked(&pm_mutex));
168 if (saved_gfp_mask) {
169 gfp_allowed_mask = saved_gfp_mask;
174 void pm_restrict_gfp_mask(void)
176 WARN_ON(!mutex_is_locked(&pm_mutex));
177 WARN_ON(saved_gfp_mask);
178 saved_gfp_mask = gfp_allowed_mask;
179 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
182 bool pm_suspended_storage(void)
184 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
188 #endif /* CONFIG_PM_SLEEP */
190 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191 unsigned int pageblock_order __read_mostly;
194 static void __free_pages_ok(struct page *page, unsigned int order);
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
207 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
208 #ifdef CONFIG_ZONE_DMA
211 #ifdef CONFIG_ZONE_DMA32
214 #ifdef CONFIG_HIGHMEM
220 EXPORT_SYMBOL(totalram_pages);
222 static char * const zone_names[MAX_NR_ZONES] = {
223 #ifdef CONFIG_ZONE_DMA
226 #ifdef CONFIG_ZONE_DMA32
230 #ifdef CONFIG_HIGHMEM
234 #ifdef CONFIG_ZONE_DEVICE
239 char * const migratetype_names[MIGRATE_TYPES] = {
247 #ifdef CONFIG_MEMORY_ISOLATION
252 compound_page_dtor * const compound_page_dtors[] = {
255 #ifdef CONFIG_HUGETLB_PAGE
258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
263 int min_free_kbytes = 1024;
264 int user_min_free_kbytes = -1;
265 int watermark_scale_factor = 10;
267 static unsigned long __meminitdata nr_kernel_pages;
268 static unsigned long __meminitdata nr_all_pages;
269 static unsigned long __meminitdata dma_reserve;
271 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
272 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
273 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
274 static unsigned long __initdata required_kernelcore;
275 static unsigned long __initdata required_movablecore;
276 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
277 static bool mirrored_kernelcore;
279 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
281 EXPORT_SYMBOL(movable_zone);
282 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
285 int nr_node_ids __read_mostly = MAX_NUMNODES;
286 int nr_online_nodes __read_mostly = 1;
287 EXPORT_SYMBOL(nr_node_ids);
288 EXPORT_SYMBOL(nr_online_nodes);
291 int page_group_by_mobility_disabled __read_mostly;
293 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
296 * Determine how many pages need to be initialized durig early boot
297 * (non-deferred initialization).
298 * The value of first_deferred_pfn will be set later, once non-deferred pages
299 * are initialized, but for now set it ULONG_MAX.
301 static inline void reset_deferred_meminit(pg_data_t *pgdat)
303 phys_addr_t start_addr, end_addr;
304 unsigned long max_pgcnt;
305 unsigned long reserved;
308 * Initialise at least 2G of a node but also take into account that
309 * two large system hashes that can take up 1GB for 0.25TB/node.
311 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
312 (pgdat->node_spanned_pages >> 8));
315 * Compensate the all the memblock reservations (e.g. crash kernel)
316 * from the initial estimation to make sure we will initialize enough
319 start_addr = PFN_PHYS(pgdat->node_start_pfn);
320 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
321 reserved = memblock_reserved_memory_within(start_addr, end_addr);
322 max_pgcnt += PHYS_PFN(reserved);
324 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
325 pgdat->first_deferred_pfn = ULONG_MAX;
328 /* Returns true if the struct page for the pfn is uninitialised */
329 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
331 int nid = early_pfn_to_nid(pfn);
333 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
340 * Returns false when the remaining initialisation should be deferred until
341 * later in the boot cycle when it can be parallelised.
343 static inline bool update_defer_init(pg_data_t *pgdat,
344 unsigned long pfn, unsigned long zone_end,
345 unsigned long *nr_initialised)
347 /* Always populate low zones for address-contrained allocations */
348 if (zone_end < pgdat_end_pfn(pgdat))
351 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
352 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
353 pgdat->first_deferred_pfn = pfn;
360 static inline void reset_deferred_meminit(pg_data_t *pgdat)
364 static inline bool early_page_uninitialised(unsigned long pfn)
369 static inline bool update_defer_init(pg_data_t *pgdat,
370 unsigned long pfn, unsigned long zone_end,
371 unsigned long *nr_initialised)
377 /* Return a pointer to the bitmap storing bits affecting a block of pages */
378 static inline unsigned long *get_pageblock_bitmap(struct page *page,
381 #ifdef CONFIG_SPARSEMEM
382 return __pfn_to_section(pfn)->pageblock_flags;
384 return page_zone(page)->pageblock_flags;
385 #endif /* CONFIG_SPARSEMEM */
388 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
390 #ifdef CONFIG_SPARSEMEM
391 pfn &= (PAGES_PER_SECTION-1);
392 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
394 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
395 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
396 #endif /* CONFIG_SPARSEMEM */
400 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
401 * @page: The page within the block of interest
402 * @pfn: The target page frame number
403 * @end_bitidx: The last bit of interest to retrieve
404 * @mask: mask of bits that the caller is interested in
406 * Return: pageblock_bits flags
408 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
410 unsigned long end_bitidx,
413 unsigned long *bitmap;
414 unsigned long bitidx, word_bitidx;
417 bitmap = get_pageblock_bitmap(page, pfn);
418 bitidx = pfn_to_bitidx(page, pfn);
419 word_bitidx = bitidx / BITS_PER_LONG;
420 bitidx &= (BITS_PER_LONG-1);
422 word = bitmap[word_bitidx];
423 bitidx += end_bitidx;
424 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
427 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
428 unsigned long end_bitidx,
431 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
434 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
436 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
440 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
441 * @page: The page within the block of interest
442 * @flags: The flags to set
443 * @pfn: The target page frame number
444 * @end_bitidx: The last bit of interest
445 * @mask: mask of bits that the caller is interested in
447 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
449 unsigned long end_bitidx,
452 unsigned long *bitmap;
453 unsigned long bitidx, word_bitidx;
454 unsigned long old_word, word;
456 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
458 bitmap = get_pageblock_bitmap(page, pfn);
459 bitidx = pfn_to_bitidx(page, pfn);
460 word_bitidx = bitidx / BITS_PER_LONG;
461 bitidx &= (BITS_PER_LONG-1);
463 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
465 bitidx += end_bitidx;
466 mask <<= (BITS_PER_LONG - bitidx - 1);
467 flags <<= (BITS_PER_LONG - bitidx - 1);
469 word = READ_ONCE(bitmap[word_bitidx]);
471 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
472 if (word == old_word)
478 void set_pageblock_migratetype(struct page *page, int migratetype)
480 if (unlikely(page_group_by_mobility_disabled &&
481 migratetype < MIGRATE_PCPTYPES))
482 migratetype = MIGRATE_UNMOVABLE;
484 set_pageblock_flags_group(page, (unsigned long)migratetype,
485 PB_migrate, PB_migrate_end);
488 #ifdef CONFIG_DEBUG_VM
489 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
493 unsigned long pfn = page_to_pfn(page);
494 unsigned long sp, start_pfn;
497 seq = zone_span_seqbegin(zone);
498 start_pfn = zone->zone_start_pfn;
499 sp = zone->spanned_pages;
500 if (!zone_spans_pfn(zone, pfn))
502 } while (zone_span_seqretry(zone, seq));
505 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
506 pfn, zone_to_nid(zone), zone->name,
507 start_pfn, start_pfn + sp);
512 static int page_is_consistent(struct zone *zone, struct page *page)
514 if (!pfn_valid_within(page_to_pfn(page)))
516 if (zone != page_zone(page))
522 * Temporary debugging check for pages not lying within a given zone.
524 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
526 if (page_outside_zone_boundaries(zone, page))
528 if (!page_is_consistent(zone, page))
534 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
540 static void bad_page(struct page *page, const char *reason,
541 unsigned long bad_flags)
543 static unsigned long resume;
544 static unsigned long nr_shown;
545 static unsigned long nr_unshown;
548 * Allow a burst of 60 reports, then keep quiet for that minute;
549 * or allow a steady drip of one report per second.
551 if (nr_shown == 60) {
552 if (time_before(jiffies, resume)) {
558 "BUG: Bad page state: %lu messages suppressed\n",
565 resume = jiffies + 60 * HZ;
567 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
568 current->comm, page_to_pfn(page));
569 __dump_page(page, reason);
570 bad_flags &= page->flags;
572 pr_alert("bad because of flags: %#lx(%pGp)\n",
573 bad_flags, &bad_flags);
574 dump_page_owner(page);
579 /* Leave bad fields for debug, except PageBuddy could make trouble */
580 page_mapcount_reset(page); /* remove PageBuddy */
581 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
585 * Higher-order pages are called "compound pages". They are structured thusly:
587 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
589 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
590 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
592 * The first tail page's ->compound_dtor holds the offset in array of compound
593 * page destructors. See compound_page_dtors.
595 * The first tail page's ->compound_order holds the order of allocation.
596 * This usage means that zero-order pages may not be compound.
599 void free_compound_page(struct page *page)
601 __free_pages_ok(page, compound_order(page));
604 void prep_compound_page(struct page *page, unsigned int order)
607 int nr_pages = 1 << order;
609 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
610 set_compound_order(page, order);
612 for (i = 1; i < nr_pages; i++) {
613 struct page *p = page + i;
614 set_page_count(p, 0);
615 p->mapping = TAIL_MAPPING;
616 set_compound_head(p, page);
618 atomic_set(compound_mapcount_ptr(page), -1);
621 #ifdef CONFIG_DEBUG_PAGEALLOC
622 unsigned int _debug_guardpage_minorder;
623 bool _debug_pagealloc_enabled __read_mostly
624 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
625 EXPORT_SYMBOL(_debug_pagealloc_enabled);
626 bool _debug_guardpage_enabled __read_mostly;
628 static int __init early_debug_pagealloc(char *buf)
632 return kstrtobool(buf, &_debug_pagealloc_enabled);
634 early_param("debug_pagealloc", early_debug_pagealloc);
636 static bool need_debug_guardpage(void)
638 /* If we don't use debug_pagealloc, we don't need guard page */
639 if (!debug_pagealloc_enabled())
642 if (!debug_guardpage_minorder())
648 static void init_debug_guardpage(void)
650 if (!debug_pagealloc_enabled())
653 if (!debug_guardpage_minorder())
656 _debug_guardpage_enabled = true;
659 struct page_ext_operations debug_guardpage_ops = {
660 .need = need_debug_guardpage,
661 .init = init_debug_guardpage,
664 static int __init debug_guardpage_minorder_setup(char *buf)
668 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
669 pr_err("Bad debug_guardpage_minorder value\n");
672 _debug_guardpage_minorder = res;
673 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
676 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
678 static inline bool set_page_guard(struct zone *zone, struct page *page,
679 unsigned int order, int migratetype)
681 struct page_ext *page_ext;
683 if (!debug_guardpage_enabled())
686 if (order >= debug_guardpage_minorder())
689 page_ext = lookup_page_ext(page);
690 if (unlikely(!page_ext))
693 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
695 INIT_LIST_HEAD(&page->lru);
696 set_page_private(page, order);
697 /* Guard pages are not available for any usage */
698 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
703 static inline void clear_page_guard(struct zone *zone, struct page *page,
704 unsigned int order, int migratetype)
706 struct page_ext *page_ext;
708 if (!debug_guardpage_enabled())
711 page_ext = lookup_page_ext(page);
712 if (unlikely(!page_ext))
715 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
717 set_page_private(page, 0);
718 if (!is_migrate_isolate(migratetype))
719 __mod_zone_freepage_state(zone, (1 << order), migratetype);
722 struct page_ext_operations debug_guardpage_ops;
723 static inline bool set_page_guard(struct zone *zone, struct page *page,
724 unsigned int order, int migratetype) { return false; }
725 static inline void clear_page_guard(struct zone *zone, struct page *page,
726 unsigned int order, int migratetype) {}
729 static inline void set_page_order(struct page *page, unsigned int order)
731 set_page_private(page, order);
732 __SetPageBuddy(page);
735 static inline void rmv_page_order(struct page *page)
737 __ClearPageBuddy(page);
738 set_page_private(page, 0);
742 * This function checks whether a page is free && is the buddy
743 * we can do coalesce a page and its buddy if
744 * (a) the buddy is not in a hole (check before calling!) &&
745 * (b) the buddy is in the buddy system &&
746 * (c) a page and its buddy have the same order &&
747 * (d) a page and its buddy are in the same zone.
749 * For recording whether a page is in the buddy system, we set ->_mapcount
750 * PAGE_BUDDY_MAPCOUNT_VALUE.
751 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
752 * serialized by zone->lock.
754 * For recording page's order, we use page_private(page).
756 static inline int page_is_buddy(struct page *page, struct page *buddy,
759 if (page_is_guard(buddy) && page_order(buddy) == order) {
760 if (page_zone_id(page) != page_zone_id(buddy))
763 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
768 if (PageBuddy(buddy) && page_order(buddy) == order) {
770 * zone check is done late to avoid uselessly
771 * calculating zone/node ids for pages that could
774 if (page_zone_id(page) != page_zone_id(buddy))
777 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
785 * Freeing function for a buddy system allocator.
787 * The concept of a buddy system is to maintain direct-mapped table
788 * (containing bit values) for memory blocks of various "orders".
789 * The bottom level table contains the map for the smallest allocatable
790 * units of memory (here, pages), and each level above it describes
791 * pairs of units from the levels below, hence, "buddies".
792 * At a high level, all that happens here is marking the table entry
793 * at the bottom level available, and propagating the changes upward
794 * as necessary, plus some accounting needed to play nicely with other
795 * parts of the VM system.
796 * At each level, we keep a list of pages, which are heads of continuous
797 * free pages of length of (1 << order) and marked with _mapcount
798 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
800 * So when we are allocating or freeing one, we can derive the state of the
801 * other. That is, if we allocate a small block, and both were
802 * free, the remainder of the region must be split into blocks.
803 * If a block is freed, and its buddy is also free, then this
804 * triggers coalescing into a block of larger size.
809 static inline void __free_one_page(struct page *page,
811 struct zone *zone, unsigned int order,
814 unsigned long combined_pfn;
815 unsigned long uninitialized_var(buddy_pfn);
817 unsigned int max_order;
819 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
821 VM_BUG_ON(!zone_is_initialized(zone));
822 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
824 VM_BUG_ON(migratetype == -1);
825 if (likely(!is_migrate_isolate(migratetype)))
826 __mod_zone_freepage_state(zone, 1 << order, migratetype);
828 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
829 VM_BUG_ON_PAGE(bad_range(zone, page), page);
832 while (order < max_order - 1) {
833 buddy_pfn = __find_buddy_pfn(pfn, order);
834 buddy = page + (buddy_pfn - pfn);
836 if (!pfn_valid_within(buddy_pfn))
838 if (!page_is_buddy(page, buddy, order))
841 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
842 * merge with it and move up one order.
844 if (page_is_guard(buddy)) {
845 clear_page_guard(zone, buddy, order, migratetype);
847 list_del(&buddy->lru);
848 zone->free_area[order].nr_free--;
849 rmv_page_order(buddy);
851 combined_pfn = buddy_pfn & pfn;
852 page = page + (combined_pfn - pfn);
856 if (max_order < MAX_ORDER) {
857 /* If we are here, it means order is >= pageblock_order.
858 * We want to prevent merge between freepages on isolate
859 * pageblock and normal pageblock. Without this, pageblock
860 * isolation could cause incorrect freepage or CMA accounting.
862 * We don't want to hit this code for the more frequent
865 if (unlikely(has_isolate_pageblock(zone))) {
868 buddy_pfn = __find_buddy_pfn(pfn, order);
869 buddy = page + (buddy_pfn - pfn);
870 buddy_mt = get_pageblock_migratetype(buddy);
872 if (migratetype != buddy_mt
873 && (is_migrate_isolate(migratetype) ||
874 is_migrate_isolate(buddy_mt)))
878 goto continue_merging;
882 set_page_order(page, order);
885 * If this is not the largest possible page, check if the buddy
886 * of the next-highest order is free. If it is, it's possible
887 * that pages are being freed that will coalesce soon. In case,
888 * that is happening, add the free page to the tail of the list
889 * so it's less likely to be used soon and more likely to be merged
890 * as a higher order page
892 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
893 struct page *higher_page, *higher_buddy;
894 combined_pfn = buddy_pfn & pfn;
895 higher_page = page + (combined_pfn - pfn);
896 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
897 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
898 if (pfn_valid_within(buddy_pfn) &&
899 page_is_buddy(higher_page, higher_buddy, order + 1)) {
900 list_add_tail(&page->lru,
901 &zone->free_area[order].free_list[migratetype]);
906 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
908 zone->free_area[order].nr_free++;
912 * A bad page could be due to a number of fields. Instead of multiple branches,
913 * try and check multiple fields with one check. The caller must do a detailed
914 * check if necessary.
916 static inline bool page_expected_state(struct page *page,
917 unsigned long check_flags)
919 if (unlikely(atomic_read(&page->_mapcount) != -1))
922 if (unlikely((unsigned long)page->mapping |
923 page_ref_count(page) |
925 (unsigned long)page->mem_cgroup |
927 (page->flags & check_flags)))
933 static void free_pages_check_bad(struct page *page)
935 const char *bad_reason;
936 unsigned long bad_flags;
941 if (unlikely(atomic_read(&page->_mapcount) != -1))
942 bad_reason = "nonzero mapcount";
943 if (unlikely(page->mapping != NULL))
944 bad_reason = "non-NULL mapping";
945 if (unlikely(page_ref_count(page) != 0))
946 bad_reason = "nonzero _refcount";
947 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
948 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
949 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
952 if (unlikely(page->mem_cgroup))
953 bad_reason = "page still charged to cgroup";
955 bad_page(page, bad_reason, bad_flags);
958 static inline int free_pages_check(struct page *page)
960 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
963 /* Something has gone sideways, find it */
964 free_pages_check_bad(page);
968 static int free_tail_pages_check(struct page *head_page, struct page *page)
973 * We rely page->lru.next never has bit 0 set, unless the page
974 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
976 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
978 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
982 switch (page - head_page) {
984 /* the first tail page: ->mapping is compound_mapcount() */
985 if (unlikely(compound_mapcount(page))) {
986 bad_page(page, "nonzero compound_mapcount", 0);
992 * the second tail page: ->mapping is
993 * page_deferred_list().next -- ignore value.
997 if (page->mapping != TAIL_MAPPING) {
998 bad_page(page, "corrupted mapping in tail page", 0);
1003 if (unlikely(!PageTail(page))) {
1004 bad_page(page, "PageTail not set", 0);
1007 if (unlikely(compound_head(page) != head_page)) {
1008 bad_page(page, "compound_head not consistent", 0);
1013 page->mapping = NULL;
1014 clear_compound_head(page);
1018 static __always_inline bool free_pages_prepare(struct page *page,
1019 unsigned int order, bool check_free)
1023 VM_BUG_ON_PAGE(PageTail(page), page);
1025 trace_mm_page_free(page, order);
1028 * Check tail pages before head page information is cleared to
1029 * avoid checking PageCompound for order-0 pages.
1031 if (unlikely(order)) {
1032 bool compound = PageCompound(page);
1035 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1038 ClearPageDoubleMap(page);
1039 for (i = 1; i < (1 << order); i++) {
1041 bad += free_tail_pages_check(page, page + i);
1042 if (unlikely(free_pages_check(page + i))) {
1046 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 if (PageMappingFlags(page))
1050 page->mapping = NULL;
1051 if (memcg_kmem_enabled() && PageKmemcg(page))
1052 memcg_kmem_uncharge(page, order);
1054 bad += free_pages_check(page);
1058 page_cpupid_reset_last(page);
1059 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1060 reset_page_owner(page, order);
1062 if (!PageHighMem(page)) {
1063 debug_check_no_locks_freed(page_address(page),
1064 PAGE_SIZE << order);
1065 debug_check_no_obj_freed(page_address(page),
1066 PAGE_SIZE << order);
1068 arch_free_page(page, order);
1069 kernel_poison_pages(page, 1 << order, 0);
1070 kernel_map_pages(page, 1 << order, 0);
1071 kasan_free_pages(page, order);
1076 #ifdef CONFIG_DEBUG_VM
1077 static inline bool free_pcp_prepare(struct page *page)
1079 return free_pages_prepare(page, 0, true);
1082 static inline bool bulkfree_pcp_prepare(struct page *page)
1087 static bool free_pcp_prepare(struct page *page)
1089 return free_pages_prepare(page, 0, false);
1092 static bool bulkfree_pcp_prepare(struct page *page)
1094 return free_pages_check(page);
1096 #endif /* CONFIG_DEBUG_VM */
1099 * Frees a number of pages from the PCP lists
1100 * Assumes all pages on list are in same zone, and of same order.
1101 * count is the number of pages to free.
1103 * If the zone was previously in an "all pages pinned" state then look to
1104 * see if this freeing clears that state.
1106 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1107 * pinned" detection logic.
1109 static void free_pcppages_bulk(struct zone *zone, int count,
1110 struct per_cpu_pages *pcp)
1112 int migratetype = 0;
1114 bool isolated_pageblocks;
1116 spin_lock(&zone->lock);
1117 isolated_pageblocks = has_isolate_pageblock(zone);
1121 struct list_head *list;
1124 * Remove pages from lists in a round-robin fashion. A
1125 * batch_free count is maintained that is incremented when an
1126 * empty list is encountered. This is so more pages are freed
1127 * off fuller lists instead of spinning excessively around empty
1132 if (++migratetype == MIGRATE_PCPTYPES)
1134 list = &pcp->lists[migratetype];
1135 } while (list_empty(list));
1137 /* This is the only non-empty list. Free them all. */
1138 if (batch_free == MIGRATE_PCPTYPES)
1142 int mt; /* migratetype of the to-be-freed page */
1144 page = list_last_entry(list, struct page, lru);
1145 /* must delete as __free_one_page list manipulates */
1146 list_del(&page->lru);
1148 mt = get_pcppage_migratetype(page);
1149 /* MIGRATE_ISOLATE page should not go to pcplists */
1150 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1151 /* Pageblock could have been isolated meanwhile */
1152 if (unlikely(isolated_pageblocks))
1153 mt = get_pageblock_migratetype(page);
1155 if (bulkfree_pcp_prepare(page))
1158 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1159 trace_mm_page_pcpu_drain(page, 0, mt);
1160 } while (--count && --batch_free && !list_empty(list));
1162 spin_unlock(&zone->lock);
1165 static void free_one_page(struct zone *zone,
1166 struct page *page, unsigned long pfn,
1170 spin_lock(&zone->lock);
1171 if (unlikely(has_isolate_pageblock(zone) ||
1172 is_migrate_isolate(migratetype))) {
1173 migratetype = get_pfnblock_migratetype(page, pfn);
1175 __free_one_page(page, pfn, zone, order, migratetype);
1176 spin_unlock(&zone->lock);
1179 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1180 unsigned long zone, int nid)
1182 mm_zero_struct_page(page);
1183 set_page_links(page, zone, nid, pfn);
1184 init_page_count(page);
1185 page_mapcount_reset(page);
1186 page_cpupid_reset_last(page);
1188 INIT_LIST_HEAD(&page->lru);
1189 #ifdef WANT_PAGE_VIRTUAL
1190 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1191 if (!is_highmem_idx(zone))
1192 set_page_address(page, __va(pfn << PAGE_SHIFT));
1196 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1199 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1202 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1203 static void __meminit init_reserved_page(unsigned long pfn)
1208 if (!early_page_uninitialised(pfn))
1211 nid = early_pfn_to_nid(pfn);
1212 pgdat = NODE_DATA(nid);
1214 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1215 struct zone *zone = &pgdat->node_zones[zid];
1217 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1220 __init_single_pfn(pfn, zid, nid);
1223 static inline void init_reserved_page(unsigned long pfn)
1226 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1229 * Initialised pages do not have PageReserved set. This function is
1230 * called for each range allocated by the bootmem allocator and
1231 * marks the pages PageReserved. The remaining valid pages are later
1232 * sent to the buddy page allocator.
1234 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1236 unsigned long start_pfn = PFN_DOWN(start);
1237 unsigned long end_pfn = PFN_UP(end);
1239 for (; start_pfn < end_pfn; start_pfn++) {
1240 if (pfn_valid(start_pfn)) {
1241 struct page *page = pfn_to_page(start_pfn);
1243 init_reserved_page(start_pfn);
1245 /* Avoid false-positive PageTail() */
1246 INIT_LIST_HEAD(&page->lru);
1248 SetPageReserved(page);
1253 static void __free_pages_ok(struct page *page, unsigned int order)
1255 unsigned long flags;
1257 unsigned long pfn = page_to_pfn(page);
1259 if (!free_pages_prepare(page, order, true))
1262 migratetype = get_pfnblock_migratetype(page, pfn);
1263 local_irq_save(flags);
1264 __count_vm_events(PGFREE, 1 << order);
1265 free_one_page(page_zone(page), page, pfn, order, migratetype);
1266 local_irq_restore(flags);
1269 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1271 unsigned int nr_pages = 1 << order;
1272 struct page *p = page;
1276 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1278 __ClearPageReserved(p);
1279 set_page_count(p, 0);
1281 __ClearPageReserved(p);
1282 set_page_count(p, 0);
1284 page_zone(page)->managed_pages += nr_pages;
1285 set_page_refcounted(page);
1286 __free_pages(page, order);
1289 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1290 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1292 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1294 int __meminit early_pfn_to_nid(unsigned long pfn)
1296 static DEFINE_SPINLOCK(early_pfn_lock);
1299 spin_lock(&early_pfn_lock);
1300 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1302 nid = first_online_node;
1303 spin_unlock(&early_pfn_lock);
1309 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1310 static inline bool __meminit __maybe_unused
1311 meminit_pfn_in_nid(unsigned long pfn, int node,
1312 struct mminit_pfnnid_cache *state)
1316 nid = __early_pfn_to_nid(pfn, state);
1317 if (nid >= 0 && nid != node)
1322 /* Only safe to use early in boot when initialisation is single-threaded */
1323 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1325 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1330 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1334 static inline bool __meminit __maybe_unused
1335 meminit_pfn_in_nid(unsigned long pfn, int node,
1336 struct mminit_pfnnid_cache *state)
1343 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1346 if (early_page_uninitialised(pfn))
1348 return __free_pages_boot_core(page, order);
1352 * Check that the whole (or subset of) a pageblock given by the interval of
1353 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1354 * with the migration of free compaction scanner. The scanners then need to
1355 * use only pfn_valid_within() check for arches that allow holes within
1358 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1360 * It's possible on some configurations to have a setup like node0 node1 node0
1361 * i.e. it's possible that all pages within a zones range of pages do not
1362 * belong to a single zone. We assume that a border between node0 and node1
1363 * can occur within a single pageblock, but not a node0 node1 node0
1364 * interleaving within a single pageblock. It is therefore sufficient to check
1365 * the first and last page of a pageblock and avoid checking each individual
1366 * page in a pageblock.
1368 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1369 unsigned long end_pfn, struct zone *zone)
1371 struct page *start_page;
1372 struct page *end_page;
1374 /* end_pfn is one past the range we are checking */
1377 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1380 start_page = pfn_to_online_page(start_pfn);
1384 if (page_zone(start_page) != zone)
1387 end_page = pfn_to_page(end_pfn);
1389 /* This gives a shorter code than deriving page_zone(end_page) */
1390 if (page_zone_id(start_page) != page_zone_id(end_page))
1396 void set_zone_contiguous(struct zone *zone)
1398 unsigned long block_start_pfn = zone->zone_start_pfn;
1399 unsigned long block_end_pfn;
1401 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1402 for (; block_start_pfn < zone_end_pfn(zone);
1403 block_start_pfn = block_end_pfn,
1404 block_end_pfn += pageblock_nr_pages) {
1406 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1408 if (!__pageblock_pfn_to_page(block_start_pfn,
1409 block_end_pfn, zone))
1413 /* We confirm that there is no hole */
1414 zone->contiguous = true;
1417 void clear_zone_contiguous(struct zone *zone)
1419 zone->contiguous = false;
1422 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1423 static void __init deferred_free_range(unsigned long pfn,
1424 unsigned long nr_pages)
1432 page = pfn_to_page(pfn);
1434 /* Free a large naturally-aligned chunk if possible */
1435 if (nr_pages == pageblock_nr_pages &&
1436 (pfn & (pageblock_nr_pages - 1)) == 0) {
1437 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1438 __free_pages_boot_core(page, pageblock_order);
1442 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1443 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1444 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1445 __free_pages_boot_core(page, 0);
1449 /* Completion tracking for deferred_init_memmap() threads */
1450 static atomic_t pgdat_init_n_undone __initdata;
1451 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1453 static inline void __init pgdat_init_report_one_done(void)
1455 if (atomic_dec_and_test(&pgdat_init_n_undone))
1456 complete(&pgdat_init_all_done_comp);
1460 * Helper for deferred_init_range, free the given range, reset the counters, and
1461 * return number of pages freed.
1463 static inline unsigned long __init __def_free(unsigned long *nr_free,
1464 unsigned long *free_base_pfn,
1467 unsigned long nr = *nr_free;
1469 deferred_free_range(*free_base_pfn, nr);
1477 static unsigned long __init deferred_init_range(int nid, int zid,
1478 unsigned long start_pfn,
1479 unsigned long end_pfn)
1481 struct mminit_pfnnid_cache nid_init_state = { };
1482 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1483 unsigned long free_base_pfn = 0;
1484 unsigned long nr_pages = 0;
1485 unsigned long nr_free = 0;
1486 struct page *page = NULL;
1490 * First we check if pfn is valid on architectures where it is possible
1491 * to have holes within pageblock_nr_pages. On systems where it is not
1492 * possible, this function is optimized out.
1494 * Then, we check if a current large page is valid by only checking the
1495 * validity of the head pfn.
1497 * meminit_pfn_in_nid is checked on systems where pfns can interleave
1498 * within a node: a pfn is between start and end of a node, but does not
1499 * belong to this memory node.
1501 * Finally, we minimize pfn page lookups and scheduler checks by
1502 * performing it only once every pageblock_nr_pages.
1504 * We do it in two loops: first we initialize struct page, than free to
1505 * buddy allocator, becuse while we are freeing pages we can access
1506 * pages that are ahead (computing buddy page in __free_one_page()).
1508 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1509 if (!pfn_valid_within(pfn))
1511 if ((pfn & nr_pgmask) || pfn_valid(pfn)) {
1512 if (meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1513 if (page && (pfn & nr_pgmask))
1516 page = pfn_to_page(pfn);
1517 __init_single_page(page, pfn, zid, nid);
1524 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
1525 if (!pfn_valid_within(pfn)) {
1526 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1527 } else if (!(pfn & nr_pgmask) && !pfn_valid(pfn)) {
1528 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1529 } else if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1530 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1531 } else if (page && (pfn & nr_pgmask)) {
1535 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1536 page = pfn_to_page(pfn);
1537 free_base_pfn = pfn;
1542 /* Free the last block of pages to allocator */
1543 nr_pages += __def_free(&nr_free, &free_base_pfn, &page);
1548 /* Initialise remaining memory on a node */
1549 static int __init deferred_init_memmap(void *data)
1551 pg_data_t *pgdat = data;
1552 int nid = pgdat->node_id;
1553 unsigned long start = jiffies;
1554 unsigned long nr_pages = 0;
1555 unsigned long spfn, epfn;
1556 phys_addr_t spa, epa;
1559 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1560 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1563 if (first_init_pfn == ULONG_MAX) {
1564 pgdat_init_report_one_done();
1568 /* Bind memory initialisation thread to a local node if possible */
1569 if (!cpumask_empty(cpumask))
1570 set_cpus_allowed_ptr(current, cpumask);
1572 /* Sanity check boundaries */
1573 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1574 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1575 pgdat->first_deferred_pfn = ULONG_MAX;
1577 /* Only the highest zone is deferred so find it */
1578 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1579 zone = pgdat->node_zones + zid;
1580 if (first_init_pfn < zone_end_pfn(zone))
1583 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1585 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1586 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1587 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1588 nr_pages += deferred_init_range(nid, zid, spfn, epfn);
1591 /* Sanity check that the next zone really is unpopulated */
1592 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1594 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1595 jiffies_to_msecs(jiffies - start));
1597 pgdat_init_report_one_done();
1600 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1602 void __init page_alloc_init_late(void)
1606 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1609 /* There will be num_node_state(N_MEMORY) threads */
1610 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1611 for_each_node_state(nid, N_MEMORY) {
1612 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1615 /* Block until all are initialised */
1616 wait_for_completion(&pgdat_init_all_done_comp);
1618 /* Reinit limits that are based on free pages after the kernel is up */
1619 files_maxfiles_init();
1621 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1622 /* Discard memblock private memory */
1626 for_each_populated_zone(zone)
1627 set_zone_contiguous(zone);
1631 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1632 void __init init_cma_reserved_pageblock(struct page *page)
1634 unsigned i = pageblock_nr_pages;
1635 struct page *p = page;
1638 __ClearPageReserved(p);
1639 set_page_count(p, 0);
1642 set_pageblock_migratetype(page, MIGRATE_CMA);
1644 if (pageblock_order >= MAX_ORDER) {
1645 i = pageblock_nr_pages;
1648 set_page_refcounted(p);
1649 __free_pages(p, MAX_ORDER - 1);
1650 p += MAX_ORDER_NR_PAGES;
1651 } while (i -= MAX_ORDER_NR_PAGES);
1653 set_page_refcounted(page);
1654 __free_pages(page, pageblock_order);
1657 adjust_managed_page_count(page, pageblock_nr_pages);
1662 * The order of subdivision here is critical for the IO subsystem.
1663 * Please do not alter this order without good reasons and regression
1664 * testing. Specifically, as large blocks of memory are subdivided,
1665 * the order in which smaller blocks are delivered depends on the order
1666 * they're subdivided in this function. This is the primary factor
1667 * influencing the order in which pages are delivered to the IO
1668 * subsystem according to empirical testing, and this is also justified
1669 * by considering the behavior of a buddy system containing a single
1670 * large block of memory acted on by a series of small allocations.
1671 * This behavior is a critical factor in sglist merging's success.
1675 static inline void expand(struct zone *zone, struct page *page,
1676 int low, int high, struct free_area *area,
1679 unsigned long size = 1 << high;
1681 while (high > low) {
1685 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1688 * Mark as guard pages (or page), that will allow to
1689 * merge back to allocator when buddy will be freed.
1690 * Corresponding page table entries will not be touched,
1691 * pages will stay not present in virtual address space
1693 if (set_page_guard(zone, &page[size], high, migratetype))
1696 list_add(&page[size].lru, &area->free_list[migratetype]);
1698 set_page_order(&page[size], high);
1702 static void check_new_page_bad(struct page *page)
1704 const char *bad_reason = NULL;
1705 unsigned long bad_flags = 0;
1707 if (unlikely(atomic_read(&page->_mapcount) != -1))
1708 bad_reason = "nonzero mapcount";
1709 if (unlikely(page->mapping != NULL))
1710 bad_reason = "non-NULL mapping";
1711 if (unlikely(page_ref_count(page) != 0))
1712 bad_reason = "nonzero _count";
1713 if (unlikely(page->flags & __PG_HWPOISON)) {
1714 bad_reason = "HWPoisoned (hardware-corrupted)";
1715 bad_flags = __PG_HWPOISON;
1716 /* Don't complain about hwpoisoned pages */
1717 page_mapcount_reset(page); /* remove PageBuddy */
1720 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1721 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1722 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1725 if (unlikely(page->mem_cgroup))
1726 bad_reason = "page still charged to cgroup";
1728 bad_page(page, bad_reason, bad_flags);
1732 * This page is about to be returned from the page allocator
1734 static inline int check_new_page(struct page *page)
1736 if (likely(page_expected_state(page,
1737 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1740 check_new_page_bad(page);
1744 static inline bool free_pages_prezeroed(void)
1746 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1747 page_poisoning_enabled();
1750 #ifdef CONFIG_DEBUG_VM
1751 static bool check_pcp_refill(struct page *page)
1756 static bool check_new_pcp(struct page *page)
1758 return check_new_page(page);
1761 static bool check_pcp_refill(struct page *page)
1763 return check_new_page(page);
1765 static bool check_new_pcp(struct page *page)
1769 #endif /* CONFIG_DEBUG_VM */
1771 static bool check_new_pages(struct page *page, unsigned int order)
1774 for (i = 0; i < (1 << order); i++) {
1775 struct page *p = page + i;
1777 if (unlikely(check_new_page(p)))
1784 inline void post_alloc_hook(struct page *page, unsigned int order,
1787 set_page_private(page, 0);
1788 set_page_refcounted(page);
1790 arch_alloc_page(page, order);
1791 kernel_map_pages(page, 1 << order, 1);
1792 kernel_poison_pages(page, 1 << order, 1);
1793 kasan_alloc_pages(page, order);
1794 set_page_owner(page, order, gfp_flags);
1797 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1798 unsigned int alloc_flags)
1802 post_alloc_hook(page, order, gfp_flags);
1804 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1805 for (i = 0; i < (1 << order); i++)
1806 clear_highpage(page + i);
1808 if (order && (gfp_flags & __GFP_COMP))
1809 prep_compound_page(page, order);
1812 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1813 * allocate the page. The expectation is that the caller is taking
1814 * steps that will free more memory. The caller should avoid the page
1815 * being used for !PFMEMALLOC purposes.
1817 if (alloc_flags & ALLOC_NO_WATERMARKS)
1818 set_page_pfmemalloc(page);
1820 clear_page_pfmemalloc(page);
1824 * Go through the free lists for the given migratetype and remove
1825 * the smallest available page from the freelists
1827 static __always_inline
1828 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1831 unsigned int current_order;
1832 struct free_area *area;
1835 /* Find a page of the appropriate size in the preferred list */
1836 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1837 area = &(zone->free_area[current_order]);
1838 page = list_first_entry_or_null(&area->free_list[migratetype],
1842 list_del(&page->lru);
1843 rmv_page_order(page);
1845 expand(zone, page, order, current_order, area, migratetype);
1846 set_pcppage_migratetype(page, migratetype);
1855 * This array describes the order lists are fallen back to when
1856 * the free lists for the desirable migrate type are depleted
1858 static int fallbacks[MIGRATE_TYPES][4] = {
1859 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1860 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1861 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1863 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1865 #ifdef CONFIG_MEMORY_ISOLATION
1866 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1871 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1874 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1877 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1878 unsigned int order) { return NULL; }
1882 * Move the free pages in a range to the free lists of the requested type.
1883 * Note that start_page and end_pages are not aligned on a pageblock
1884 * boundary. If alignment is required, use move_freepages_block()
1886 static int move_freepages(struct zone *zone,
1887 struct page *start_page, struct page *end_page,
1888 int migratetype, int *num_movable)
1892 int pages_moved = 0;
1894 #ifndef CONFIG_HOLES_IN_ZONE
1896 * page_zone is not safe to call in this context when
1897 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1898 * anyway as we check zone boundaries in move_freepages_block().
1899 * Remove at a later date when no bug reports exist related to
1900 * grouping pages by mobility
1902 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1908 for (page = start_page; page <= end_page;) {
1909 if (!pfn_valid_within(page_to_pfn(page))) {
1914 /* Make sure we are not inadvertently changing nodes */
1915 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1917 if (!PageBuddy(page)) {
1919 * We assume that pages that could be isolated for
1920 * migration are movable. But we don't actually try
1921 * isolating, as that would be expensive.
1924 (PageLRU(page) || __PageMovable(page)))
1931 order = page_order(page);
1932 list_move(&page->lru,
1933 &zone->free_area[order].free_list[migratetype]);
1935 pages_moved += 1 << order;
1941 int move_freepages_block(struct zone *zone, struct page *page,
1942 int migratetype, int *num_movable)
1944 unsigned long start_pfn, end_pfn;
1945 struct page *start_page, *end_page;
1947 start_pfn = page_to_pfn(page);
1948 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1949 start_page = pfn_to_page(start_pfn);
1950 end_page = start_page + pageblock_nr_pages - 1;
1951 end_pfn = start_pfn + pageblock_nr_pages - 1;
1953 /* Do not cross zone boundaries */
1954 if (!zone_spans_pfn(zone, start_pfn))
1956 if (!zone_spans_pfn(zone, end_pfn))
1959 return move_freepages(zone, start_page, end_page, migratetype,
1963 static void change_pageblock_range(struct page *pageblock_page,
1964 int start_order, int migratetype)
1966 int nr_pageblocks = 1 << (start_order - pageblock_order);
1968 while (nr_pageblocks--) {
1969 set_pageblock_migratetype(pageblock_page, migratetype);
1970 pageblock_page += pageblock_nr_pages;
1975 * When we are falling back to another migratetype during allocation, try to
1976 * steal extra free pages from the same pageblocks to satisfy further
1977 * allocations, instead of polluting multiple pageblocks.
1979 * If we are stealing a relatively large buddy page, it is likely there will
1980 * be more free pages in the pageblock, so try to steal them all. For
1981 * reclaimable and unmovable allocations, we steal regardless of page size,
1982 * as fragmentation caused by those allocations polluting movable pageblocks
1983 * is worse than movable allocations stealing from unmovable and reclaimable
1986 static bool can_steal_fallback(unsigned int order, int start_mt)
1989 * Leaving this order check is intended, although there is
1990 * relaxed order check in next check. The reason is that
1991 * we can actually steal whole pageblock if this condition met,
1992 * but, below check doesn't guarantee it and that is just heuristic
1993 * so could be changed anytime.
1995 if (order >= pageblock_order)
1998 if (order >= pageblock_order / 2 ||
1999 start_mt == MIGRATE_RECLAIMABLE ||
2000 start_mt == MIGRATE_UNMOVABLE ||
2001 page_group_by_mobility_disabled)
2008 * This function implements actual steal behaviour. If order is large enough,
2009 * we can steal whole pageblock. If not, we first move freepages in this
2010 * pageblock to our migratetype and determine how many already-allocated pages
2011 * are there in the pageblock with a compatible migratetype. If at least half
2012 * of pages are free or compatible, we can change migratetype of the pageblock
2013 * itself, so pages freed in the future will be put on the correct free list.
2015 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2016 int start_type, bool whole_block)
2018 unsigned int current_order = page_order(page);
2019 struct free_area *area;
2020 int free_pages, movable_pages, alike_pages;
2023 old_block_type = get_pageblock_migratetype(page);
2026 * This can happen due to races and we want to prevent broken
2027 * highatomic accounting.
2029 if (is_migrate_highatomic(old_block_type))
2032 /* Take ownership for orders >= pageblock_order */
2033 if (current_order >= pageblock_order) {
2034 change_pageblock_range(page, current_order, start_type);
2038 /* We are not allowed to try stealing from the whole block */
2042 free_pages = move_freepages_block(zone, page, start_type,
2045 * Determine how many pages are compatible with our allocation.
2046 * For movable allocation, it's the number of movable pages which
2047 * we just obtained. For other types it's a bit more tricky.
2049 if (start_type == MIGRATE_MOVABLE) {
2050 alike_pages = movable_pages;
2053 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2054 * to MOVABLE pageblock, consider all non-movable pages as
2055 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2056 * vice versa, be conservative since we can't distinguish the
2057 * exact migratetype of non-movable pages.
2059 if (old_block_type == MIGRATE_MOVABLE)
2060 alike_pages = pageblock_nr_pages
2061 - (free_pages + movable_pages);
2066 /* moving whole block can fail due to zone boundary conditions */
2071 * If a sufficient number of pages in the block are either free or of
2072 * comparable migratability as our allocation, claim the whole block.
2074 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2075 page_group_by_mobility_disabled)
2076 set_pageblock_migratetype(page, start_type);
2081 area = &zone->free_area[current_order];
2082 list_move(&page->lru, &area->free_list[start_type]);
2086 * Check whether there is a suitable fallback freepage with requested order.
2087 * If only_stealable is true, this function returns fallback_mt only if
2088 * we can steal other freepages all together. This would help to reduce
2089 * fragmentation due to mixed migratetype pages in one pageblock.
2091 int find_suitable_fallback(struct free_area *area, unsigned int order,
2092 int migratetype, bool only_stealable, bool *can_steal)
2097 if (area->nr_free == 0)
2102 fallback_mt = fallbacks[migratetype][i];
2103 if (fallback_mt == MIGRATE_TYPES)
2106 if (list_empty(&area->free_list[fallback_mt]))
2109 if (can_steal_fallback(order, migratetype))
2112 if (!only_stealable)
2123 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2124 * there are no empty page blocks that contain a page with a suitable order
2126 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2127 unsigned int alloc_order)
2130 unsigned long max_managed, flags;
2133 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2134 * Check is race-prone but harmless.
2136 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2137 if (zone->nr_reserved_highatomic >= max_managed)
2140 spin_lock_irqsave(&zone->lock, flags);
2142 /* Recheck the nr_reserved_highatomic limit under the lock */
2143 if (zone->nr_reserved_highatomic >= max_managed)
2147 mt = get_pageblock_migratetype(page);
2148 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2149 && !is_migrate_cma(mt)) {
2150 zone->nr_reserved_highatomic += pageblock_nr_pages;
2151 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2152 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2156 spin_unlock_irqrestore(&zone->lock, flags);
2160 * Used when an allocation is about to fail under memory pressure. This
2161 * potentially hurts the reliability of high-order allocations when under
2162 * intense memory pressure but failed atomic allocations should be easier
2163 * to recover from than an OOM.
2165 * If @force is true, try to unreserve a pageblock even though highatomic
2166 * pageblock is exhausted.
2168 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2171 struct zonelist *zonelist = ac->zonelist;
2172 unsigned long flags;
2179 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2182 * Preserve at least one pageblock unless memory pressure
2185 if (!force && zone->nr_reserved_highatomic <=
2189 spin_lock_irqsave(&zone->lock, flags);
2190 for (order = 0; order < MAX_ORDER; order++) {
2191 struct free_area *area = &(zone->free_area[order]);
2193 page = list_first_entry_or_null(
2194 &area->free_list[MIGRATE_HIGHATOMIC],
2200 * In page freeing path, migratetype change is racy so
2201 * we can counter several free pages in a pageblock
2202 * in this loop althoug we changed the pageblock type
2203 * from highatomic to ac->migratetype. So we should
2204 * adjust the count once.
2206 if (is_migrate_highatomic_page(page)) {
2208 * It should never happen but changes to
2209 * locking could inadvertently allow a per-cpu
2210 * drain to add pages to MIGRATE_HIGHATOMIC
2211 * while unreserving so be safe and watch for
2214 zone->nr_reserved_highatomic -= min(
2216 zone->nr_reserved_highatomic);
2220 * Convert to ac->migratetype and avoid the normal
2221 * pageblock stealing heuristics. Minimally, the caller
2222 * is doing the work and needs the pages. More
2223 * importantly, if the block was always converted to
2224 * MIGRATE_UNMOVABLE or another type then the number
2225 * of pageblocks that cannot be completely freed
2228 set_pageblock_migratetype(page, ac->migratetype);
2229 ret = move_freepages_block(zone, page, ac->migratetype,
2232 spin_unlock_irqrestore(&zone->lock, flags);
2236 spin_unlock_irqrestore(&zone->lock, flags);
2243 * Try finding a free buddy page on the fallback list and put it on the free
2244 * list of requested migratetype, possibly along with other pages from the same
2245 * block, depending on fragmentation avoidance heuristics. Returns true if
2246 * fallback was found so that __rmqueue_smallest() can grab it.
2248 * The use of signed ints for order and current_order is a deliberate
2249 * deviation from the rest of this file, to make the for loop
2250 * condition simpler.
2252 static __always_inline bool
2253 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2255 struct free_area *area;
2262 * Find the largest available free page in the other list. This roughly
2263 * approximates finding the pageblock with the most free pages, which
2264 * would be too costly to do exactly.
2266 for (current_order = MAX_ORDER - 1; current_order >= order;
2268 area = &(zone->free_area[current_order]);
2269 fallback_mt = find_suitable_fallback(area, current_order,
2270 start_migratetype, false, &can_steal);
2271 if (fallback_mt == -1)
2275 * We cannot steal all free pages from the pageblock and the
2276 * requested migratetype is movable. In that case it's better to
2277 * steal and split the smallest available page instead of the
2278 * largest available page, because even if the next movable
2279 * allocation falls back into a different pageblock than this
2280 * one, it won't cause permanent fragmentation.
2282 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2283 && current_order > order)
2292 for (current_order = order; current_order < MAX_ORDER;
2294 area = &(zone->free_area[current_order]);
2295 fallback_mt = find_suitable_fallback(area, current_order,
2296 start_migratetype, false, &can_steal);
2297 if (fallback_mt != -1)
2302 * This should not happen - we already found a suitable fallback
2303 * when looking for the largest page.
2305 VM_BUG_ON(current_order == MAX_ORDER);
2308 page = list_first_entry(&area->free_list[fallback_mt],
2311 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2313 trace_mm_page_alloc_extfrag(page, order, current_order,
2314 start_migratetype, fallback_mt);
2321 * Do the hard work of removing an element from the buddy allocator.
2322 * Call me with the zone->lock already held.
2324 static __always_inline struct page *
2325 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2330 page = __rmqueue_smallest(zone, order, migratetype);
2331 if (unlikely(!page)) {
2332 if (migratetype == MIGRATE_MOVABLE)
2333 page = __rmqueue_cma_fallback(zone, order);
2335 if (!page && __rmqueue_fallback(zone, order, migratetype))
2339 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2344 * Obtain a specified number of elements from the buddy allocator, all under
2345 * a single hold of the lock, for efficiency. Add them to the supplied list.
2346 * Returns the number of new pages which were placed at *list.
2348 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2349 unsigned long count, struct list_head *list,
2354 spin_lock(&zone->lock);
2355 for (i = 0; i < count; ++i) {
2356 struct page *page = __rmqueue(zone, order, migratetype);
2357 if (unlikely(page == NULL))
2360 if (unlikely(check_pcp_refill(page)))
2364 * Split buddy pages returned by expand() are received here in
2365 * physical page order. The page is added to the tail of
2366 * caller's list. From the callers perspective, the linked list
2367 * is ordered by page number under some conditions. This is
2368 * useful for IO devices that can forward direction from the
2369 * head, thus also in the physical page order. This is useful
2370 * for IO devices that can merge IO requests if the physical
2371 * pages are ordered properly.
2373 list_add_tail(&page->lru, list);
2375 if (is_migrate_cma(get_pcppage_migratetype(page)))
2376 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2381 * i pages were removed from the buddy list even if some leak due
2382 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2383 * on i. Do not confuse with 'alloced' which is the number of
2384 * pages added to the pcp list.
2386 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2387 spin_unlock(&zone->lock);
2393 * Called from the vmstat counter updater to drain pagesets of this
2394 * currently executing processor on remote nodes after they have
2397 * Note that this function must be called with the thread pinned to
2398 * a single processor.
2400 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2402 unsigned long flags;
2403 int to_drain, batch;
2405 local_irq_save(flags);
2406 batch = READ_ONCE(pcp->batch);
2407 to_drain = min(pcp->count, batch);
2409 free_pcppages_bulk(zone, to_drain, pcp);
2410 pcp->count -= to_drain;
2412 local_irq_restore(flags);
2417 * Drain pcplists of the indicated processor and zone.
2419 * The processor must either be the current processor and the
2420 * thread pinned to the current processor or a processor that
2423 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2425 unsigned long flags;
2426 struct per_cpu_pageset *pset;
2427 struct per_cpu_pages *pcp;
2429 local_irq_save(flags);
2430 pset = per_cpu_ptr(zone->pageset, cpu);
2434 free_pcppages_bulk(zone, pcp->count, pcp);
2437 local_irq_restore(flags);
2441 * Drain pcplists of all zones on the indicated processor.
2443 * The processor must either be the current processor and the
2444 * thread pinned to the current processor or a processor that
2447 static void drain_pages(unsigned int cpu)
2451 for_each_populated_zone(zone) {
2452 drain_pages_zone(cpu, zone);
2457 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2459 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2460 * the single zone's pages.
2462 void drain_local_pages(struct zone *zone)
2464 int cpu = smp_processor_id();
2467 drain_pages_zone(cpu, zone);
2472 static void drain_local_pages_wq(struct work_struct *work)
2475 * drain_all_pages doesn't use proper cpu hotplug protection so
2476 * we can race with cpu offline when the WQ can move this from
2477 * a cpu pinned worker to an unbound one. We can operate on a different
2478 * cpu which is allright but we also have to make sure to not move to
2482 drain_local_pages(NULL);
2487 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2489 * When zone parameter is non-NULL, spill just the single zone's pages.
2491 * Note that this can be extremely slow as the draining happens in a workqueue.
2493 void drain_all_pages(struct zone *zone)
2498 * Allocate in the BSS so we wont require allocation in
2499 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2501 static cpumask_t cpus_with_pcps;
2504 * Make sure nobody triggers this path before mm_percpu_wq is fully
2507 if (WARN_ON_ONCE(!mm_percpu_wq))
2511 * Do not drain if one is already in progress unless it's specific to
2512 * a zone. Such callers are primarily CMA and memory hotplug and need
2513 * the drain to be complete when the call returns.
2515 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2518 mutex_lock(&pcpu_drain_mutex);
2522 * We don't care about racing with CPU hotplug event
2523 * as offline notification will cause the notified
2524 * cpu to drain that CPU pcps and on_each_cpu_mask
2525 * disables preemption as part of its processing
2527 for_each_online_cpu(cpu) {
2528 struct per_cpu_pageset *pcp;
2530 bool has_pcps = false;
2533 pcp = per_cpu_ptr(zone->pageset, cpu);
2537 for_each_populated_zone(z) {
2538 pcp = per_cpu_ptr(z->pageset, cpu);
2539 if (pcp->pcp.count) {
2547 cpumask_set_cpu(cpu, &cpus_with_pcps);
2549 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2552 for_each_cpu(cpu, &cpus_with_pcps) {
2553 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2554 INIT_WORK(work, drain_local_pages_wq);
2555 queue_work_on(cpu, mm_percpu_wq, work);
2557 for_each_cpu(cpu, &cpus_with_pcps)
2558 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2560 mutex_unlock(&pcpu_drain_mutex);
2563 #ifdef CONFIG_HIBERNATION
2566 * Touch the watchdog for every WD_PAGE_COUNT pages.
2568 #define WD_PAGE_COUNT (128*1024)
2570 void mark_free_pages(struct zone *zone)
2572 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2573 unsigned long flags;
2574 unsigned int order, t;
2577 if (zone_is_empty(zone))
2580 spin_lock_irqsave(&zone->lock, flags);
2582 max_zone_pfn = zone_end_pfn(zone);
2583 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2584 if (pfn_valid(pfn)) {
2585 page = pfn_to_page(pfn);
2587 if (!--page_count) {
2588 touch_nmi_watchdog();
2589 page_count = WD_PAGE_COUNT;
2592 if (page_zone(page) != zone)
2595 if (!swsusp_page_is_forbidden(page))
2596 swsusp_unset_page_free(page);
2599 for_each_migratetype_order(order, t) {
2600 list_for_each_entry(page,
2601 &zone->free_area[order].free_list[t], lru) {
2604 pfn = page_to_pfn(page);
2605 for (i = 0; i < (1UL << order); i++) {
2606 if (!--page_count) {
2607 touch_nmi_watchdog();
2608 page_count = WD_PAGE_COUNT;
2610 swsusp_set_page_free(pfn_to_page(pfn + i));
2614 spin_unlock_irqrestore(&zone->lock, flags);
2616 #endif /* CONFIG_PM */
2618 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2622 if (!free_pcp_prepare(page))
2625 migratetype = get_pfnblock_migratetype(page, pfn);
2626 set_pcppage_migratetype(page, migratetype);
2630 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2632 struct zone *zone = page_zone(page);
2633 struct per_cpu_pages *pcp;
2636 migratetype = get_pcppage_migratetype(page);
2637 __count_vm_event(PGFREE);
2640 * We only track unmovable, reclaimable and movable on pcp lists.
2641 * Free ISOLATE pages back to the allocator because they are being
2642 * offlined but treat HIGHATOMIC as movable pages so we can get those
2643 * areas back if necessary. Otherwise, we may have to free
2644 * excessively into the page allocator
2646 if (migratetype >= MIGRATE_PCPTYPES) {
2647 if (unlikely(is_migrate_isolate(migratetype))) {
2648 free_one_page(zone, page, pfn, 0, migratetype);
2651 migratetype = MIGRATE_MOVABLE;
2654 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2655 list_add(&page->lru, &pcp->lists[migratetype]);
2657 if (pcp->count >= pcp->high) {
2658 unsigned long batch = READ_ONCE(pcp->batch);
2659 free_pcppages_bulk(zone, batch, pcp);
2660 pcp->count -= batch;
2665 * Free a 0-order page
2667 void free_unref_page(struct page *page)
2669 unsigned long flags;
2670 unsigned long pfn = page_to_pfn(page);
2672 if (!free_unref_page_prepare(page, pfn))
2675 local_irq_save(flags);
2676 free_unref_page_commit(page, pfn);
2677 local_irq_restore(flags);
2681 * Free a list of 0-order pages
2683 void free_unref_page_list(struct list_head *list)
2685 struct page *page, *next;
2686 unsigned long flags, pfn;
2688 /* Prepare pages for freeing */
2689 list_for_each_entry_safe(page, next, list, lru) {
2690 pfn = page_to_pfn(page);
2691 if (!free_unref_page_prepare(page, pfn))
2692 list_del(&page->lru);
2693 set_page_private(page, pfn);
2696 local_irq_save(flags);
2697 list_for_each_entry_safe(page, next, list, lru) {
2698 unsigned long pfn = page_private(page);
2700 set_page_private(page, 0);
2701 trace_mm_page_free_batched(page);
2702 free_unref_page_commit(page, pfn);
2704 local_irq_restore(flags);
2708 * split_page takes a non-compound higher-order page, and splits it into
2709 * n (1<<order) sub-pages: page[0..n]
2710 * Each sub-page must be freed individually.
2712 * Note: this is probably too low level an operation for use in drivers.
2713 * Please consult with lkml before using this in your driver.
2715 void split_page(struct page *page, unsigned int order)
2719 VM_BUG_ON_PAGE(PageCompound(page), page);
2720 VM_BUG_ON_PAGE(!page_count(page), page);
2722 for (i = 1; i < (1 << order); i++)
2723 set_page_refcounted(page + i);
2724 split_page_owner(page, order);
2726 EXPORT_SYMBOL_GPL(split_page);
2728 int __isolate_free_page(struct page *page, unsigned int order)
2730 unsigned long watermark;
2734 BUG_ON(!PageBuddy(page));
2736 zone = page_zone(page);
2737 mt = get_pageblock_migratetype(page);
2739 if (!is_migrate_isolate(mt)) {
2741 * Obey watermarks as if the page was being allocated. We can
2742 * emulate a high-order watermark check with a raised order-0
2743 * watermark, because we already know our high-order page
2746 watermark = min_wmark_pages(zone) + (1UL << order);
2747 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2750 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2753 /* Remove page from free list */
2754 list_del(&page->lru);
2755 zone->free_area[order].nr_free--;
2756 rmv_page_order(page);
2759 * Set the pageblock if the isolated page is at least half of a
2762 if (order >= pageblock_order - 1) {
2763 struct page *endpage = page + (1 << order) - 1;
2764 for (; page < endpage; page += pageblock_nr_pages) {
2765 int mt = get_pageblock_migratetype(page);
2766 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2767 && !is_migrate_highatomic(mt))
2768 set_pageblock_migratetype(page,
2774 return 1UL << order;
2778 * Update NUMA hit/miss statistics
2780 * Must be called with interrupts disabled.
2782 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2785 enum numa_stat_item local_stat = NUMA_LOCAL;
2787 /* skip numa counters update if numa stats is disabled */
2788 if (!static_branch_likely(&vm_numa_stat_key))
2791 if (z->node != numa_node_id())
2792 local_stat = NUMA_OTHER;
2794 if (z->node == preferred_zone->node)
2795 __inc_numa_state(z, NUMA_HIT);
2797 __inc_numa_state(z, NUMA_MISS);
2798 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2800 __inc_numa_state(z, local_stat);
2804 /* Remove page from the per-cpu list, caller must protect the list */
2805 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2806 struct per_cpu_pages *pcp,
2807 struct list_head *list)
2812 if (list_empty(list)) {
2813 pcp->count += rmqueue_bulk(zone, 0,
2816 if (unlikely(list_empty(list)))
2820 page = list_first_entry(list, struct page, lru);
2821 list_del(&page->lru);
2823 } while (check_new_pcp(page));
2828 /* Lock and remove page from the per-cpu list */
2829 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2830 struct zone *zone, unsigned int order,
2831 gfp_t gfp_flags, int migratetype)
2833 struct per_cpu_pages *pcp;
2834 struct list_head *list;
2836 unsigned long flags;
2838 local_irq_save(flags);
2839 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2840 list = &pcp->lists[migratetype];
2841 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2843 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2844 zone_statistics(preferred_zone, zone);
2846 local_irq_restore(flags);
2851 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2854 struct page *rmqueue(struct zone *preferred_zone,
2855 struct zone *zone, unsigned int order,
2856 gfp_t gfp_flags, unsigned int alloc_flags,
2859 unsigned long flags;
2862 if (likely(order == 0)) {
2863 page = rmqueue_pcplist(preferred_zone, zone, order,
2864 gfp_flags, migratetype);
2869 * We most definitely don't want callers attempting to
2870 * allocate greater than order-1 page units with __GFP_NOFAIL.
2872 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2873 spin_lock_irqsave(&zone->lock, flags);
2877 if (alloc_flags & ALLOC_HARDER) {
2878 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2880 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2883 page = __rmqueue(zone, order, migratetype);
2884 } while (page && check_new_pages(page, order));
2885 spin_unlock(&zone->lock);
2888 __mod_zone_freepage_state(zone, -(1 << order),
2889 get_pcppage_migratetype(page));
2891 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2892 zone_statistics(preferred_zone, zone);
2893 local_irq_restore(flags);
2896 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2900 local_irq_restore(flags);
2904 #ifdef CONFIG_FAIL_PAGE_ALLOC
2907 struct fault_attr attr;
2909 bool ignore_gfp_highmem;
2910 bool ignore_gfp_reclaim;
2912 } fail_page_alloc = {
2913 .attr = FAULT_ATTR_INITIALIZER,
2914 .ignore_gfp_reclaim = true,
2915 .ignore_gfp_highmem = true,
2919 static int __init setup_fail_page_alloc(char *str)
2921 return setup_fault_attr(&fail_page_alloc.attr, str);
2923 __setup("fail_page_alloc=", setup_fail_page_alloc);
2925 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2927 if (order < fail_page_alloc.min_order)
2929 if (gfp_mask & __GFP_NOFAIL)
2931 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2933 if (fail_page_alloc.ignore_gfp_reclaim &&
2934 (gfp_mask & __GFP_DIRECT_RECLAIM))
2937 return should_fail(&fail_page_alloc.attr, 1 << order);
2940 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2942 static int __init fail_page_alloc_debugfs(void)
2944 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2947 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2948 &fail_page_alloc.attr);
2950 return PTR_ERR(dir);
2952 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2953 &fail_page_alloc.ignore_gfp_reclaim))
2955 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2956 &fail_page_alloc.ignore_gfp_highmem))
2958 if (!debugfs_create_u32("min-order", mode, dir,
2959 &fail_page_alloc.min_order))
2964 debugfs_remove_recursive(dir);
2969 late_initcall(fail_page_alloc_debugfs);
2971 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2973 #else /* CONFIG_FAIL_PAGE_ALLOC */
2975 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2980 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2983 * Return true if free base pages are above 'mark'. For high-order checks it
2984 * will return true of the order-0 watermark is reached and there is at least
2985 * one free page of a suitable size. Checking now avoids taking the zone lock
2986 * to check in the allocation paths if no pages are free.
2988 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2989 int classzone_idx, unsigned int alloc_flags,
2994 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2996 /* free_pages may go negative - that's OK */
2997 free_pages -= (1 << order) - 1;
2999 if (alloc_flags & ALLOC_HIGH)
3003 * If the caller does not have rights to ALLOC_HARDER then subtract
3004 * the high-atomic reserves. This will over-estimate the size of the
3005 * atomic reserve but it avoids a search.
3007 if (likely(!alloc_harder)) {
3008 free_pages -= z->nr_reserved_highatomic;
3011 * OOM victims can try even harder than normal ALLOC_HARDER
3012 * users on the grounds that it's definitely going to be in
3013 * the exit path shortly and free memory. Any allocation it
3014 * makes during the free path will be small and short-lived.
3016 if (alloc_flags & ALLOC_OOM)
3024 /* If allocation can't use CMA areas don't use free CMA pages */
3025 if (!(alloc_flags & ALLOC_CMA))
3026 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3030 * Check watermarks for an order-0 allocation request. If these
3031 * are not met, then a high-order request also cannot go ahead
3032 * even if a suitable page happened to be free.
3034 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3037 /* If this is an order-0 request then the watermark is fine */
3041 /* For a high-order request, check at least one suitable page is free */
3042 for (o = order; o < MAX_ORDER; o++) {
3043 struct free_area *area = &z->free_area[o];
3049 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3050 if (!list_empty(&area->free_list[mt]))
3055 if ((alloc_flags & ALLOC_CMA) &&
3056 !list_empty(&area->free_list[MIGRATE_CMA])) {
3061 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3067 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3068 int classzone_idx, unsigned int alloc_flags)
3070 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3071 zone_page_state(z, NR_FREE_PAGES));
3074 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3075 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3077 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3081 /* If allocation can't use CMA areas don't use free CMA pages */
3082 if (!(alloc_flags & ALLOC_CMA))
3083 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3087 * Fast check for order-0 only. If this fails then the reserves
3088 * need to be calculated. There is a corner case where the check
3089 * passes but only the high-order atomic reserve are free. If
3090 * the caller is !atomic then it'll uselessly search the free
3091 * list. That corner case is then slower but it is harmless.
3093 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3096 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3100 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3101 unsigned long mark, int classzone_idx)
3103 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3105 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3106 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3108 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3113 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3115 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3118 #else /* CONFIG_NUMA */
3119 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3123 #endif /* CONFIG_NUMA */
3126 * get_page_from_freelist goes through the zonelist trying to allocate
3129 static struct page *
3130 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3131 const struct alloc_context *ac)
3133 struct zoneref *z = ac->preferred_zoneref;
3135 struct pglist_data *last_pgdat_dirty_limit = NULL;
3138 * Scan zonelist, looking for a zone with enough free.
3139 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3141 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3146 if (cpusets_enabled() &&
3147 (alloc_flags & ALLOC_CPUSET) &&
3148 !__cpuset_zone_allowed(zone, gfp_mask))
3151 * When allocating a page cache page for writing, we
3152 * want to get it from a node that is within its dirty
3153 * limit, such that no single node holds more than its
3154 * proportional share of globally allowed dirty pages.
3155 * The dirty limits take into account the node's
3156 * lowmem reserves and high watermark so that kswapd
3157 * should be able to balance it without having to
3158 * write pages from its LRU list.
3160 * XXX: For now, allow allocations to potentially
3161 * exceed the per-node dirty limit in the slowpath
3162 * (spread_dirty_pages unset) before going into reclaim,
3163 * which is important when on a NUMA setup the allowed
3164 * nodes are together not big enough to reach the
3165 * global limit. The proper fix for these situations
3166 * will require awareness of nodes in the
3167 * dirty-throttling and the flusher threads.
3169 if (ac->spread_dirty_pages) {
3170 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3173 if (!node_dirty_ok(zone->zone_pgdat)) {
3174 last_pgdat_dirty_limit = zone->zone_pgdat;
3179 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3180 if (!zone_watermark_fast(zone, order, mark,
3181 ac_classzone_idx(ac), alloc_flags)) {
3184 /* Checked here to keep the fast path fast */
3185 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3186 if (alloc_flags & ALLOC_NO_WATERMARKS)
3189 if (node_reclaim_mode == 0 ||
3190 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3193 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3195 case NODE_RECLAIM_NOSCAN:
3198 case NODE_RECLAIM_FULL:
3199 /* scanned but unreclaimable */
3202 /* did we reclaim enough */
3203 if (zone_watermark_ok(zone, order, mark,
3204 ac_classzone_idx(ac), alloc_flags))
3212 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3213 gfp_mask, alloc_flags, ac->migratetype);
3215 prep_new_page(page, order, gfp_mask, alloc_flags);
3218 * If this is a high-order atomic allocation then check
3219 * if the pageblock should be reserved for the future
3221 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3222 reserve_highatomic_pageblock(page, zone, order);
3232 * Large machines with many possible nodes should not always dump per-node
3233 * meminfo in irq context.
3235 static inline bool should_suppress_show_mem(void)
3240 ret = in_interrupt();
3245 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3247 unsigned int filter = SHOW_MEM_FILTER_NODES;
3248 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3250 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3254 * This documents exceptions given to allocations in certain
3255 * contexts that are allowed to allocate outside current's set
3258 if (!(gfp_mask & __GFP_NOMEMALLOC))
3259 if (tsk_is_oom_victim(current) ||
3260 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3261 filter &= ~SHOW_MEM_FILTER_NODES;
3262 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3263 filter &= ~SHOW_MEM_FILTER_NODES;
3265 show_mem(filter, nodemask);
3268 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3270 struct va_format vaf;
3272 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3273 DEFAULT_RATELIMIT_BURST);
3275 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3278 va_start(args, fmt);
3281 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3282 current->comm, &vaf, gfp_mask, &gfp_mask,
3283 nodemask_pr_args(nodemask));
3286 cpuset_print_current_mems_allowed();
3289 warn_alloc_show_mem(gfp_mask, nodemask);
3292 static inline struct page *
3293 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3294 unsigned int alloc_flags,
3295 const struct alloc_context *ac)
3299 page = get_page_from_freelist(gfp_mask, order,
3300 alloc_flags|ALLOC_CPUSET, ac);
3302 * fallback to ignore cpuset restriction if our nodes
3306 page = get_page_from_freelist(gfp_mask, order,
3312 static inline struct page *
3313 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3314 const struct alloc_context *ac, unsigned long *did_some_progress)
3316 struct oom_control oc = {
3317 .zonelist = ac->zonelist,
3318 .nodemask = ac->nodemask,
3320 .gfp_mask = gfp_mask,
3325 *did_some_progress = 0;
3328 * Acquire the oom lock. If that fails, somebody else is
3329 * making progress for us.
3331 if (!mutex_trylock(&oom_lock)) {
3332 *did_some_progress = 1;
3333 schedule_timeout_uninterruptible(1);
3338 * Go through the zonelist yet one more time, keep very high watermark
3339 * here, this is only to catch a parallel oom killing, we must fail if
3340 * we're still under heavy pressure. But make sure that this reclaim
3341 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3342 * allocation which will never fail due to oom_lock already held.
3344 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3345 ~__GFP_DIRECT_RECLAIM, order,
3346 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3350 /* Coredumps can quickly deplete all memory reserves */
3351 if (current->flags & PF_DUMPCORE)
3353 /* The OOM killer will not help higher order allocs */
3354 if (order > PAGE_ALLOC_COSTLY_ORDER)
3357 * We have already exhausted all our reclaim opportunities without any
3358 * success so it is time to admit defeat. We will skip the OOM killer
3359 * because it is very likely that the caller has a more reasonable
3360 * fallback than shooting a random task.
3362 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3364 /* The OOM killer does not needlessly kill tasks for lowmem */
3365 if (ac->high_zoneidx < ZONE_NORMAL)
3367 if (pm_suspended_storage())
3370 * XXX: GFP_NOFS allocations should rather fail than rely on
3371 * other request to make a forward progress.
3372 * We are in an unfortunate situation where out_of_memory cannot
3373 * do much for this context but let's try it to at least get
3374 * access to memory reserved if the current task is killed (see
3375 * out_of_memory). Once filesystems are ready to handle allocation
3376 * failures more gracefully we should just bail out here.
3379 /* The OOM killer may not free memory on a specific node */
3380 if (gfp_mask & __GFP_THISNODE)
3383 /* Exhausted what can be done so it's blamo time */
3384 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3385 *did_some_progress = 1;
3388 * Help non-failing allocations by giving them access to memory
3391 if (gfp_mask & __GFP_NOFAIL)
3392 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3393 ALLOC_NO_WATERMARKS, ac);
3396 mutex_unlock(&oom_lock);
3401 * Maximum number of compaction retries wit a progress before OOM
3402 * killer is consider as the only way to move forward.
3404 #define MAX_COMPACT_RETRIES 16
3406 #ifdef CONFIG_COMPACTION
3407 /* Try memory compaction for high-order allocations before reclaim */
3408 static struct page *
3409 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3410 unsigned int alloc_flags, const struct alloc_context *ac,
3411 enum compact_priority prio, enum compact_result *compact_result)
3414 unsigned int noreclaim_flag;
3419 noreclaim_flag = memalloc_noreclaim_save();
3420 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3422 memalloc_noreclaim_restore(noreclaim_flag);
3424 if (*compact_result <= COMPACT_INACTIVE)
3428 * At least in one zone compaction wasn't deferred or skipped, so let's
3429 * count a compaction stall
3431 count_vm_event(COMPACTSTALL);
3433 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3436 struct zone *zone = page_zone(page);
3438 zone->compact_blockskip_flush = false;
3439 compaction_defer_reset(zone, order, true);
3440 count_vm_event(COMPACTSUCCESS);
3445 * It's bad if compaction run occurs and fails. The most likely reason
3446 * is that pages exist, but not enough to satisfy watermarks.
3448 count_vm_event(COMPACTFAIL);
3456 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3457 enum compact_result compact_result,
3458 enum compact_priority *compact_priority,
3459 int *compaction_retries)
3461 int max_retries = MAX_COMPACT_RETRIES;
3464 int retries = *compaction_retries;
3465 enum compact_priority priority = *compact_priority;
3470 if (compaction_made_progress(compact_result))
3471 (*compaction_retries)++;
3474 * compaction considers all the zone as desperately out of memory
3475 * so it doesn't really make much sense to retry except when the
3476 * failure could be caused by insufficient priority
3478 if (compaction_failed(compact_result))
3479 goto check_priority;
3482 * make sure the compaction wasn't deferred or didn't bail out early
3483 * due to locks contention before we declare that we should give up.
3484 * But do not retry if the given zonelist is not suitable for
3487 if (compaction_withdrawn(compact_result)) {
3488 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3493 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3494 * costly ones because they are de facto nofail and invoke OOM
3495 * killer to move on while costly can fail and users are ready
3496 * to cope with that. 1/4 retries is rather arbitrary but we
3497 * would need much more detailed feedback from compaction to
3498 * make a better decision.
3500 if (order > PAGE_ALLOC_COSTLY_ORDER)
3502 if (*compaction_retries <= max_retries) {
3508 * Make sure there are attempts at the highest priority if we exhausted
3509 * all retries or failed at the lower priorities.
3512 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3513 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3515 if (*compact_priority > min_priority) {
3516 (*compact_priority)--;
3517 *compaction_retries = 0;
3521 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3525 static inline struct page *
3526 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3527 unsigned int alloc_flags, const struct alloc_context *ac,
3528 enum compact_priority prio, enum compact_result *compact_result)
3530 *compact_result = COMPACT_SKIPPED;
3535 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3536 enum compact_result compact_result,
3537 enum compact_priority *compact_priority,
3538 int *compaction_retries)
3543 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3547 * There are setups with compaction disabled which would prefer to loop
3548 * inside the allocator rather than hit the oom killer prematurely.
3549 * Let's give them a good hope and keep retrying while the order-0
3550 * watermarks are OK.
3552 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3554 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3555 ac_classzone_idx(ac), alloc_flags))
3560 #endif /* CONFIG_COMPACTION */
3562 #ifdef CONFIG_LOCKDEP
3563 struct lockdep_map __fs_reclaim_map =
3564 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3566 static bool __need_fs_reclaim(gfp_t gfp_mask)
3568 gfp_mask = current_gfp_context(gfp_mask);
3570 /* no reclaim without waiting on it */
3571 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3574 /* this guy won't enter reclaim */
3575 if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
3578 /* We're only interested __GFP_FS allocations for now */
3579 if (!(gfp_mask & __GFP_FS))
3582 if (gfp_mask & __GFP_NOLOCKDEP)
3588 void fs_reclaim_acquire(gfp_t gfp_mask)
3590 if (__need_fs_reclaim(gfp_mask))
3591 lock_map_acquire(&__fs_reclaim_map);
3593 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3595 void fs_reclaim_release(gfp_t gfp_mask)
3597 if (__need_fs_reclaim(gfp_mask))
3598 lock_map_release(&__fs_reclaim_map);
3600 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3603 /* Perform direct synchronous page reclaim */
3605 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3606 const struct alloc_context *ac)
3608 struct reclaim_state reclaim_state;
3610 unsigned int noreclaim_flag;
3614 /* We now go into synchronous reclaim */
3615 cpuset_memory_pressure_bump();
3616 noreclaim_flag = memalloc_noreclaim_save();
3617 fs_reclaim_acquire(gfp_mask);
3618 reclaim_state.reclaimed_slab = 0;
3619 current->reclaim_state = &reclaim_state;
3621 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3624 current->reclaim_state = NULL;
3625 fs_reclaim_release(gfp_mask);
3626 memalloc_noreclaim_restore(noreclaim_flag);
3633 /* The really slow allocator path where we enter direct reclaim */
3634 static inline struct page *
3635 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3636 unsigned int alloc_flags, const struct alloc_context *ac,
3637 unsigned long *did_some_progress)
3639 struct page *page = NULL;
3640 bool drained = false;
3642 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3643 if (unlikely(!(*did_some_progress)))
3647 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3650 * If an allocation failed after direct reclaim, it could be because
3651 * pages are pinned on the per-cpu lists or in high alloc reserves.
3652 * Shrink them them and try again
3654 if (!page && !drained) {
3655 unreserve_highatomic_pageblock(ac, false);
3656 drain_all_pages(NULL);
3664 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3668 pg_data_t *last_pgdat = NULL;
3670 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3671 ac->high_zoneidx, ac->nodemask) {
3672 if (last_pgdat != zone->zone_pgdat)
3673 wakeup_kswapd(zone, order, ac->high_zoneidx);
3674 last_pgdat = zone->zone_pgdat;
3678 static inline unsigned int
3679 gfp_to_alloc_flags(gfp_t gfp_mask)
3681 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3683 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3684 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3687 * The caller may dip into page reserves a bit more if the caller
3688 * cannot run direct reclaim, or if the caller has realtime scheduling
3689 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3690 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3692 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3694 if (gfp_mask & __GFP_ATOMIC) {
3696 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3697 * if it can't schedule.
3699 if (!(gfp_mask & __GFP_NOMEMALLOC))
3700 alloc_flags |= ALLOC_HARDER;
3702 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3703 * comment for __cpuset_node_allowed().
3705 alloc_flags &= ~ALLOC_CPUSET;
3706 } else if (unlikely(rt_task(current)) && !in_interrupt())
3707 alloc_flags |= ALLOC_HARDER;
3710 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3711 alloc_flags |= ALLOC_CMA;
3716 static bool oom_reserves_allowed(struct task_struct *tsk)
3718 if (!tsk_is_oom_victim(tsk))
3722 * !MMU doesn't have oom reaper so give access to memory reserves
3723 * only to the thread with TIF_MEMDIE set
3725 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3732 * Distinguish requests which really need access to full memory
3733 * reserves from oom victims which can live with a portion of it
3735 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3737 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3739 if (gfp_mask & __GFP_MEMALLOC)
3740 return ALLOC_NO_WATERMARKS;
3741 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3742 return ALLOC_NO_WATERMARKS;
3743 if (!in_interrupt()) {
3744 if (current->flags & PF_MEMALLOC)
3745 return ALLOC_NO_WATERMARKS;
3746 else if (oom_reserves_allowed(current))
3753 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3755 return !!__gfp_pfmemalloc_flags(gfp_mask);
3759 * Checks whether it makes sense to retry the reclaim to make a forward progress
3760 * for the given allocation request.
3762 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3763 * without success, or when we couldn't even meet the watermark if we
3764 * reclaimed all remaining pages on the LRU lists.
3766 * Returns true if a retry is viable or false to enter the oom path.
3769 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3770 struct alloc_context *ac, int alloc_flags,
3771 bool did_some_progress, int *no_progress_loops)
3777 * Costly allocations might have made a progress but this doesn't mean
3778 * their order will become available due to high fragmentation so
3779 * always increment the no progress counter for them
3781 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3782 *no_progress_loops = 0;
3784 (*no_progress_loops)++;
3787 * Make sure we converge to OOM if we cannot make any progress
3788 * several times in the row.
3790 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3791 /* Before OOM, exhaust highatomic_reserve */
3792 return unreserve_highatomic_pageblock(ac, true);
3796 * Keep reclaiming pages while there is a chance this will lead
3797 * somewhere. If none of the target zones can satisfy our allocation
3798 * request even if all reclaimable pages are considered then we are
3799 * screwed and have to go OOM.
3801 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3803 unsigned long available;
3804 unsigned long reclaimable;
3805 unsigned long min_wmark = min_wmark_pages(zone);
3808 available = reclaimable = zone_reclaimable_pages(zone);
3809 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3812 * Would the allocation succeed if we reclaimed all
3813 * reclaimable pages?
3815 wmark = __zone_watermark_ok(zone, order, min_wmark,
3816 ac_classzone_idx(ac), alloc_flags, available);
3817 trace_reclaim_retry_zone(z, order, reclaimable,
3818 available, min_wmark, *no_progress_loops, wmark);
3821 * If we didn't make any progress and have a lot of
3822 * dirty + writeback pages then we should wait for
3823 * an IO to complete to slow down the reclaim and
3824 * prevent from pre mature OOM
3826 if (!did_some_progress) {
3827 unsigned long write_pending;
3829 write_pending = zone_page_state_snapshot(zone,
3830 NR_ZONE_WRITE_PENDING);
3832 if (2 * write_pending > reclaimable) {
3833 congestion_wait(BLK_RW_ASYNC, HZ/10);
3839 * Memory allocation/reclaim might be called from a WQ
3840 * context and the current implementation of the WQ
3841 * concurrency control doesn't recognize that
3842 * a particular WQ is congested if the worker thread is
3843 * looping without ever sleeping. Therefore we have to
3844 * do a short sleep here rather than calling
3847 if (current->flags & PF_WQ_WORKER)
3848 schedule_timeout_uninterruptible(1);
3860 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3863 * It's possible that cpuset's mems_allowed and the nodemask from
3864 * mempolicy don't intersect. This should be normally dealt with by
3865 * policy_nodemask(), but it's possible to race with cpuset update in
3866 * such a way the check therein was true, and then it became false
3867 * before we got our cpuset_mems_cookie here.
3868 * This assumes that for all allocations, ac->nodemask can come only
3869 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3870 * when it does not intersect with the cpuset restrictions) or the
3871 * caller can deal with a violated nodemask.
3873 if (cpusets_enabled() && ac->nodemask &&
3874 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3875 ac->nodemask = NULL;
3880 * When updating a task's mems_allowed or mempolicy nodemask, it is
3881 * possible to race with parallel threads in such a way that our
3882 * allocation can fail while the mask is being updated. If we are about
3883 * to fail, check if the cpuset changed during allocation and if so,
3886 if (read_mems_allowed_retry(cpuset_mems_cookie))
3892 static inline struct page *
3893 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3894 struct alloc_context *ac)
3896 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3897 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3898 struct page *page = NULL;
3899 unsigned int alloc_flags;
3900 unsigned long did_some_progress;
3901 enum compact_priority compact_priority;
3902 enum compact_result compact_result;
3903 int compaction_retries;
3904 int no_progress_loops;
3905 unsigned int cpuset_mems_cookie;
3909 * In the slowpath, we sanity check order to avoid ever trying to
3910 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3911 * be using allocators in order of preference for an area that is
3914 if (order >= MAX_ORDER) {
3915 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3920 * We also sanity check to catch abuse of atomic reserves being used by
3921 * callers that are not in atomic context.
3923 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3924 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3925 gfp_mask &= ~__GFP_ATOMIC;
3928 compaction_retries = 0;
3929 no_progress_loops = 0;
3930 compact_priority = DEF_COMPACT_PRIORITY;
3931 cpuset_mems_cookie = read_mems_allowed_begin();
3934 * The fast path uses conservative alloc_flags to succeed only until
3935 * kswapd needs to be woken up, and to avoid the cost of setting up
3936 * alloc_flags precisely. So we do that now.
3938 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3941 * We need to recalculate the starting point for the zonelist iterator
3942 * because we might have used different nodemask in the fast path, or
3943 * there was a cpuset modification and we are retrying - otherwise we
3944 * could end up iterating over non-eligible zones endlessly.
3946 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3947 ac->high_zoneidx, ac->nodemask);
3948 if (!ac->preferred_zoneref->zone)
3951 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3952 wake_all_kswapds(order, ac);
3955 * The adjusted alloc_flags might result in immediate success, so try
3958 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3963 * For costly allocations, try direct compaction first, as it's likely
3964 * that we have enough base pages and don't need to reclaim. For non-
3965 * movable high-order allocations, do that as well, as compaction will
3966 * try prevent permanent fragmentation by migrating from blocks of the
3968 * Don't try this for allocations that are allowed to ignore
3969 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3971 if (can_direct_reclaim &&
3973 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3974 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3975 page = __alloc_pages_direct_compact(gfp_mask, order,
3977 INIT_COMPACT_PRIORITY,
3983 * Checks for costly allocations with __GFP_NORETRY, which
3984 * includes THP page fault allocations
3986 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3988 * If compaction is deferred for high-order allocations,
3989 * it is because sync compaction recently failed. If
3990 * this is the case and the caller requested a THP
3991 * allocation, we do not want to heavily disrupt the
3992 * system, so we fail the allocation instead of entering
3995 if (compact_result == COMPACT_DEFERRED)
3999 * Looks like reclaim/compaction is worth trying, but
4000 * sync compaction could be very expensive, so keep
4001 * using async compaction.
4003 compact_priority = INIT_COMPACT_PRIORITY;
4008 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4009 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4010 wake_all_kswapds(order, ac);
4012 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4014 alloc_flags = reserve_flags;
4017 * Reset the zonelist iterators if memory policies can be ignored.
4018 * These allocations are high priority and system rather than user
4021 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4022 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4023 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4024 ac->high_zoneidx, ac->nodemask);
4027 /* Attempt with potentially adjusted zonelist and alloc_flags */
4028 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4032 /* Caller is not willing to reclaim, we can't balance anything */
4033 if (!can_direct_reclaim)
4036 /* Avoid recursion of direct reclaim */
4037 if (current->flags & PF_MEMALLOC)
4040 /* Try direct reclaim and then allocating */
4041 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4042 &did_some_progress);
4046 /* Try direct compaction and then allocating */
4047 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4048 compact_priority, &compact_result);
4052 /* Do not loop if specifically requested */
4053 if (gfp_mask & __GFP_NORETRY)
4057 * Do not retry costly high order allocations unless they are
4058 * __GFP_RETRY_MAYFAIL
4060 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4063 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4064 did_some_progress > 0, &no_progress_loops))
4068 * It doesn't make any sense to retry for the compaction if the order-0
4069 * reclaim is not able to make any progress because the current
4070 * implementation of the compaction depends on the sufficient amount
4071 * of free memory (see __compaction_suitable)
4073 if (did_some_progress > 0 &&
4074 should_compact_retry(ac, order, alloc_flags,
4075 compact_result, &compact_priority,
4076 &compaction_retries))
4080 /* Deal with possible cpuset update races before we start OOM killing */
4081 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4084 /* Reclaim has failed us, start killing things */
4085 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4089 /* Avoid allocations with no watermarks from looping endlessly */
4090 if (tsk_is_oom_victim(current) &&
4091 (alloc_flags == ALLOC_OOM ||
4092 (gfp_mask & __GFP_NOMEMALLOC)))
4095 /* Retry as long as the OOM killer is making progress */
4096 if (did_some_progress) {
4097 no_progress_loops = 0;
4102 /* Deal with possible cpuset update races before we fail */
4103 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4107 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4110 if (gfp_mask & __GFP_NOFAIL) {
4112 * All existing users of the __GFP_NOFAIL are blockable, so warn
4113 * of any new users that actually require GFP_NOWAIT
4115 if (WARN_ON_ONCE(!can_direct_reclaim))
4119 * PF_MEMALLOC request from this context is rather bizarre
4120 * because we cannot reclaim anything and only can loop waiting
4121 * for somebody to do a work for us
4123 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4126 * non failing costly orders are a hard requirement which we
4127 * are not prepared for much so let's warn about these users
4128 * so that we can identify them and convert them to something
4131 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4134 * Help non-failing allocations by giving them access to memory
4135 * reserves but do not use ALLOC_NO_WATERMARKS because this
4136 * could deplete whole memory reserves which would just make
4137 * the situation worse
4139 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4147 warn_alloc(gfp_mask, ac->nodemask,
4148 "page allocation failure: order:%u", order);
4153 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4154 int preferred_nid, nodemask_t *nodemask,
4155 struct alloc_context *ac, gfp_t *alloc_mask,
4156 unsigned int *alloc_flags)
4158 ac->high_zoneidx = gfp_zone(gfp_mask);
4159 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4160 ac->nodemask = nodemask;
4161 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4163 if (cpusets_enabled()) {
4164 *alloc_mask |= __GFP_HARDWALL;
4166 ac->nodemask = &cpuset_current_mems_allowed;
4168 *alloc_flags |= ALLOC_CPUSET;
4171 fs_reclaim_acquire(gfp_mask);
4172 fs_reclaim_release(gfp_mask);
4174 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4176 if (should_fail_alloc_page(gfp_mask, order))
4179 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4180 *alloc_flags |= ALLOC_CMA;
4185 /* Determine whether to spread dirty pages and what the first usable zone */
4186 static inline void finalise_ac(gfp_t gfp_mask,
4187 unsigned int order, struct alloc_context *ac)
4189 /* Dirty zone balancing only done in the fast path */
4190 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4193 * The preferred zone is used for statistics but crucially it is
4194 * also used as the starting point for the zonelist iterator. It
4195 * may get reset for allocations that ignore memory policies.
4197 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4198 ac->high_zoneidx, ac->nodemask);
4202 * This is the 'heart' of the zoned buddy allocator.
4205 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4206 nodemask_t *nodemask)
4209 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4210 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4211 struct alloc_context ac = { };
4213 gfp_mask &= gfp_allowed_mask;
4214 alloc_mask = gfp_mask;
4215 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4218 finalise_ac(gfp_mask, order, &ac);
4220 /* First allocation attempt */
4221 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4226 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4227 * resp. GFP_NOIO which has to be inherited for all allocation requests
4228 * from a particular context which has been marked by
4229 * memalloc_no{fs,io}_{save,restore}.
4231 alloc_mask = current_gfp_context(gfp_mask);
4232 ac.spread_dirty_pages = false;
4235 * Restore the original nodemask if it was potentially replaced with
4236 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4238 if (unlikely(ac.nodemask != nodemask))
4239 ac.nodemask = nodemask;
4241 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4244 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4245 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4246 __free_pages(page, order);
4250 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4254 EXPORT_SYMBOL(__alloc_pages_nodemask);
4257 * Common helper functions.
4259 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4264 * __get_free_pages() returns a 32-bit address, which cannot represent
4267 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4269 page = alloc_pages(gfp_mask, order);
4272 return (unsigned long) page_address(page);
4274 EXPORT_SYMBOL(__get_free_pages);
4276 unsigned long get_zeroed_page(gfp_t gfp_mask)
4278 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4280 EXPORT_SYMBOL(get_zeroed_page);
4282 void __free_pages(struct page *page, unsigned int order)
4284 if (put_page_testzero(page)) {
4286 free_unref_page(page);
4288 __free_pages_ok(page, order);
4292 EXPORT_SYMBOL(__free_pages);
4294 void free_pages(unsigned long addr, unsigned int order)
4297 VM_BUG_ON(!virt_addr_valid((void *)addr));
4298 __free_pages(virt_to_page((void *)addr), order);
4302 EXPORT_SYMBOL(free_pages);
4306 * An arbitrary-length arbitrary-offset area of memory which resides
4307 * within a 0 or higher order page. Multiple fragments within that page
4308 * are individually refcounted, in the page's reference counter.
4310 * The page_frag functions below provide a simple allocation framework for
4311 * page fragments. This is used by the network stack and network device
4312 * drivers to provide a backing region of memory for use as either an
4313 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4315 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4318 struct page *page = NULL;
4319 gfp_t gfp = gfp_mask;
4321 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4322 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4324 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4325 PAGE_FRAG_CACHE_MAX_ORDER);
4326 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4328 if (unlikely(!page))
4329 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4331 nc->va = page ? page_address(page) : NULL;
4336 void __page_frag_cache_drain(struct page *page, unsigned int count)
4338 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4340 if (page_ref_sub_and_test(page, count)) {
4341 unsigned int order = compound_order(page);
4344 free_unref_page(page);
4346 __free_pages_ok(page, order);
4349 EXPORT_SYMBOL(__page_frag_cache_drain);
4351 void *page_frag_alloc(struct page_frag_cache *nc,
4352 unsigned int fragsz, gfp_t gfp_mask)
4354 unsigned int size = PAGE_SIZE;
4358 if (unlikely(!nc->va)) {
4360 page = __page_frag_cache_refill(nc, gfp_mask);
4364 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4365 /* if size can vary use size else just use PAGE_SIZE */
4368 /* Even if we own the page, we do not use atomic_set().
4369 * This would break get_page_unless_zero() users.
4371 page_ref_add(page, size - 1);
4373 /* reset page count bias and offset to start of new frag */
4374 nc->pfmemalloc = page_is_pfmemalloc(page);
4375 nc->pagecnt_bias = size;
4379 offset = nc->offset - fragsz;
4380 if (unlikely(offset < 0)) {
4381 page = virt_to_page(nc->va);
4383 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4386 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4387 /* if size can vary use size else just use PAGE_SIZE */
4390 /* OK, page count is 0, we can safely set it */
4391 set_page_count(page, size);
4393 /* reset page count bias and offset to start of new frag */
4394 nc->pagecnt_bias = size;
4395 offset = size - fragsz;
4399 nc->offset = offset;
4401 return nc->va + offset;
4403 EXPORT_SYMBOL(page_frag_alloc);
4406 * Frees a page fragment allocated out of either a compound or order 0 page.
4408 void page_frag_free(void *addr)
4410 struct page *page = virt_to_head_page(addr);
4412 if (unlikely(put_page_testzero(page)))
4413 __free_pages_ok(page, compound_order(page));
4415 EXPORT_SYMBOL(page_frag_free);
4417 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4421 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4422 unsigned long used = addr + PAGE_ALIGN(size);
4424 split_page(virt_to_page((void *)addr), order);
4425 while (used < alloc_end) {
4430 return (void *)addr;
4434 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4435 * @size: the number of bytes to allocate
4436 * @gfp_mask: GFP flags for the allocation
4438 * This function is similar to alloc_pages(), except that it allocates the
4439 * minimum number of pages to satisfy the request. alloc_pages() can only
4440 * allocate memory in power-of-two pages.
4442 * This function is also limited by MAX_ORDER.
4444 * Memory allocated by this function must be released by free_pages_exact().
4446 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4448 unsigned int order = get_order(size);
4451 addr = __get_free_pages(gfp_mask, order);
4452 return make_alloc_exact(addr, order, size);
4454 EXPORT_SYMBOL(alloc_pages_exact);
4457 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4459 * @nid: the preferred node ID where memory should be allocated
4460 * @size: the number of bytes to allocate
4461 * @gfp_mask: GFP flags for the allocation
4463 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4466 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4468 unsigned int order = get_order(size);
4469 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4472 return make_alloc_exact((unsigned long)page_address(p), order, size);
4476 * free_pages_exact - release memory allocated via alloc_pages_exact()
4477 * @virt: the value returned by alloc_pages_exact.
4478 * @size: size of allocation, same value as passed to alloc_pages_exact().
4480 * Release the memory allocated by a previous call to alloc_pages_exact.
4482 void free_pages_exact(void *virt, size_t size)
4484 unsigned long addr = (unsigned long)virt;
4485 unsigned long end = addr + PAGE_ALIGN(size);
4487 while (addr < end) {
4492 EXPORT_SYMBOL(free_pages_exact);
4495 * nr_free_zone_pages - count number of pages beyond high watermark
4496 * @offset: The zone index of the highest zone
4498 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4499 * high watermark within all zones at or below a given zone index. For each
4500 * zone, the number of pages is calculated as:
4502 * nr_free_zone_pages = managed_pages - high_pages
4504 static unsigned long nr_free_zone_pages(int offset)
4509 /* Just pick one node, since fallback list is circular */
4510 unsigned long sum = 0;
4512 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4514 for_each_zone_zonelist(zone, z, zonelist, offset) {
4515 unsigned long size = zone->managed_pages;
4516 unsigned long high = high_wmark_pages(zone);
4525 * nr_free_buffer_pages - count number of pages beyond high watermark
4527 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4528 * watermark within ZONE_DMA and ZONE_NORMAL.
4530 unsigned long nr_free_buffer_pages(void)
4532 return nr_free_zone_pages(gfp_zone(GFP_USER));
4534 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4537 * nr_free_pagecache_pages - count number of pages beyond high watermark
4539 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4540 * high watermark within all zones.
4542 unsigned long nr_free_pagecache_pages(void)
4544 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4547 static inline void show_node(struct zone *zone)
4549 if (IS_ENABLED(CONFIG_NUMA))
4550 printk("Node %d ", zone_to_nid(zone));
4553 long si_mem_available(void)
4556 unsigned long pagecache;
4557 unsigned long wmark_low = 0;
4558 unsigned long pages[NR_LRU_LISTS];
4562 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4563 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4566 wmark_low += zone->watermark[WMARK_LOW];
4569 * Estimate the amount of memory available for userspace allocations,
4570 * without causing swapping.
4572 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4575 * Not all the page cache can be freed, otherwise the system will
4576 * start swapping. Assume at least half of the page cache, or the
4577 * low watermark worth of cache, needs to stay.
4579 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4580 pagecache -= min(pagecache / 2, wmark_low);
4581 available += pagecache;
4584 * Part of the reclaimable slab consists of items that are in use,
4585 * and cannot be freed. Cap this estimate at the low watermark.
4587 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4588 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4595 EXPORT_SYMBOL_GPL(si_mem_available);
4597 void si_meminfo(struct sysinfo *val)
4599 val->totalram = totalram_pages;
4600 val->sharedram = global_node_page_state(NR_SHMEM);
4601 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4602 val->bufferram = nr_blockdev_pages();
4603 val->totalhigh = totalhigh_pages;
4604 val->freehigh = nr_free_highpages();
4605 val->mem_unit = PAGE_SIZE;
4608 EXPORT_SYMBOL(si_meminfo);
4611 void si_meminfo_node(struct sysinfo *val, int nid)
4613 int zone_type; /* needs to be signed */
4614 unsigned long managed_pages = 0;
4615 unsigned long managed_highpages = 0;
4616 unsigned long free_highpages = 0;
4617 pg_data_t *pgdat = NODE_DATA(nid);
4619 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4620 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4621 val->totalram = managed_pages;
4622 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4623 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4624 #ifdef CONFIG_HIGHMEM
4625 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4626 struct zone *zone = &pgdat->node_zones[zone_type];
4628 if (is_highmem(zone)) {
4629 managed_highpages += zone->managed_pages;
4630 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4633 val->totalhigh = managed_highpages;
4634 val->freehigh = free_highpages;
4636 val->totalhigh = managed_highpages;
4637 val->freehigh = free_highpages;
4639 val->mem_unit = PAGE_SIZE;
4644 * Determine whether the node should be displayed or not, depending on whether
4645 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4647 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4649 if (!(flags & SHOW_MEM_FILTER_NODES))
4653 * no node mask - aka implicit memory numa policy. Do not bother with
4654 * the synchronization - read_mems_allowed_begin - because we do not
4655 * have to be precise here.
4658 nodemask = &cpuset_current_mems_allowed;
4660 return !node_isset(nid, *nodemask);
4663 #define K(x) ((x) << (PAGE_SHIFT-10))
4665 static void show_migration_types(unsigned char type)
4667 static const char types[MIGRATE_TYPES] = {
4668 [MIGRATE_UNMOVABLE] = 'U',
4669 [MIGRATE_MOVABLE] = 'M',
4670 [MIGRATE_RECLAIMABLE] = 'E',
4671 [MIGRATE_HIGHATOMIC] = 'H',
4673 [MIGRATE_CMA] = 'C',
4675 #ifdef CONFIG_MEMORY_ISOLATION
4676 [MIGRATE_ISOLATE] = 'I',
4679 char tmp[MIGRATE_TYPES + 1];
4683 for (i = 0; i < MIGRATE_TYPES; i++) {
4684 if (type & (1 << i))
4689 printk(KERN_CONT "(%s) ", tmp);
4693 * Show free area list (used inside shift_scroll-lock stuff)
4694 * We also calculate the percentage fragmentation. We do this by counting the
4695 * memory on each free list with the exception of the first item on the list.
4698 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4701 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4703 unsigned long free_pcp = 0;
4708 for_each_populated_zone(zone) {
4709 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4712 for_each_online_cpu(cpu)
4713 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4716 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4717 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4718 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4719 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4720 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4721 " free:%lu free_pcp:%lu free_cma:%lu\n",
4722 global_node_page_state(NR_ACTIVE_ANON),
4723 global_node_page_state(NR_INACTIVE_ANON),
4724 global_node_page_state(NR_ISOLATED_ANON),
4725 global_node_page_state(NR_ACTIVE_FILE),
4726 global_node_page_state(NR_INACTIVE_FILE),
4727 global_node_page_state(NR_ISOLATED_FILE),
4728 global_node_page_state(NR_UNEVICTABLE),
4729 global_node_page_state(NR_FILE_DIRTY),
4730 global_node_page_state(NR_WRITEBACK),
4731 global_node_page_state(NR_UNSTABLE_NFS),
4732 global_node_page_state(NR_SLAB_RECLAIMABLE),
4733 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4734 global_node_page_state(NR_FILE_MAPPED),
4735 global_node_page_state(NR_SHMEM),
4736 global_zone_page_state(NR_PAGETABLE),
4737 global_zone_page_state(NR_BOUNCE),
4738 global_zone_page_state(NR_FREE_PAGES),
4740 global_zone_page_state(NR_FREE_CMA_PAGES));
4742 for_each_online_pgdat(pgdat) {
4743 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4747 " active_anon:%lukB"
4748 " inactive_anon:%lukB"
4749 " active_file:%lukB"
4750 " inactive_file:%lukB"
4751 " unevictable:%lukB"
4752 " isolated(anon):%lukB"
4753 " isolated(file):%lukB"
4758 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4760 " shmem_pmdmapped: %lukB"
4763 " writeback_tmp:%lukB"
4765 " all_unreclaimable? %s"
4768 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4769 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4770 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4771 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4772 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4773 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4774 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4775 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4776 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4777 K(node_page_state(pgdat, NR_WRITEBACK)),
4778 K(node_page_state(pgdat, NR_SHMEM)),
4779 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4780 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4781 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4783 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4785 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4786 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4787 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4791 for_each_populated_zone(zone) {
4794 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4798 for_each_online_cpu(cpu)
4799 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4808 " active_anon:%lukB"
4809 " inactive_anon:%lukB"
4810 " active_file:%lukB"
4811 " inactive_file:%lukB"
4812 " unevictable:%lukB"
4813 " writepending:%lukB"
4817 " kernel_stack:%lukB"
4825 K(zone_page_state(zone, NR_FREE_PAGES)),
4826 K(min_wmark_pages(zone)),
4827 K(low_wmark_pages(zone)),
4828 K(high_wmark_pages(zone)),
4829 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4830 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4831 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4832 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4833 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4834 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4835 K(zone->present_pages),
4836 K(zone->managed_pages),
4837 K(zone_page_state(zone, NR_MLOCK)),
4838 zone_page_state(zone, NR_KERNEL_STACK_KB),
4839 K(zone_page_state(zone, NR_PAGETABLE)),
4840 K(zone_page_state(zone, NR_BOUNCE)),
4842 K(this_cpu_read(zone->pageset->pcp.count)),
4843 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4844 printk("lowmem_reserve[]:");
4845 for (i = 0; i < MAX_NR_ZONES; i++)
4846 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4847 printk(KERN_CONT "\n");
4850 for_each_populated_zone(zone) {
4852 unsigned long nr[MAX_ORDER], flags, total = 0;
4853 unsigned char types[MAX_ORDER];
4855 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4858 printk(KERN_CONT "%s: ", zone->name);
4860 spin_lock_irqsave(&zone->lock, flags);
4861 for (order = 0; order < MAX_ORDER; order++) {
4862 struct free_area *area = &zone->free_area[order];
4865 nr[order] = area->nr_free;
4866 total += nr[order] << order;
4869 for (type = 0; type < MIGRATE_TYPES; type++) {
4870 if (!list_empty(&area->free_list[type]))
4871 types[order] |= 1 << type;
4874 spin_unlock_irqrestore(&zone->lock, flags);
4875 for (order = 0; order < MAX_ORDER; order++) {
4876 printk(KERN_CONT "%lu*%lukB ",
4877 nr[order], K(1UL) << order);
4879 show_migration_types(types[order]);
4881 printk(KERN_CONT "= %lukB\n", K(total));
4884 hugetlb_show_meminfo();
4886 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4888 show_swap_cache_info();
4891 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4893 zoneref->zone = zone;
4894 zoneref->zone_idx = zone_idx(zone);
4898 * Builds allocation fallback zone lists.
4900 * Add all populated zones of a node to the zonelist.
4902 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4905 enum zone_type zone_type = MAX_NR_ZONES;
4910 zone = pgdat->node_zones + zone_type;
4911 if (managed_zone(zone)) {
4912 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4913 check_highest_zone(zone_type);
4915 } while (zone_type);
4922 static int __parse_numa_zonelist_order(char *s)
4925 * We used to support different zonlists modes but they turned
4926 * out to be just not useful. Let's keep the warning in place
4927 * if somebody still use the cmd line parameter so that we do
4928 * not fail it silently
4930 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4931 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4937 static __init int setup_numa_zonelist_order(char *s)
4942 return __parse_numa_zonelist_order(s);
4944 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4946 char numa_zonelist_order[] = "Node";
4949 * sysctl handler for numa_zonelist_order
4951 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4952 void __user *buffer, size_t *length,
4959 return proc_dostring(table, write, buffer, length, ppos);
4960 str = memdup_user_nul(buffer, 16);
4962 return PTR_ERR(str);
4964 ret = __parse_numa_zonelist_order(str);
4970 #define MAX_NODE_LOAD (nr_online_nodes)
4971 static int node_load[MAX_NUMNODES];
4974 * find_next_best_node - find the next node that should appear in a given node's fallback list
4975 * @node: node whose fallback list we're appending
4976 * @used_node_mask: nodemask_t of already used nodes
4978 * We use a number of factors to determine which is the next node that should
4979 * appear on a given node's fallback list. The node should not have appeared
4980 * already in @node's fallback list, and it should be the next closest node
4981 * according to the distance array (which contains arbitrary distance values
4982 * from each node to each node in the system), and should also prefer nodes
4983 * with no CPUs, since presumably they'll have very little allocation pressure
4984 * on them otherwise.
4985 * It returns -1 if no node is found.
4987 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4990 int min_val = INT_MAX;
4991 int best_node = NUMA_NO_NODE;
4992 const struct cpumask *tmp = cpumask_of_node(0);
4994 /* Use the local node if we haven't already */
4995 if (!node_isset(node, *used_node_mask)) {
4996 node_set(node, *used_node_mask);
5000 for_each_node_state(n, N_MEMORY) {
5002 /* Don't want a node to appear more than once */
5003 if (node_isset(n, *used_node_mask))
5006 /* Use the distance array to find the distance */
5007 val = node_distance(node, n);
5009 /* Penalize nodes under us ("prefer the next node") */
5012 /* Give preference to headless and unused nodes */
5013 tmp = cpumask_of_node(n);
5014 if (!cpumask_empty(tmp))
5015 val += PENALTY_FOR_NODE_WITH_CPUS;
5017 /* Slight preference for less loaded node */
5018 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5019 val += node_load[n];
5021 if (val < min_val) {
5028 node_set(best_node, *used_node_mask);
5035 * Build zonelists ordered by node and zones within node.
5036 * This results in maximum locality--normal zone overflows into local
5037 * DMA zone, if any--but risks exhausting DMA zone.
5039 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5042 struct zoneref *zonerefs;
5045 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5047 for (i = 0; i < nr_nodes; i++) {
5050 pg_data_t *node = NODE_DATA(node_order[i]);
5052 nr_zones = build_zonerefs_node(node, zonerefs);
5053 zonerefs += nr_zones;
5055 zonerefs->zone = NULL;
5056 zonerefs->zone_idx = 0;
5060 * Build gfp_thisnode zonelists
5062 static void build_thisnode_zonelists(pg_data_t *pgdat)
5064 struct zoneref *zonerefs;
5067 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5068 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5069 zonerefs += nr_zones;
5070 zonerefs->zone = NULL;
5071 zonerefs->zone_idx = 0;
5075 * Build zonelists ordered by zone and nodes within zones.
5076 * This results in conserving DMA zone[s] until all Normal memory is
5077 * exhausted, but results in overflowing to remote node while memory
5078 * may still exist in local DMA zone.
5081 static void build_zonelists(pg_data_t *pgdat)
5083 static int node_order[MAX_NUMNODES];
5084 int node, load, nr_nodes = 0;
5085 nodemask_t used_mask;
5086 int local_node, prev_node;
5088 /* NUMA-aware ordering of nodes */
5089 local_node = pgdat->node_id;
5090 load = nr_online_nodes;
5091 prev_node = local_node;
5092 nodes_clear(used_mask);
5094 memset(node_order, 0, sizeof(node_order));
5095 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5097 * We don't want to pressure a particular node.
5098 * So adding penalty to the first node in same
5099 * distance group to make it round-robin.
5101 if (node_distance(local_node, node) !=
5102 node_distance(local_node, prev_node))
5103 node_load[node] = load;
5105 node_order[nr_nodes++] = node;
5110 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5111 build_thisnode_zonelists(pgdat);
5114 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5116 * Return node id of node used for "local" allocations.
5117 * I.e., first node id of first zone in arg node's generic zonelist.
5118 * Used for initializing percpu 'numa_mem', which is used primarily
5119 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5121 int local_memory_node(int node)
5125 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5126 gfp_zone(GFP_KERNEL),
5128 return z->zone->node;
5132 static void setup_min_unmapped_ratio(void);
5133 static void setup_min_slab_ratio(void);
5134 #else /* CONFIG_NUMA */
5136 static void build_zonelists(pg_data_t *pgdat)
5138 int node, local_node;
5139 struct zoneref *zonerefs;
5142 local_node = pgdat->node_id;
5144 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5145 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5146 zonerefs += nr_zones;
5149 * Now we build the zonelist so that it contains the zones
5150 * of all the other nodes.
5151 * We don't want to pressure a particular node, so when
5152 * building the zones for node N, we make sure that the
5153 * zones coming right after the local ones are those from
5154 * node N+1 (modulo N)
5156 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5157 if (!node_online(node))
5159 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5160 zonerefs += nr_zones;
5162 for (node = 0; node < local_node; node++) {
5163 if (!node_online(node))
5165 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5166 zonerefs += nr_zones;
5169 zonerefs->zone = NULL;
5170 zonerefs->zone_idx = 0;
5173 #endif /* CONFIG_NUMA */
5176 * Boot pageset table. One per cpu which is going to be used for all
5177 * zones and all nodes. The parameters will be set in such a way
5178 * that an item put on a list will immediately be handed over to
5179 * the buddy list. This is safe since pageset manipulation is done
5180 * with interrupts disabled.
5182 * The boot_pagesets must be kept even after bootup is complete for
5183 * unused processors and/or zones. They do play a role for bootstrapping
5184 * hotplugged processors.
5186 * zoneinfo_show() and maybe other functions do
5187 * not check if the processor is online before following the pageset pointer.
5188 * Other parts of the kernel may not check if the zone is available.
5190 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5191 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5192 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5194 static void __build_all_zonelists(void *data)
5197 int __maybe_unused cpu;
5198 pg_data_t *self = data;
5199 static DEFINE_SPINLOCK(lock);
5204 memset(node_load, 0, sizeof(node_load));
5208 * This node is hotadded and no memory is yet present. So just
5209 * building zonelists is fine - no need to touch other nodes.
5211 if (self && !node_online(self->node_id)) {
5212 build_zonelists(self);
5214 for_each_online_node(nid) {
5215 pg_data_t *pgdat = NODE_DATA(nid);
5217 build_zonelists(pgdat);
5220 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5222 * We now know the "local memory node" for each node--
5223 * i.e., the node of the first zone in the generic zonelist.
5224 * Set up numa_mem percpu variable for on-line cpus. During
5225 * boot, only the boot cpu should be on-line; we'll init the
5226 * secondary cpus' numa_mem as they come on-line. During
5227 * node/memory hotplug, we'll fixup all on-line cpus.
5229 for_each_online_cpu(cpu)
5230 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5237 static noinline void __init
5238 build_all_zonelists_init(void)
5242 __build_all_zonelists(NULL);
5245 * Initialize the boot_pagesets that are going to be used
5246 * for bootstrapping processors. The real pagesets for
5247 * each zone will be allocated later when the per cpu
5248 * allocator is available.
5250 * boot_pagesets are used also for bootstrapping offline
5251 * cpus if the system is already booted because the pagesets
5252 * are needed to initialize allocators on a specific cpu too.
5253 * F.e. the percpu allocator needs the page allocator which
5254 * needs the percpu allocator in order to allocate its pagesets
5255 * (a chicken-egg dilemma).
5257 for_each_possible_cpu(cpu)
5258 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5260 mminit_verify_zonelist();
5261 cpuset_init_current_mems_allowed();
5265 * unless system_state == SYSTEM_BOOTING.
5267 * __ref due to call of __init annotated helper build_all_zonelists_init
5268 * [protected by SYSTEM_BOOTING].
5270 void __ref build_all_zonelists(pg_data_t *pgdat)
5272 if (system_state == SYSTEM_BOOTING) {
5273 build_all_zonelists_init();
5275 __build_all_zonelists(pgdat);
5276 /* cpuset refresh routine should be here */
5278 vm_total_pages = nr_free_pagecache_pages();
5280 * Disable grouping by mobility if the number of pages in the
5281 * system is too low to allow the mechanism to work. It would be
5282 * more accurate, but expensive to check per-zone. This check is
5283 * made on memory-hotadd so a system can start with mobility
5284 * disabled and enable it later
5286 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5287 page_group_by_mobility_disabled = 1;
5289 page_group_by_mobility_disabled = 0;
5291 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5293 page_group_by_mobility_disabled ? "off" : "on",
5296 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5301 * Initially all pages are reserved - free ones are freed
5302 * up by free_all_bootmem() once the early boot process is
5303 * done. Non-atomic initialization, single-pass.
5305 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5306 unsigned long start_pfn, enum memmap_context context)
5308 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5309 unsigned long end_pfn = start_pfn + size;
5310 pg_data_t *pgdat = NODE_DATA(nid);
5312 unsigned long nr_initialised = 0;
5313 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5314 struct memblock_region *r = NULL, *tmp;
5317 if (highest_memmap_pfn < end_pfn - 1)
5318 highest_memmap_pfn = end_pfn - 1;
5321 * Honor reservation requested by the driver for this ZONE_DEVICE
5324 if (altmap && start_pfn == altmap->base_pfn)
5325 start_pfn += altmap->reserve;
5327 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5329 * There can be holes in boot-time mem_map[]s handed to this
5330 * function. They do not exist on hotplugged memory.
5332 if (context != MEMMAP_EARLY)
5335 if (!early_pfn_valid(pfn)) {
5336 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5338 * Skip to the pfn preceding the next valid one (or
5339 * end_pfn), such that we hit a valid pfn (or end_pfn)
5340 * on our next iteration of the loop.
5342 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5346 if (!early_pfn_in_nid(pfn, nid))
5348 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5351 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5353 * Check given memblock attribute by firmware which can affect
5354 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5355 * mirrored, it's an overlapped memmap init. skip it.
5357 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5358 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5359 for_each_memblock(memory, tmp)
5360 if (pfn < memblock_region_memory_end_pfn(tmp))
5364 if (pfn >= memblock_region_memory_base_pfn(r) &&
5365 memblock_is_mirror(r)) {
5366 /* already initialized as NORMAL */
5367 pfn = memblock_region_memory_end_pfn(r);
5375 * Mark the block movable so that blocks are reserved for
5376 * movable at startup. This will force kernel allocations
5377 * to reserve their blocks rather than leaking throughout
5378 * the address space during boot when many long-lived
5379 * kernel allocations are made.
5381 * bitmap is created for zone's valid pfn range. but memmap
5382 * can be created for invalid pages (for alignment)
5383 * check here not to call set_pageblock_migratetype() against
5386 if (!(pfn & (pageblock_nr_pages - 1))) {
5387 struct page *page = pfn_to_page(pfn);
5389 __init_single_page(page, pfn, zone, nid);
5390 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5393 __init_single_pfn(pfn, zone, nid);
5398 static void __meminit zone_init_free_lists(struct zone *zone)
5400 unsigned int order, t;
5401 for_each_migratetype_order(order, t) {
5402 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5403 zone->free_area[order].nr_free = 0;
5407 #ifndef __HAVE_ARCH_MEMMAP_INIT
5408 #define memmap_init(size, nid, zone, start_pfn) \
5409 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5412 static int zone_batchsize(struct zone *zone)
5418 * The per-cpu-pages pools are set to around 1000th of the
5419 * size of the zone. But no more than 1/2 of a meg.
5421 * OK, so we don't know how big the cache is. So guess.
5423 batch = zone->managed_pages / 1024;
5424 if (batch * PAGE_SIZE > 512 * 1024)
5425 batch = (512 * 1024) / PAGE_SIZE;
5426 batch /= 4; /* We effectively *= 4 below */
5431 * Clamp the batch to a 2^n - 1 value. Having a power
5432 * of 2 value was found to be more likely to have
5433 * suboptimal cache aliasing properties in some cases.
5435 * For example if 2 tasks are alternately allocating
5436 * batches of pages, one task can end up with a lot
5437 * of pages of one half of the possible page colors
5438 * and the other with pages of the other colors.
5440 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5445 /* The deferral and batching of frees should be suppressed under NOMMU
5448 * The problem is that NOMMU needs to be able to allocate large chunks
5449 * of contiguous memory as there's no hardware page translation to
5450 * assemble apparent contiguous memory from discontiguous pages.
5452 * Queueing large contiguous runs of pages for batching, however,
5453 * causes the pages to actually be freed in smaller chunks. As there
5454 * can be a significant delay between the individual batches being
5455 * recycled, this leads to the once large chunks of space being
5456 * fragmented and becoming unavailable for high-order allocations.
5463 * pcp->high and pcp->batch values are related and dependent on one another:
5464 * ->batch must never be higher then ->high.
5465 * The following function updates them in a safe manner without read side
5468 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5469 * those fields changing asynchronously (acording the the above rule).
5471 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5472 * outside of boot time (or some other assurance that no concurrent updaters
5475 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5476 unsigned long batch)
5478 /* start with a fail safe value for batch */
5482 /* Update high, then batch, in order */
5489 /* a companion to pageset_set_high() */
5490 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5492 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5495 static void pageset_init(struct per_cpu_pageset *p)
5497 struct per_cpu_pages *pcp;
5500 memset(p, 0, sizeof(*p));
5504 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5505 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5508 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5511 pageset_set_batch(p, batch);
5515 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5516 * to the value high for the pageset p.
5518 static void pageset_set_high(struct per_cpu_pageset *p,
5521 unsigned long batch = max(1UL, high / 4);
5522 if ((high / 4) > (PAGE_SHIFT * 8))
5523 batch = PAGE_SHIFT * 8;
5525 pageset_update(&p->pcp, high, batch);
5528 static void pageset_set_high_and_batch(struct zone *zone,
5529 struct per_cpu_pageset *pcp)
5531 if (percpu_pagelist_fraction)
5532 pageset_set_high(pcp,
5533 (zone->managed_pages /
5534 percpu_pagelist_fraction));
5536 pageset_set_batch(pcp, zone_batchsize(zone));
5539 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5541 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5544 pageset_set_high_and_batch(zone, pcp);
5547 void __meminit setup_zone_pageset(struct zone *zone)
5550 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5551 for_each_possible_cpu(cpu)
5552 zone_pageset_init(zone, cpu);
5556 * Allocate per cpu pagesets and initialize them.
5557 * Before this call only boot pagesets were available.
5559 void __init setup_per_cpu_pageset(void)
5561 struct pglist_data *pgdat;
5564 for_each_populated_zone(zone)
5565 setup_zone_pageset(zone);
5567 for_each_online_pgdat(pgdat)
5568 pgdat->per_cpu_nodestats =
5569 alloc_percpu(struct per_cpu_nodestat);
5572 static __meminit void zone_pcp_init(struct zone *zone)
5575 * per cpu subsystem is not up at this point. The following code
5576 * relies on the ability of the linker to provide the
5577 * offset of a (static) per cpu variable into the per cpu area.
5579 zone->pageset = &boot_pageset;
5581 if (populated_zone(zone))
5582 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5583 zone->name, zone->present_pages,
5584 zone_batchsize(zone));
5587 void __meminit init_currently_empty_zone(struct zone *zone,
5588 unsigned long zone_start_pfn,
5591 struct pglist_data *pgdat = zone->zone_pgdat;
5593 pgdat->nr_zones = zone_idx(zone) + 1;
5595 zone->zone_start_pfn = zone_start_pfn;
5597 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5598 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5600 (unsigned long)zone_idx(zone),
5601 zone_start_pfn, (zone_start_pfn + size));
5603 zone_init_free_lists(zone);
5604 zone->initialized = 1;
5607 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5608 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5611 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5613 int __meminit __early_pfn_to_nid(unsigned long pfn,
5614 struct mminit_pfnnid_cache *state)
5616 unsigned long start_pfn, end_pfn;
5619 if (state->last_start <= pfn && pfn < state->last_end)
5620 return state->last_nid;
5622 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5624 state->last_start = start_pfn;
5625 state->last_end = end_pfn;
5626 state->last_nid = nid;
5631 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5634 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5635 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5636 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5638 * If an architecture guarantees that all ranges registered contain no holes
5639 * and may be freed, this this function may be used instead of calling
5640 * memblock_free_early_nid() manually.
5642 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5644 unsigned long start_pfn, end_pfn;
5647 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5648 start_pfn = min(start_pfn, max_low_pfn);
5649 end_pfn = min(end_pfn, max_low_pfn);
5651 if (start_pfn < end_pfn)
5652 memblock_free_early_nid(PFN_PHYS(start_pfn),
5653 (end_pfn - start_pfn) << PAGE_SHIFT,
5659 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5660 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5662 * If an architecture guarantees that all ranges registered contain no holes and may
5663 * be freed, this function may be used instead of calling memory_present() manually.
5665 void __init sparse_memory_present_with_active_regions(int nid)
5667 unsigned long start_pfn, end_pfn;
5670 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5671 memory_present(this_nid, start_pfn, end_pfn);
5675 * get_pfn_range_for_nid - Return the start and end page frames for a node
5676 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5677 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5678 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5680 * It returns the start and end page frame of a node based on information
5681 * provided by memblock_set_node(). If called for a node
5682 * with no available memory, a warning is printed and the start and end
5685 void __meminit get_pfn_range_for_nid(unsigned int nid,
5686 unsigned long *start_pfn, unsigned long *end_pfn)
5688 unsigned long this_start_pfn, this_end_pfn;
5694 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5695 *start_pfn = min(*start_pfn, this_start_pfn);
5696 *end_pfn = max(*end_pfn, this_end_pfn);
5699 if (*start_pfn == -1UL)
5704 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5705 * assumption is made that zones within a node are ordered in monotonic
5706 * increasing memory addresses so that the "highest" populated zone is used
5708 static void __init find_usable_zone_for_movable(void)
5711 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5712 if (zone_index == ZONE_MOVABLE)
5715 if (arch_zone_highest_possible_pfn[zone_index] >
5716 arch_zone_lowest_possible_pfn[zone_index])
5720 VM_BUG_ON(zone_index == -1);
5721 movable_zone = zone_index;
5725 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5726 * because it is sized independent of architecture. Unlike the other zones,
5727 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5728 * in each node depending on the size of each node and how evenly kernelcore
5729 * is distributed. This helper function adjusts the zone ranges
5730 * provided by the architecture for a given node by using the end of the
5731 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5732 * zones within a node are in order of monotonic increases memory addresses
5734 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5735 unsigned long zone_type,
5736 unsigned long node_start_pfn,
5737 unsigned long node_end_pfn,
5738 unsigned long *zone_start_pfn,
5739 unsigned long *zone_end_pfn)
5741 /* Only adjust if ZONE_MOVABLE is on this node */
5742 if (zone_movable_pfn[nid]) {
5743 /* Size ZONE_MOVABLE */
5744 if (zone_type == ZONE_MOVABLE) {
5745 *zone_start_pfn = zone_movable_pfn[nid];
5746 *zone_end_pfn = min(node_end_pfn,
5747 arch_zone_highest_possible_pfn[movable_zone]);
5749 /* Adjust for ZONE_MOVABLE starting within this range */
5750 } else if (!mirrored_kernelcore &&
5751 *zone_start_pfn < zone_movable_pfn[nid] &&
5752 *zone_end_pfn > zone_movable_pfn[nid]) {
5753 *zone_end_pfn = zone_movable_pfn[nid];
5755 /* Check if this whole range is within ZONE_MOVABLE */
5756 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5757 *zone_start_pfn = *zone_end_pfn;
5762 * Return the number of pages a zone spans in a node, including holes
5763 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5765 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5766 unsigned long zone_type,
5767 unsigned long node_start_pfn,
5768 unsigned long node_end_pfn,
5769 unsigned long *zone_start_pfn,
5770 unsigned long *zone_end_pfn,
5771 unsigned long *ignored)
5773 /* When hotadd a new node from cpu_up(), the node should be empty */
5774 if (!node_start_pfn && !node_end_pfn)
5777 /* Get the start and end of the zone */
5778 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5779 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5780 adjust_zone_range_for_zone_movable(nid, zone_type,
5781 node_start_pfn, node_end_pfn,
5782 zone_start_pfn, zone_end_pfn);
5784 /* Check that this node has pages within the zone's required range */
5785 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5788 /* Move the zone boundaries inside the node if necessary */
5789 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5790 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5792 /* Return the spanned pages */
5793 return *zone_end_pfn - *zone_start_pfn;
5797 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5798 * then all holes in the requested range will be accounted for.
5800 unsigned long __meminit __absent_pages_in_range(int nid,
5801 unsigned long range_start_pfn,
5802 unsigned long range_end_pfn)
5804 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5805 unsigned long start_pfn, end_pfn;
5808 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5809 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5810 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5811 nr_absent -= end_pfn - start_pfn;
5817 * absent_pages_in_range - Return number of page frames in holes within a range
5818 * @start_pfn: The start PFN to start searching for holes
5819 * @end_pfn: The end PFN to stop searching for holes
5821 * It returns the number of pages frames in memory holes within a range.
5823 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5824 unsigned long end_pfn)
5826 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5829 /* Return the number of page frames in holes in a zone on a node */
5830 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5831 unsigned long zone_type,
5832 unsigned long node_start_pfn,
5833 unsigned long node_end_pfn,
5834 unsigned long *ignored)
5836 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5837 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5838 unsigned long zone_start_pfn, zone_end_pfn;
5839 unsigned long nr_absent;
5841 /* When hotadd a new node from cpu_up(), the node should be empty */
5842 if (!node_start_pfn && !node_end_pfn)
5845 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5846 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5848 adjust_zone_range_for_zone_movable(nid, zone_type,
5849 node_start_pfn, node_end_pfn,
5850 &zone_start_pfn, &zone_end_pfn);
5851 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5854 * ZONE_MOVABLE handling.
5855 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5858 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5859 unsigned long start_pfn, end_pfn;
5860 struct memblock_region *r;
5862 for_each_memblock(memory, r) {
5863 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5864 zone_start_pfn, zone_end_pfn);
5865 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5866 zone_start_pfn, zone_end_pfn);
5868 if (zone_type == ZONE_MOVABLE &&
5869 memblock_is_mirror(r))
5870 nr_absent += end_pfn - start_pfn;
5872 if (zone_type == ZONE_NORMAL &&
5873 !memblock_is_mirror(r))
5874 nr_absent += end_pfn - start_pfn;
5881 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5882 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5883 unsigned long zone_type,
5884 unsigned long node_start_pfn,
5885 unsigned long node_end_pfn,
5886 unsigned long *zone_start_pfn,
5887 unsigned long *zone_end_pfn,
5888 unsigned long *zones_size)
5892 *zone_start_pfn = node_start_pfn;
5893 for (zone = 0; zone < zone_type; zone++)
5894 *zone_start_pfn += zones_size[zone];
5896 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5898 return zones_size[zone_type];
5901 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5902 unsigned long zone_type,
5903 unsigned long node_start_pfn,
5904 unsigned long node_end_pfn,
5905 unsigned long *zholes_size)
5910 return zholes_size[zone_type];
5913 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5915 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5916 unsigned long node_start_pfn,
5917 unsigned long node_end_pfn,
5918 unsigned long *zones_size,
5919 unsigned long *zholes_size)
5921 unsigned long realtotalpages = 0, totalpages = 0;
5924 for (i = 0; i < MAX_NR_ZONES; i++) {
5925 struct zone *zone = pgdat->node_zones + i;
5926 unsigned long zone_start_pfn, zone_end_pfn;
5927 unsigned long size, real_size;
5929 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5935 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5936 node_start_pfn, node_end_pfn,
5939 zone->zone_start_pfn = zone_start_pfn;
5941 zone->zone_start_pfn = 0;
5942 zone->spanned_pages = size;
5943 zone->present_pages = real_size;
5946 realtotalpages += real_size;
5949 pgdat->node_spanned_pages = totalpages;
5950 pgdat->node_present_pages = realtotalpages;
5951 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5955 #ifndef CONFIG_SPARSEMEM
5957 * Calculate the size of the zone->blockflags rounded to an unsigned long
5958 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5959 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5960 * round what is now in bits to nearest long in bits, then return it in
5963 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5965 unsigned long usemapsize;
5967 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5968 usemapsize = roundup(zonesize, pageblock_nr_pages);
5969 usemapsize = usemapsize >> pageblock_order;
5970 usemapsize *= NR_PAGEBLOCK_BITS;
5971 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5973 return usemapsize / 8;
5976 static void __init setup_usemap(struct pglist_data *pgdat,
5978 unsigned long zone_start_pfn,
5979 unsigned long zonesize)
5981 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5982 zone->pageblock_flags = NULL;
5984 zone->pageblock_flags =
5985 memblock_virt_alloc_node_nopanic(usemapsize,
5989 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5990 unsigned long zone_start_pfn, unsigned long zonesize) {}
5991 #endif /* CONFIG_SPARSEMEM */
5993 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5995 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5996 void __paginginit set_pageblock_order(void)
6000 /* Check that pageblock_nr_pages has not already been setup */
6001 if (pageblock_order)
6004 if (HPAGE_SHIFT > PAGE_SHIFT)
6005 order = HUGETLB_PAGE_ORDER;
6007 order = MAX_ORDER - 1;
6010 * Assume the largest contiguous order of interest is a huge page.
6011 * This value may be variable depending on boot parameters on IA64 and
6014 pageblock_order = order;
6016 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6019 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6020 * is unused as pageblock_order is set at compile-time. See
6021 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6024 void __paginginit set_pageblock_order(void)
6028 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6030 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6031 unsigned long present_pages)
6033 unsigned long pages = spanned_pages;
6036 * Provide a more accurate estimation if there are holes within
6037 * the zone and SPARSEMEM is in use. If there are holes within the
6038 * zone, each populated memory region may cost us one or two extra
6039 * memmap pages due to alignment because memmap pages for each
6040 * populated regions may not be naturally aligned on page boundary.
6041 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6043 if (spanned_pages > present_pages + (present_pages >> 4) &&
6044 IS_ENABLED(CONFIG_SPARSEMEM))
6045 pages = present_pages;
6047 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6051 * Set up the zone data structures:
6052 * - mark all pages reserved
6053 * - mark all memory queues empty
6054 * - clear the memory bitmaps
6056 * NOTE: pgdat should get zeroed by caller.
6058 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6061 int nid = pgdat->node_id;
6063 pgdat_resize_init(pgdat);
6064 #ifdef CONFIG_NUMA_BALANCING
6065 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6066 pgdat->numabalancing_migrate_nr_pages = 0;
6067 pgdat->numabalancing_migrate_next_window = jiffies;
6069 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6070 spin_lock_init(&pgdat->split_queue_lock);
6071 INIT_LIST_HEAD(&pgdat->split_queue);
6072 pgdat->split_queue_len = 0;
6074 init_waitqueue_head(&pgdat->kswapd_wait);
6075 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6076 #ifdef CONFIG_COMPACTION
6077 init_waitqueue_head(&pgdat->kcompactd_wait);
6079 pgdat_page_ext_init(pgdat);
6080 spin_lock_init(&pgdat->lru_lock);
6081 lruvec_init(node_lruvec(pgdat));
6083 pgdat->per_cpu_nodestats = &boot_nodestats;
6085 for (j = 0; j < MAX_NR_ZONES; j++) {
6086 struct zone *zone = pgdat->node_zones + j;
6087 unsigned long size, realsize, freesize, memmap_pages;
6088 unsigned long zone_start_pfn = zone->zone_start_pfn;
6090 size = zone->spanned_pages;
6091 realsize = freesize = zone->present_pages;
6094 * Adjust freesize so that it accounts for how much memory
6095 * is used by this zone for memmap. This affects the watermark
6096 * and per-cpu initialisations
6098 memmap_pages = calc_memmap_size(size, realsize);
6099 if (!is_highmem_idx(j)) {
6100 if (freesize >= memmap_pages) {
6101 freesize -= memmap_pages;
6104 " %s zone: %lu pages used for memmap\n",
6105 zone_names[j], memmap_pages);
6107 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6108 zone_names[j], memmap_pages, freesize);
6111 /* Account for reserved pages */
6112 if (j == 0 && freesize > dma_reserve) {
6113 freesize -= dma_reserve;
6114 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6115 zone_names[0], dma_reserve);
6118 if (!is_highmem_idx(j))
6119 nr_kernel_pages += freesize;
6120 /* Charge for highmem memmap if there are enough kernel pages */
6121 else if (nr_kernel_pages > memmap_pages * 2)
6122 nr_kernel_pages -= memmap_pages;
6123 nr_all_pages += freesize;
6126 * Set an approximate value for lowmem here, it will be adjusted
6127 * when the bootmem allocator frees pages into the buddy system.
6128 * And all highmem pages will be managed by the buddy system.
6130 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6134 zone->name = zone_names[j];
6135 zone->zone_pgdat = pgdat;
6136 spin_lock_init(&zone->lock);
6137 zone_seqlock_init(zone);
6138 zone_pcp_init(zone);
6143 set_pageblock_order();
6144 setup_usemap(pgdat, zone, zone_start_pfn, size);
6145 init_currently_empty_zone(zone, zone_start_pfn, size);
6146 memmap_init(size, nid, j, zone_start_pfn);
6150 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6151 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6153 unsigned long __maybe_unused start = 0;
6154 unsigned long __maybe_unused offset = 0;
6156 /* Skip empty nodes */
6157 if (!pgdat->node_spanned_pages)
6160 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6161 offset = pgdat->node_start_pfn - start;
6162 /* ia64 gets its own node_mem_map, before this, without bootmem */
6163 if (!pgdat->node_mem_map) {
6164 unsigned long size, end;
6168 * The zone's endpoints aren't required to be MAX_ORDER
6169 * aligned but the node_mem_map endpoints must be in order
6170 * for the buddy allocator to function correctly.
6172 end = pgdat_end_pfn(pgdat);
6173 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6174 size = (end - start) * sizeof(struct page);
6175 map = alloc_remap(pgdat->node_id, size);
6177 map = memblock_virt_alloc_node_nopanic(size,
6179 pgdat->node_mem_map = map + offset;
6181 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6182 __func__, pgdat->node_id, (unsigned long)pgdat,
6183 (unsigned long)pgdat->node_mem_map);
6184 #ifndef CONFIG_NEED_MULTIPLE_NODES
6186 * With no DISCONTIG, the global mem_map is just set as node 0's
6188 if (pgdat == NODE_DATA(0)) {
6189 mem_map = NODE_DATA(0)->node_mem_map;
6190 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6191 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6193 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6198 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6199 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6201 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6202 unsigned long node_start_pfn, unsigned long *zholes_size)
6204 pg_data_t *pgdat = NODE_DATA(nid);
6205 unsigned long start_pfn = 0;
6206 unsigned long end_pfn = 0;
6208 /* pg_data_t should be reset to zero when it's allocated */
6209 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6211 pgdat->node_id = nid;
6212 pgdat->node_start_pfn = node_start_pfn;
6213 pgdat->per_cpu_nodestats = NULL;
6214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6215 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6216 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6217 (u64)start_pfn << PAGE_SHIFT,
6218 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6220 start_pfn = node_start_pfn;
6222 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6223 zones_size, zholes_size);
6225 alloc_node_mem_map(pgdat);
6227 reset_deferred_meminit(pgdat);
6228 free_area_init_core(pgdat);
6231 #ifdef CONFIG_HAVE_MEMBLOCK
6233 * Only struct pages that are backed by physical memory are zeroed and
6234 * initialized by going through __init_single_page(). But, there are some
6235 * struct pages which are reserved in memblock allocator and their fields
6236 * may be accessed (for example page_to_pfn() on some configuration accesses
6237 * flags). We must explicitly zero those struct pages.
6239 void __paginginit zero_resv_unavail(void)
6241 phys_addr_t start, end;
6246 * Loop through ranges that are reserved, but do not have reported
6247 * physical memory backing.
6250 for_each_resv_unavail_range(i, &start, &end) {
6251 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6252 mm_zero_struct_page(pfn_to_page(pfn));
6258 * Struct pages that do not have backing memory. This could be because
6259 * firmware is using some of this memory, or for some other reasons.
6260 * Once memblock is changed so such behaviour is not allowed: i.e.
6261 * list of "reserved" memory must be a subset of list of "memory", then
6262 * this code can be removed.
6265 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6267 #endif /* CONFIG_HAVE_MEMBLOCK */
6269 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6271 #if MAX_NUMNODES > 1
6273 * Figure out the number of possible node ids.
6275 void __init setup_nr_node_ids(void)
6277 unsigned int highest;
6279 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6280 nr_node_ids = highest + 1;
6285 * node_map_pfn_alignment - determine the maximum internode alignment
6287 * This function should be called after node map is populated and sorted.
6288 * It calculates the maximum power of two alignment which can distinguish
6291 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6292 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6293 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6294 * shifted, 1GiB is enough and this function will indicate so.
6296 * This is used to test whether pfn -> nid mapping of the chosen memory
6297 * model has fine enough granularity to avoid incorrect mapping for the
6298 * populated node map.
6300 * Returns the determined alignment in pfn's. 0 if there is no alignment
6301 * requirement (single node).
6303 unsigned long __init node_map_pfn_alignment(void)
6305 unsigned long accl_mask = 0, last_end = 0;
6306 unsigned long start, end, mask;
6310 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6311 if (!start || last_nid < 0 || last_nid == nid) {
6318 * Start with a mask granular enough to pin-point to the
6319 * start pfn and tick off bits one-by-one until it becomes
6320 * too coarse to separate the current node from the last.
6322 mask = ~((1 << __ffs(start)) - 1);
6323 while (mask && last_end <= (start & (mask << 1)))
6326 /* accumulate all internode masks */
6330 /* convert mask to number of pages */
6331 return ~accl_mask + 1;
6334 /* Find the lowest pfn for a node */
6335 static unsigned long __init find_min_pfn_for_node(int nid)
6337 unsigned long min_pfn = ULONG_MAX;
6338 unsigned long start_pfn;
6341 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6342 min_pfn = min(min_pfn, start_pfn);
6344 if (min_pfn == ULONG_MAX) {
6345 pr_warn("Could not find start_pfn for node %d\n", nid);
6353 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6355 * It returns the minimum PFN based on information provided via
6356 * memblock_set_node().
6358 unsigned long __init find_min_pfn_with_active_regions(void)
6360 return find_min_pfn_for_node(MAX_NUMNODES);
6364 * early_calculate_totalpages()
6365 * Sum pages in active regions for movable zone.
6366 * Populate N_MEMORY for calculating usable_nodes.
6368 static unsigned long __init early_calculate_totalpages(void)
6370 unsigned long totalpages = 0;
6371 unsigned long start_pfn, end_pfn;
6374 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6375 unsigned long pages = end_pfn - start_pfn;
6377 totalpages += pages;
6379 node_set_state(nid, N_MEMORY);
6385 * Find the PFN the Movable zone begins in each node. Kernel memory
6386 * is spread evenly between nodes as long as the nodes have enough
6387 * memory. When they don't, some nodes will have more kernelcore than
6390 static void __init find_zone_movable_pfns_for_nodes(void)
6393 unsigned long usable_startpfn;
6394 unsigned long kernelcore_node, kernelcore_remaining;
6395 /* save the state before borrow the nodemask */
6396 nodemask_t saved_node_state = node_states[N_MEMORY];
6397 unsigned long totalpages = early_calculate_totalpages();
6398 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6399 struct memblock_region *r;
6401 /* Need to find movable_zone earlier when movable_node is specified. */
6402 find_usable_zone_for_movable();
6405 * If movable_node is specified, ignore kernelcore and movablecore
6408 if (movable_node_is_enabled()) {
6409 for_each_memblock(memory, r) {
6410 if (!memblock_is_hotpluggable(r))
6415 usable_startpfn = PFN_DOWN(r->base);
6416 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6417 min(usable_startpfn, zone_movable_pfn[nid]) :
6425 * If kernelcore=mirror is specified, ignore movablecore option
6427 if (mirrored_kernelcore) {
6428 bool mem_below_4gb_not_mirrored = false;
6430 for_each_memblock(memory, r) {
6431 if (memblock_is_mirror(r))
6436 usable_startpfn = memblock_region_memory_base_pfn(r);
6438 if (usable_startpfn < 0x100000) {
6439 mem_below_4gb_not_mirrored = true;
6443 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6444 min(usable_startpfn, zone_movable_pfn[nid]) :
6448 if (mem_below_4gb_not_mirrored)
6449 pr_warn("This configuration results in unmirrored kernel memory.");
6455 * If movablecore=nn[KMG] was specified, calculate what size of
6456 * kernelcore that corresponds so that memory usable for
6457 * any allocation type is evenly spread. If both kernelcore
6458 * and movablecore are specified, then the value of kernelcore
6459 * will be used for required_kernelcore if it's greater than
6460 * what movablecore would have allowed.
6462 if (required_movablecore) {
6463 unsigned long corepages;
6466 * Round-up so that ZONE_MOVABLE is at least as large as what
6467 * was requested by the user
6469 required_movablecore =
6470 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6471 required_movablecore = min(totalpages, required_movablecore);
6472 corepages = totalpages - required_movablecore;
6474 required_kernelcore = max(required_kernelcore, corepages);
6478 * If kernelcore was not specified or kernelcore size is larger
6479 * than totalpages, there is no ZONE_MOVABLE.
6481 if (!required_kernelcore || required_kernelcore >= totalpages)
6484 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6485 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6488 /* Spread kernelcore memory as evenly as possible throughout nodes */
6489 kernelcore_node = required_kernelcore / usable_nodes;
6490 for_each_node_state(nid, N_MEMORY) {
6491 unsigned long start_pfn, end_pfn;
6494 * Recalculate kernelcore_node if the division per node
6495 * now exceeds what is necessary to satisfy the requested
6496 * amount of memory for the kernel
6498 if (required_kernelcore < kernelcore_node)
6499 kernelcore_node = required_kernelcore / usable_nodes;
6502 * As the map is walked, we track how much memory is usable
6503 * by the kernel using kernelcore_remaining. When it is
6504 * 0, the rest of the node is usable by ZONE_MOVABLE
6506 kernelcore_remaining = kernelcore_node;
6508 /* Go through each range of PFNs within this node */
6509 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6510 unsigned long size_pages;
6512 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6513 if (start_pfn >= end_pfn)
6516 /* Account for what is only usable for kernelcore */
6517 if (start_pfn < usable_startpfn) {
6518 unsigned long kernel_pages;
6519 kernel_pages = min(end_pfn, usable_startpfn)
6522 kernelcore_remaining -= min(kernel_pages,
6523 kernelcore_remaining);
6524 required_kernelcore -= min(kernel_pages,
6525 required_kernelcore);
6527 /* Continue if range is now fully accounted */
6528 if (end_pfn <= usable_startpfn) {
6531 * Push zone_movable_pfn to the end so
6532 * that if we have to rebalance
6533 * kernelcore across nodes, we will
6534 * not double account here
6536 zone_movable_pfn[nid] = end_pfn;
6539 start_pfn = usable_startpfn;
6543 * The usable PFN range for ZONE_MOVABLE is from
6544 * start_pfn->end_pfn. Calculate size_pages as the
6545 * number of pages used as kernelcore
6547 size_pages = end_pfn - start_pfn;
6548 if (size_pages > kernelcore_remaining)
6549 size_pages = kernelcore_remaining;
6550 zone_movable_pfn[nid] = start_pfn + size_pages;
6553 * Some kernelcore has been met, update counts and
6554 * break if the kernelcore for this node has been
6557 required_kernelcore -= min(required_kernelcore,
6559 kernelcore_remaining -= size_pages;
6560 if (!kernelcore_remaining)
6566 * If there is still required_kernelcore, we do another pass with one
6567 * less node in the count. This will push zone_movable_pfn[nid] further
6568 * along on the nodes that still have memory until kernelcore is
6572 if (usable_nodes && required_kernelcore > usable_nodes)
6576 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6577 for (nid = 0; nid < MAX_NUMNODES; nid++)
6578 zone_movable_pfn[nid] =
6579 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6582 /* restore the node_state */
6583 node_states[N_MEMORY] = saved_node_state;
6586 /* Any regular or high memory on that node ? */
6587 static void check_for_memory(pg_data_t *pgdat, int nid)
6589 enum zone_type zone_type;
6591 if (N_MEMORY == N_NORMAL_MEMORY)
6594 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6595 struct zone *zone = &pgdat->node_zones[zone_type];
6596 if (populated_zone(zone)) {
6597 node_set_state(nid, N_HIGH_MEMORY);
6598 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6599 zone_type <= ZONE_NORMAL)
6600 node_set_state(nid, N_NORMAL_MEMORY);
6607 * free_area_init_nodes - Initialise all pg_data_t and zone data
6608 * @max_zone_pfn: an array of max PFNs for each zone
6610 * This will call free_area_init_node() for each active node in the system.
6611 * Using the page ranges provided by memblock_set_node(), the size of each
6612 * zone in each node and their holes is calculated. If the maximum PFN
6613 * between two adjacent zones match, it is assumed that the zone is empty.
6614 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6615 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6616 * starts where the previous one ended. For example, ZONE_DMA32 starts
6617 * at arch_max_dma_pfn.
6619 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6621 unsigned long start_pfn, end_pfn;
6624 /* Record where the zone boundaries are */
6625 memset(arch_zone_lowest_possible_pfn, 0,
6626 sizeof(arch_zone_lowest_possible_pfn));
6627 memset(arch_zone_highest_possible_pfn, 0,
6628 sizeof(arch_zone_highest_possible_pfn));
6630 start_pfn = find_min_pfn_with_active_regions();
6632 for (i = 0; i < MAX_NR_ZONES; i++) {
6633 if (i == ZONE_MOVABLE)
6636 end_pfn = max(max_zone_pfn[i], start_pfn);
6637 arch_zone_lowest_possible_pfn[i] = start_pfn;
6638 arch_zone_highest_possible_pfn[i] = end_pfn;
6640 start_pfn = end_pfn;
6643 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6644 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6645 find_zone_movable_pfns_for_nodes();
6647 /* Print out the zone ranges */
6648 pr_info("Zone ranges:\n");
6649 for (i = 0; i < MAX_NR_ZONES; i++) {
6650 if (i == ZONE_MOVABLE)
6652 pr_info(" %-8s ", zone_names[i]);
6653 if (arch_zone_lowest_possible_pfn[i] ==
6654 arch_zone_highest_possible_pfn[i])
6657 pr_cont("[mem %#018Lx-%#018Lx]\n",
6658 (u64)arch_zone_lowest_possible_pfn[i]
6660 ((u64)arch_zone_highest_possible_pfn[i]
6661 << PAGE_SHIFT) - 1);
6664 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6665 pr_info("Movable zone start for each node\n");
6666 for (i = 0; i < MAX_NUMNODES; i++) {
6667 if (zone_movable_pfn[i])
6668 pr_info(" Node %d: %#018Lx\n", i,
6669 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6672 /* Print out the early node map */
6673 pr_info("Early memory node ranges\n");
6674 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6675 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6676 (u64)start_pfn << PAGE_SHIFT,
6677 ((u64)end_pfn << PAGE_SHIFT) - 1);
6679 /* Initialise every node */
6680 mminit_verify_pageflags_layout();
6681 setup_nr_node_ids();
6682 for_each_online_node(nid) {
6683 pg_data_t *pgdat = NODE_DATA(nid);
6684 free_area_init_node(nid, NULL,
6685 find_min_pfn_for_node(nid), NULL);
6687 /* Any memory on that node */
6688 if (pgdat->node_present_pages)
6689 node_set_state(nid, N_MEMORY);
6690 check_for_memory(pgdat, nid);
6692 zero_resv_unavail();
6695 static int __init cmdline_parse_core(char *p, unsigned long *core)
6697 unsigned long long coremem;
6701 coremem = memparse(p, &p);
6702 *core = coremem >> PAGE_SHIFT;
6704 /* Paranoid check that UL is enough for the coremem value */
6705 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6711 * kernelcore=size sets the amount of memory for use for allocations that
6712 * cannot be reclaimed or migrated.
6714 static int __init cmdline_parse_kernelcore(char *p)
6716 /* parse kernelcore=mirror */
6717 if (parse_option_str(p, "mirror")) {
6718 mirrored_kernelcore = true;
6722 return cmdline_parse_core(p, &required_kernelcore);
6726 * movablecore=size sets the amount of memory for use for allocations that
6727 * can be reclaimed or migrated.
6729 static int __init cmdline_parse_movablecore(char *p)
6731 return cmdline_parse_core(p, &required_movablecore);
6734 early_param("kernelcore", cmdline_parse_kernelcore);
6735 early_param("movablecore", cmdline_parse_movablecore);
6737 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6739 void adjust_managed_page_count(struct page *page, long count)
6741 spin_lock(&managed_page_count_lock);
6742 page_zone(page)->managed_pages += count;
6743 totalram_pages += count;
6744 #ifdef CONFIG_HIGHMEM
6745 if (PageHighMem(page))
6746 totalhigh_pages += count;
6748 spin_unlock(&managed_page_count_lock);
6750 EXPORT_SYMBOL(adjust_managed_page_count);
6752 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6755 unsigned long pages = 0;
6757 start = (void *)PAGE_ALIGN((unsigned long)start);
6758 end = (void *)((unsigned long)end & PAGE_MASK);
6759 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6760 if ((unsigned int)poison <= 0xFF)
6761 memset(pos, poison, PAGE_SIZE);
6762 free_reserved_page(virt_to_page(pos));
6766 pr_info("Freeing %s memory: %ldK\n",
6767 s, pages << (PAGE_SHIFT - 10));
6771 EXPORT_SYMBOL(free_reserved_area);
6773 #ifdef CONFIG_HIGHMEM
6774 void free_highmem_page(struct page *page)
6776 __free_reserved_page(page);
6778 page_zone(page)->managed_pages++;
6784 void __init mem_init_print_info(const char *str)
6786 unsigned long physpages, codesize, datasize, rosize, bss_size;
6787 unsigned long init_code_size, init_data_size;
6789 physpages = get_num_physpages();
6790 codesize = _etext - _stext;
6791 datasize = _edata - _sdata;
6792 rosize = __end_rodata - __start_rodata;
6793 bss_size = __bss_stop - __bss_start;
6794 init_data_size = __init_end - __init_begin;
6795 init_code_size = _einittext - _sinittext;
6798 * Detect special cases and adjust section sizes accordingly:
6799 * 1) .init.* may be embedded into .data sections
6800 * 2) .init.text.* may be out of [__init_begin, __init_end],
6801 * please refer to arch/tile/kernel/vmlinux.lds.S.
6802 * 3) .rodata.* may be embedded into .text or .data sections.
6804 #define adj_init_size(start, end, size, pos, adj) \
6806 if (start <= pos && pos < end && size > adj) \
6810 adj_init_size(__init_begin, __init_end, init_data_size,
6811 _sinittext, init_code_size);
6812 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6813 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6814 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6815 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6817 #undef adj_init_size
6819 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6820 #ifdef CONFIG_HIGHMEM
6824 nr_free_pages() << (PAGE_SHIFT - 10),
6825 physpages << (PAGE_SHIFT - 10),
6826 codesize >> 10, datasize >> 10, rosize >> 10,
6827 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6828 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6829 totalcma_pages << (PAGE_SHIFT - 10),
6830 #ifdef CONFIG_HIGHMEM
6831 totalhigh_pages << (PAGE_SHIFT - 10),
6833 str ? ", " : "", str ? str : "");
6837 * set_dma_reserve - set the specified number of pages reserved in the first zone
6838 * @new_dma_reserve: The number of pages to mark reserved
6840 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6841 * In the DMA zone, a significant percentage may be consumed by kernel image
6842 * and other unfreeable allocations which can skew the watermarks badly. This
6843 * function may optionally be used to account for unfreeable pages in the
6844 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6845 * smaller per-cpu batchsize.
6847 void __init set_dma_reserve(unsigned long new_dma_reserve)
6849 dma_reserve = new_dma_reserve;
6852 void __init free_area_init(unsigned long *zones_size)
6854 free_area_init_node(0, zones_size,
6855 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6856 zero_resv_unavail();
6859 static int page_alloc_cpu_dead(unsigned int cpu)
6862 lru_add_drain_cpu(cpu);
6866 * Spill the event counters of the dead processor
6867 * into the current processors event counters.
6868 * This artificially elevates the count of the current
6871 vm_events_fold_cpu(cpu);
6874 * Zero the differential counters of the dead processor
6875 * so that the vm statistics are consistent.
6877 * This is only okay since the processor is dead and cannot
6878 * race with what we are doing.
6880 cpu_vm_stats_fold(cpu);
6884 void __init page_alloc_init(void)
6888 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6889 "mm/page_alloc:dead", NULL,
6890 page_alloc_cpu_dead);
6895 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6896 * or min_free_kbytes changes.
6898 static void calculate_totalreserve_pages(void)
6900 struct pglist_data *pgdat;
6901 unsigned long reserve_pages = 0;
6902 enum zone_type i, j;
6904 for_each_online_pgdat(pgdat) {
6906 pgdat->totalreserve_pages = 0;
6908 for (i = 0; i < MAX_NR_ZONES; i++) {
6909 struct zone *zone = pgdat->node_zones + i;
6912 /* Find valid and maximum lowmem_reserve in the zone */
6913 for (j = i; j < MAX_NR_ZONES; j++) {
6914 if (zone->lowmem_reserve[j] > max)
6915 max = zone->lowmem_reserve[j];
6918 /* we treat the high watermark as reserved pages. */
6919 max += high_wmark_pages(zone);
6921 if (max > zone->managed_pages)
6922 max = zone->managed_pages;
6924 pgdat->totalreserve_pages += max;
6926 reserve_pages += max;
6929 totalreserve_pages = reserve_pages;
6933 * setup_per_zone_lowmem_reserve - called whenever
6934 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6935 * has a correct pages reserved value, so an adequate number of
6936 * pages are left in the zone after a successful __alloc_pages().
6938 static void setup_per_zone_lowmem_reserve(void)
6940 struct pglist_data *pgdat;
6941 enum zone_type j, idx;
6943 for_each_online_pgdat(pgdat) {
6944 for (j = 0; j < MAX_NR_ZONES; j++) {
6945 struct zone *zone = pgdat->node_zones + j;
6946 unsigned long managed_pages = zone->managed_pages;
6948 zone->lowmem_reserve[j] = 0;
6952 struct zone *lower_zone;
6956 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6957 sysctl_lowmem_reserve_ratio[idx] = 1;
6959 lower_zone = pgdat->node_zones + idx;
6960 lower_zone->lowmem_reserve[j] = managed_pages /
6961 sysctl_lowmem_reserve_ratio[idx];
6962 managed_pages += lower_zone->managed_pages;
6967 /* update totalreserve_pages */
6968 calculate_totalreserve_pages();
6971 static void __setup_per_zone_wmarks(void)
6973 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6974 unsigned long lowmem_pages = 0;
6976 unsigned long flags;
6978 /* Calculate total number of !ZONE_HIGHMEM pages */
6979 for_each_zone(zone) {
6980 if (!is_highmem(zone))
6981 lowmem_pages += zone->managed_pages;
6984 for_each_zone(zone) {
6987 spin_lock_irqsave(&zone->lock, flags);
6988 tmp = (u64)pages_min * zone->managed_pages;
6989 do_div(tmp, lowmem_pages);
6990 if (is_highmem(zone)) {
6992 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6993 * need highmem pages, so cap pages_min to a small
6996 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6997 * deltas control asynch page reclaim, and so should
6998 * not be capped for highmem.
7000 unsigned long min_pages;
7002 min_pages = zone->managed_pages / 1024;
7003 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7004 zone->watermark[WMARK_MIN] = min_pages;
7007 * If it's a lowmem zone, reserve a number of pages
7008 * proportionate to the zone's size.
7010 zone->watermark[WMARK_MIN] = tmp;
7014 * Set the kswapd watermarks distance according to the
7015 * scale factor in proportion to available memory, but
7016 * ensure a minimum size on small systems.
7018 tmp = max_t(u64, tmp >> 2,
7019 mult_frac(zone->managed_pages,
7020 watermark_scale_factor, 10000));
7022 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7023 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7025 spin_unlock_irqrestore(&zone->lock, flags);
7028 /* update totalreserve_pages */
7029 calculate_totalreserve_pages();
7033 * setup_per_zone_wmarks - called when min_free_kbytes changes
7034 * or when memory is hot-{added|removed}
7036 * Ensures that the watermark[min,low,high] values for each zone are set
7037 * correctly with respect to min_free_kbytes.
7039 void setup_per_zone_wmarks(void)
7041 static DEFINE_SPINLOCK(lock);
7044 __setup_per_zone_wmarks();
7049 * Initialise min_free_kbytes.
7051 * For small machines we want it small (128k min). For large machines
7052 * we want it large (64MB max). But it is not linear, because network
7053 * bandwidth does not increase linearly with machine size. We use
7055 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7056 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7072 int __meminit init_per_zone_wmark_min(void)
7074 unsigned long lowmem_kbytes;
7075 int new_min_free_kbytes;
7077 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7078 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7080 if (new_min_free_kbytes > user_min_free_kbytes) {
7081 min_free_kbytes = new_min_free_kbytes;
7082 if (min_free_kbytes < 128)
7083 min_free_kbytes = 128;
7084 if (min_free_kbytes > 65536)
7085 min_free_kbytes = 65536;
7087 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7088 new_min_free_kbytes, user_min_free_kbytes);
7090 setup_per_zone_wmarks();
7091 refresh_zone_stat_thresholds();
7092 setup_per_zone_lowmem_reserve();
7095 setup_min_unmapped_ratio();
7096 setup_min_slab_ratio();
7101 core_initcall(init_per_zone_wmark_min)
7104 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7105 * that we can call two helper functions whenever min_free_kbytes
7108 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7109 void __user *buffer, size_t *length, loff_t *ppos)
7113 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7118 user_min_free_kbytes = min_free_kbytes;
7119 setup_per_zone_wmarks();
7124 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7125 void __user *buffer, size_t *length, loff_t *ppos)
7129 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7134 setup_per_zone_wmarks();
7140 static void setup_min_unmapped_ratio(void)
7145 for_each_online_pgdat(pgdat)
7146 pgdat->min_unmapped_pages = 0;
7149 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7150 sysctl_min_unmapped_ratio) / 100;
7154 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7155 void __user *buffer, size_t *length, loff_t *ppos)
7159 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7163 setup_min_unmapped_ratio();
7168 static void setup_min_slab_ratio(void)
7173 for_each_online_pgdat(pgdat)
7174 pgdat->min_slab_pages = 0;
7177 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7178 sysctl_min_slab_ratio) / 100;
7181 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7182 void __user *buffer, size_t *length, loff_t *ppos)
7186 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7190 setup_min_slab_ratio();
7197 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7198 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7199 * whenever sysctl_lowmem_reserve_ratio changes.
7201 * The reserve ratio obviously has absolutely no relation with the
7202 * minimum watermarks. The lowmem reserve ratio can only make sense
7203 * if in function of the boot time zone sizes.
7205 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7206 void __user *buffer, size_t *length, loff_t *ppos)
7208 proc_dointvec_minmax(table, write, buffer, length, ppos);
7209 setup_per_zone_lowmem_reserve();
7214 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7215 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7216 * pagelist can have before it gets flushed back to buddy allocator.
7218 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7219 void __user *buffer, size_t *length, loff_t *ppos)
7222 int old_percpu_pagelist_fraction;
7225 mutex_lock(&pcp_batch_high_lock);
7226 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7228 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7229 if (!write || ret < 0)
7232 /* Sanity checking to avoid pcp imbalance */
7233 if (percpu_pagelist_fraction &&
7234 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7235 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7241 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7244 for_each_populated_zone(zone) {
7247 for_each_possible_cpu(cpu)
7248 pageset_set_high_and_batch(zone,
7249 per_cpu_ptr(zone->pageset, cpu));
7252 mutex_unlock(&pcp_batch_high_lock);
7257 int hashdist = HASHDIST_DEFAULT;
7259 static int __init set_hashdist(char *str)
7263 hashdist = simple_strtoul(str, &str, 0);
7266 __setup("hashdist=", set_hashdist);
7269 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7271 * Returns the number of pages that arch has reserved but
7272 * is not known to alloc_large_system_hash().
7274 static unsigned long __init arch_reserved_kernel_pages(void)
7281 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7282 * machines. As memory size is increased the scale is also increased but at
7283 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7284 * quadruples the scale is increased by one, which means the size of hash table
7285 * only doubles, instead of quadrupling as well.
7286 * Because 32-bit systems cannot have large physical memory, where this scaling
7287 * makes sense, it is disabled on such platforms.
7289 #if __BITS_PER_LONG > 32
7290 #define ADAPT_SCALE_BASE (64ul << 30)
7291 #define ADAPT_SCALE_SHIFT 2
7292 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7296 * allocate a large system hash table from bootmem
7297 * - it is assumed that the hash table must contain an exact power-of-2
7298 * quantity of entries
7299 * - limit is the number of hash buckets, not the total allocation size
7301 void *__init alloc_large_system_hash(const char *tablename,
7302 unsigned long bucketsize,
7303 unsigned long numentries,
7306 unsigned int *_hash_shift,
7307 unsigned int *_hash_mask,
7308 unsigned long low_limit,
7309 unsigned long high_limit)
7311 unsigned long long max = high_limit;
7312 unsigned long log2qty, size;
7316 /* allow the kernel cmdline to have a say */
7318 /* round applicable memory size up to nearest megabyte */
7319 numentries = nr_kernel_pages;
7320 numentries -= arch_reserved_kernel_pages();
7322 /* It isn't necessary when PAGE_SIZE >= 1MB */
7323 if (PAGE_SHIFT < 20)
7324 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7326 #if __BITS_PER_LONG > 32
7328 unsigned long adapt;
7330 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7331 adapt <<= ADAPT_SCALE_SHIFT)
7336 /* limit to 1 bucket per 2^scale bytes of low memory */
7337 if (scale > PAGE_SHIFT)
7338 numentries >>= (scale - PAGE_SHIFT);
7340 numentries <<= (PAGE_SHIFT - scale);
7342 /* Make sure we've got at least a 0-order allocation.. */
7343 if (unlikely(flags & HASH_SMALL)) {
7344 /* Makes no sense without HASH_EARLY */
7345 WARN_ON(!(flags & HASH_EARLY));
7346 if (!(numentries >> *_hash_shift)) {
7347 numentries = 1UL << *_hash_shift;
7348 BUG_ON(!numentries);
7350 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7351 numentries = PAGE_SIZE / bucketsize;
7353 numentries = roundup_pow_of_two(numentries);
7355 /* limit allocation size to 1/16 total memory by default */
7357 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7358 do_div(max, bucketsize);
7360 max = min(max, 0x80000000ULL);
7362 if (numentries < low_limit)
7363 numentries = low_limit;
7364 if (numentries > max)
7367 log2qty = ilog2(numentries);
7369 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7371 size = bucketsize << log2qty;
7372 if (flags & HASH_EARLY) {
7373 if (flags & HASH_ZERO)
7374 table = memblock_virt_alloc_nopanic(size, 0);
7376 table = memblock_virt_alloc_raw(size, 0);
7377 } else if (hashdist) {
7378 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7381 * If bucketsize is not a power-of-two, we may free
7382 * some pages at the end of hash table which
7383 * alloc_pages_exact() automatically does
7385 if (get_order(size) < MAX_ORDER) {
7386 table = alloc_pages_exact(size, gfp_flags);
7387 kmemleak_alloc(table, size, 1, gfp_flags);
7390 } while (!table && size > PAGE_SIZE && --log2qty);
7393 panic("Failed to allocate %s hash table\n", tablename);
7395 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7396 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7399 *_hash_shift = log2qty;
7401 *_hash_mask = (1 << log2qty) - 1;
7407 * This function checks whether pageblock includes unmovable pages or not.
7408 * If @count is not zero, it is okay to include less @count unmovable pages
7410 * PageLRU check without isolation or lru_lock could race so that
7411 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7412 * check without lock_page also may miss some movable non-lru pages at
7413 * race condition. So you can't expect this function should be exact.
7415 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7417 bool skip_hwpoisoned_pages)
7419 unsigned long pfn, iter, found;
7422 * For avoiding noise data, lru_add_drain_all() should be called
7423 * If ZONE_MOVABLE, the zone never contains unmovable pages
7425 if (zone_idx(zone) == ZONE_MOVABLE)
7429 * CMA allocations (alloc_contig_range) really need to mark isolate
7430 * CMA pageblocks even when they are not movable in fact so consider
7431 * them movable here.
7433 if (is_migrate_cma(migratetype) &&
7434 is_migrate_cma(get_pageblock_migratetype(page)))
7437 pfn = page_to_pfn(page);
7438 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7439 unsigned long check = pfn + iter;
7441 if (!pfn_valid_within(check))
7444 page = pfn_to_page(check);
7446 if (PageReserved(page))
7450 * Hugepages are not in LRU lists, but they're movable.
7451 * We need not scan over tail pages bacause we don't
7452 * handle each tail page individually in migration.
7454 if (PageHuge(page)) {
7455 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7460 * We can't use page_count without pin a page
7461 * because another CPU can free compound page.
7462 * This check already skips compound tails of THP
7463 * because their page->_refcount is zero at all time.
7465 if (!page_ref_count(page)) {
7466 if (PageBuddy(page))
7467 iter += (1 << page_order(page)) - 1;
7472 * The HWPoisoned page may be not in buddy system, and
7473 * page_count() is not 0.
7475 if (skip_hwpoisoned_pages && PageHWPoison(page))
7478 if (__PageMovable(page))
7484 * If there are RECLAIMABLE pages, we need to check
7485 * it. But now, memory offline itself doesn't call
7486 * shrink_node_slabs() and it still to be fixed.
7489 * If the page is not RAM, page_count()should be 0.
7490 * we don't need more check. This is an _used_ not-movable page.
7492 * The problematic thing here is PG_reserved pages. PG_reserved
7493 * is set to both of a memory hole page and a _used_ kernel
7502 bool is_pageblock_removable_nolock(struct page *page)
7508 * We have to be careful here because we are iterating over memory
7509 * sections which are not zone aware so we might end up outside of
7510 * the zone but still within the section.
7511 * We have to take care about the node as well. If the node is offline
7512 * its NODE_DATA will be NULL - see page_zone.
7514 if (!node_online(page_to_nid(page)))
7517 zone = page_zone(page);
7518 pfn = page_to_pfn(page);
7519 if (!zone_spans_pfn(zone, pfn))
7522 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7525 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7527 static unsigned long pfn_max_align_down(unsigned long pfn)
7529 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7530 pageblock_nr_pages) - 1);
7533 static unsigned long pfn_max_align_up(unsigned long pfn)
7535 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7536 pageblock_nr_pages));
7539 /* [start, end) must belong to a single zone. */
7540 static int __alloc_contig_migrate_range(struct compact_control *cc,
7541 unsigned long start, unsigned long end)
7543 /* This function is based on compact_zone() from compaction.c. */
7544 unsigned long nr_reclaimed;
7545 unsigned long pfn = start;
7546 unsigned int tries = 0;
7551 while (pfn < end || !list_empty(&cc->migratepages)) {
7552 if (fatal_signal_pending(current)) {
7557 if (list_empty(&cc->migratepages)) {
7558 cc->nr_migratepages = 0;
7559 pfn = isolate_migratepages_range(cc, pfn, end);
7565 } else if (++tries == 5) {
7566 ret = ret < 0 ? ret : -EBUSY;
7570 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7572 cc->nr_migratepages -= nr_reclaimed;
7574 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7575 NULL, 0, cc->mode, MR_CMA);
7578 putback_movable_pages(&cc->migratepages);
7585 * alloc_contig_range() -- tries to allocate given range of pages
7586 * @start: start PFN to allocate
7587 * @end: one-past-the-last PFN to allocate
7588 * @migratetype: migratetype of the underlaying pageblocks (either
7589 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7590 * in range must have the same migratetype and it must
7591 * be either of the two.
7592 * @gfp_mask: GFP mask to use during compaction
7594 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7595 * aligned, however it's the caller's responsibility to guarantee that
7596 * we are the only thread that changes migrate type of pageblocks the
7599 * The PFN range must belong to a single zone.
7601 * Returns zero on success or negative error code. On success all
7602 * pages which PFN is in [start, end) are allocated for the caller and
7603 * need to be freed with free_contig_range().
7605 int alloc_contig_range(unsigned long start, unsigned long end,
7606 unsigned migratetype, gfp_t gfp_mask)
7608 unsigned long outer_start, outer_end;
7612 struct compact_control cc = {
7613 .nr_migratepages = 0,
7615 .zone = page_zone(pfn_to_page(start)),
7616 .mode = MIGRATE_SYNC,
7617 .ignore_skip_hint = true,
7618 .no_set_skip_hint = true,
7619 .gfp_mask = current_gfp_context(gfp_mask),
7621 INIT_LIST_HEAD(&cc.migratepages);
7624 * What we do here is we mark all pageblocks in range as
7625 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7626 * have different sizes, and due to the way page allocator
7627 * work, we align the range to biggest of the two pages so
7628 * that page allocator won't try to merge buddies from
7629 * different pageblocks and change MIGRATE_ISOLATE to some
7630 * other migration type.
7632 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7633 * migrate the pages from an unaligned range (ie. pages that
7634 * we are interested in). This will put all the pages in
7635 * range back to page allocator as MIGRATE_ISOLATE.
7637 * When this is done, we take the pages in range from page
7638 * allocator removing them from the buddy system. This way
7639 * page allocator will never consider using them.
7641 * This lets us mark the pageblocks back as
7642 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7643 * aligned range but not in the unaligned, original range are
7644 * put back to page allocator so that buddy can use them.
7647 ret = start_isolate_page_range(pfn_max_align_down(start),
7648 pfn_max_align_up(end), migratetype,
7654 * In case of -EBUSY, we'd like to know which page causes problem.
7655 * So, just fall through. test_pages_isolated() has a tracepoint
7656 * which will report the busy page.
7658 * It is possible that busy pages could become available before
7659 * the call to test_pages_isolated, and the range will actually be
7660 * allocated. So, if we fall through be sure to clear ret so that
7661 * -EBUSY is not accidentally used or returned to caller.
7663 ret = __alloc_contig_migrate_range(&cc, start, end);
7664 if (ret && ret != -EBUSY)
7669 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7670 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7671 * more, all pages in [start, end) are free in page allocator.
7672 * What we are going to do is to allocate all pages from
7673 * [start, end) (that is remove them from page allocator).
7675 * The only problem is that pages at the beginning and at the
7676 * end of interesting range may be not aligned with pages that
7677 * page allocator holds, ie. they can be part of higher order
7678 * pages. Because of this, we reserve the bigger range and
7679 * once this is done free the pages we are not interested in.
7681 * We don't have to hold zone->lock here because the pages are
7682 * isolated thus they won't get removed from buddy.
7685 lru_add_drain_all();
7686 drain_all_pages(cc.zone);
7689 outer_start = start;
7690 while (!PageBuddy(pfn_to_page(outer_start))) {
7691 if (++order >= MAX_ORDER) {
7692 outer_start = start;
7695 outer_start &= ~0UL << order;
7698 if (outer_start != start) {
7699 order = page_order(pfn_to_page(outer_start));
7702 * outer_start page could be small order buddy page and
7703 * it doesn't include start page. Adjust outer_start
7704 * in this case to report failed page properly
7705 * on tracepoint in test_pages_isolated()
7707 if (outer_start + (1UL << order) <= start)
7708 outer_start = start;
7711 /* Make sure the range is really isolated. */
7712 if (test_pages_isolated(outer_start, end, false)) {
7713 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7714 __func__, outer_start, end);
7719 /* Grab isolated pages from freelists. */
7720 outer_end = isolate_freepages_range(&cc, outer_start, end);
7726 /* Free head and tail (if any) */
7727 if (start != outer_start)
7728 free_contig_range(outer_start, start - outer_start);
7729 if (end != outer_end)
7730 free_contig_range(end, outer_end - end);
7733 undo_isolate_page_range(pfn_max_align_down(start),
7734 pfn_max_align_up(end), migratetype);
7738 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7740 unsigned int count = 0;
7742 for (; nr_pages--; pfn++) {
7743 struct page *page = pfn_to_page(pfn);
7745 count += page_count(page) != 1;
7748 WARN(count != 0, "%d pages are still in use!\n", count);
7752 #ifdef CONFIG_MEMORY_HOTPLUG
7754 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7755 * page high values need to be recalulated.
7757 void __meminit zone_pcp_update(struct zone *zone)
7760 mutex_lock(&pcp_batch_high_lock);
7761 for_each_possible_cpu(cpu)
7762 pageset_set_high_and_batch(zone,
7763 per_cpu_ptr(zone->pageset, cpu));
7764 mutex_unlock(&pcp_batch_high_lock);
7768 void zone_pcp_reset(struct zone *zone)
7770 unsigned long flags;
7772 struct per_cpu_pageset *pset;
7774 /* avoid races with drain_pages() */
7775 local_irq_save(flags);
7776 if (zone->pageset != &boot_pageset) {
7777 for_each_online_cpu(cpu) {
7778 pset = per_cpu_ptr(zone->pageset, cpu);
7779 drain_zonestat(zone, pset);
7781 free_percpu(zone->pageset);
7782 zone->pageset = &boot_pageset;
7784 local_irq_restore(flags);
7787 #ifdef CONFIG_MEMORY_HOTREMOVE
7789 * All pages in the range must be in a single zone and isolated
7790 * before calling this.
7793 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7797 unsigned int order, i;
7799 unsigned long flags;
7800 /* find the first valid pfn */
7801 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7806 offline_mem_sections(pfn, end_pfn);
7807 zone = page_zone(pfn_to_page(pfn));
7808 spin_lock_irqsave(&zone->lock, flags);
7810 while (pfn < end_pfn) {
7811 if (!pfn_valid(pfn)) {
7815 page = pfn_to_page(pfn);
7817 * The HWPoisoned page may be not in buddy system, and
7818 * page_count() is not 0.
7820 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7822 SetPageReserved(page);
7826 BUG_ON(page_count(page));
7827 BUG_ON(!PageBuddy(page));
7828 order = page_order(page);
7829 #ifdef CONFIG_DEBUG_VM
7830 pr_info("remove from free list %lx %d %lx\n",
7831 pfn, 1 << order, end_pfn);
7833 list_del(&page->lru);
7834 rmv_page_order(page);
7835 zone->free_area[order].nr_free--;
7836 for (i = 0; i < (1 << order); i++)
7837 SetPageReserved((page+i));
7838 pfn += (1 << order);
7840 spin_unlock_irqrestore(&zone->lock, flags);
7844 bool is_free_buddy_page(struct page *page)
7846 struct zone *zone = page_zone(page);
7847 unsigned long pfn = page_to_pfn(page);
7848 unsigned long flags;
7851 spin_lock_irqsave(&zone->lock, flags);
7852 for (order = 0; order < MAX_ORDER; order++) {
7853 struct page *page_head = page - (pfn & ((1 << order) - 1));
7855 if (PageBuddy(page_head) && page_order(page_head) >= order)
7858 spin_unlock_irqrestore(&zone->lock, flags);
7860 return order < MAX_ORDER;